Appendix C - Estimating The Economic Cost of Load Shedding in South Africa Report

Estimating the economic cost
of load shedding in South Africa

A report for Eskom Holdings (SOC) Ltd.
03 December 2020
Authors
Kay Walsh, MCom (economics).

Rachel Theron, MCom (economics)
Ahmed Seedat BBus Sci (finance) CA (SA).

Chris Reeders, B.Eng (industrial)
The authors would like to thank Associate Professor Rulof Burger from the University of Stellenbosch
for giving input to the econometric approach to estimating the economic cost of load shedding
and Associate Professor Roula Inglesi-Lotz from the University of Pretoria for her comments on the
study design. . In addition, we would like to acknowledge the assistance of several Eskom personnel
who assisted us by providing insight and information during the course of the study. These include
Dr Ulrich Minaar, Mr Deon Joubert Mr Nelson Nunes, Mr Hendri Bower, Ms Stefanie Visser, Mr
Andries Gildenhuys and Mr Dhamesh Bhana.
The opinions expressed herein are those of the authors and do not necessarily represent the views
of Eskom SOC Holdings or its employees.
Nova Economics
16 Elektron Road
Technopark, Stellenbosch
www.novaeconomics.com
© Nova Economics – economic and strategy consulting
Novꭤ Economics
Table of contents
Executive summary ..................................................................................................................................................viii
1. Background and context .................................................................................................................................. 1
1.1. Purpose of the study and key objectives ......................................................................................... 1
1.2. What were the events that precipitated the power crisis in South Africa? ............................ 1
1.3. When and how often did load shedding occur? .......................................................................... 3
1.4. Why does load shedding occur? The reserve margin in the context of long-term
electricity supply shortages ............................................................................................................................... 5
2. Electricity supply interruptions – the key concepts ................................................................................. 8
2.1. Introduction ............................................................................................................................................. 8
2.2. Distinguishing between planned and unplanned supply interruptions ................................. 8
2.3. What measures have been used to estimate the cost of outages in South Africa? .......... 9
2.3.1. What can the CoUE and CoLS measures be used for? ........................................................10
2.4. The costs associated with power outages ..................................................................................... 11
3. Estimating the magnitude of load shedding............................................................................................14
3.1. Introduction ............................................................................................................................................14
3.2. Two approaches to estimating the magnitude of load shedding ..........................................14
3.2.1. Top-down estimate of load shedding by Eskom’s system operator ................................15
3.2.2. Bottom-up estimate of load shedding ......................................................................................15
3.3. Comparison of the two estimates of load shedding ..................................................................19
3.4. Load shedding as a percentage of total electricity sales ..........................................................21
4. Methods used to estimate the cost of load shedding ......................................................................... 22
4.1. Broad analytical approaches that can be taken to estimating the cost of load shedding
22
4.1.1. Forward-looking approach to estimating the cost of planned outages ........................ 22
4.1.2. Retrospective approach to estimating the cost of planned outages .............................. 22
4.2. Previous estimates of the cost of load shedding in South Africa .......................................... 25
4.3. Previous econometric studies of the cost of load shedding ................................................... 27
4.4. Our approach to estimating the CoLS .......................................................................................... 28
4.5. Estimation of the CoLS based on the expenditure-side models of GDP ............................ 29
4.5.1. Introduction ...................................................................................................................................... 29
4.5.2. Overview of data and data sources........................................................................................... 29
ii
4.5.3. Approach to estimating the CoLS based on a classic linear regression model
(CLRM) 30
4.5.4. Approach to estimating the CoLS based on an auto-regressive distributed lag
model 32
4.6. Estimation of the CoLS based on an energy-augmented Cobb-Douglas production
function ................................................................................................................................................................. 34
4.6.1. Introduction ...................................................................................................................................... 34
4.6.2. Overview of data and data sources........................................................................................... 36
5. Results ................................................................................................................................................................. 40
5.1. Introduction ........................................................................................................................................... 40
5.2. National estimates of the cost of load shedding........................................................................ 40
5.2.1. The cost of load shedding in rand per kilowatt-hour .......................................................... 40
5.2.2. The impact of load shedding on the South African economy ...........................................41
5.2.3. Alternative estimates of the CoLS .............................................................................................. 43
5.2.4. Comparison with previous load shedding estimates ........................................................... 45
5.3. The impact of load shedding by industry ..................................................................................... 45
5.3.1. Cost of load shedding by industry (in R/kWh) ....................................................................... 45
5.3.2. Normalised costs of load shedding ........................................................................................... 46
5.3.3. Impact of load shedding on GDP growth by industry ......................................................... 47
5.3.4. Impact of load shedding on GDP, by industry in billions of rand .................................... 48
5.3.5. Comparison of sectoral estimates of the CoLS with the CoUE ......................................... 49
5.3.6. Putting the economic cost of load shedding in perspective ..............................................51
5.4. Caveats to the results.......................................................................................................................... 52
5.4.1. Our estimates of the CoLS are conservative because they exclude longer-term
impacts of load shedding on potential GDP growth ............................................................................ 52
5.4.2. We were only able to estimate the costs of load shedding at a macroeconomic
level 52
6. Key insights and recommendations........................................................................................................... 53
6.1. The cost of load shedding – key insights...................................................................................... 53
6.2. The potential uses of the CoLS estimates ..................................................................................... 54
Estimating the magnitude of load shedding. .................................................................... 59
Detailed approach to producing the top-down estimate ....................................................... 59
Bottom-up estimate ............................................................................................................................ 62
Review of previous studies on the cost of power outages............................................ 66
iii
Summary of previous studies on the cost of load shedding in South Africa ..................... 66
Review of forward-looking studies focusing on those based on general equilibrium
models................................................................................................................................................................... 68
Review of retrospective studies with a focus on econometric methods ............................. 75
Energy-augmented production function ............................................................................ 79
The method used to derive the quarterly capital stock series................................................ 82
The method used to derive the quarterly series on electricity sales by sector ................. 82
Expenditure side regression output ..................................................................................... 85
Impact of load shedding on GDP growth by sector ....................................................... 87
Industry classification by SIC code ........................................................................................ 92
iv
List of figures
Figure 1 Monthly data on the incidence of load shedding, 2007 until 2019 ............................................. 4
Figure 2 Relationship between the operating reserve margin and load shedding ................................ 7
Figure 3 Load shedding incidences by quarter, 2007-2020 .........................................................................14
Figure 4 Composition of the bottom-up estimate of load shedding ........................................................18
Figure 5 Comparison of the two estimates of the magnitude of load shedding .................................. 20
Figure 6 Load shedding as a percentage of sales, quarterly .......................................................................21
Figure 7 Load shedding as a percentage of sales, monthly .........................................................................21
Figure 8 Analytical approaches and empirical methods used to estimate the cost of load shedding
....................................................................................................................................................................................... 23
Figure 9 Scatterplot of economic growth (GDP q/q % change) against load shedding (as % of
sales) ............................................................................................................................................................................. 25
Figure 10 Approach to estimating CoLS ............................................................................................................ 28
Figure 11 Change in GDP against load shedding as a percentage of total electricity sales ................31
Figure 12 Trend in real GDP and load shedding in GWh, 2007 to 2019 .................................................. 32
Figure 13 Manufacturing industry production function variables, normalised (2003=100) ................ 35
Figure 14 Trend in quarterly fixed capital stock, by industry, 2000 to 2019 ............................................ 37
Figure 15 Electricity sales by industry, seasonally adjusted .......................................................................... 39
Figure 16 The impact of load shedding on quarterly GDP growth, 2007 to 2019 .................................41
Figure 17 Impact of load shedding on GDP (rand billion, constant 2020) by period ...........................41
Figure 18 Impact of load shedding on GDP growth q/q .............................................................................. 42
Figure 19 Impact of load shedding on GDP (rand billion, constant 2020) .............................................. 42
Figure 20 Cost of load shedding by periods and estimation method (R/kWh, 2020 constant prices)
....................................................................................................................................................................................... 44
Figure 21 Contribution of each sector to the total cost of load shedding (R/kWh) ............................. 46
Figure 22 Contribution to total GDP by sector (%) ........................................................................................ 47
Figure 23 CoLS: actual and normalised values (R/kWh, 2020 values) ...................................................... 47
Figure 24 Impact of load shedding GDP growth by industry, 2007 to 2019 .......................................... 48
Figure 25 Decrease in GDP attributed to load shedding, per period and by industry (2020 prices)
....................................................................................................................................................................................... 49
Figure 26 Estimate of CoUE by sector ............................................................................................................... 50
Figure 27 Estimate of CoLS by sector ................................................................................................................ 50
Figure 28 Impact of load shedding on GDP growth, compared to Covid-19 and the financial crisis
........................................................................................................................................................................................51
v
List of tables
Table 1 The distinction between the two main types of electricity supply interruptions....................... 8
Table 2 Potential applications of measures of planned and unplanned outages.................................. 11
Table 3 Factors that impact the cost of electricity outages ..........................................................................12
Table 4 Key differences in approach and data sources .................................................................................15
Table 5 Data and variables used for the expenditure-side model of GDP ............................................. 30
Table 6 Data sources and variables- Cobb-Douglas production function ............................................. 36
Table 7 Primary estimate of the cost of load shedding, 2007/08 to 2018/19 ......................................... 40
Table 8 Cost of load shedding (R/kWh, 2020 constant prices) .................................................................. 43
Table 9 Comparison of our primary estimate of the CoLS with Deloitte 2009...................................... 45
Table 10 Types of supply-side and demand-side interventions given persistent outages ................. 55
vi
List of abbreviations
Abbreviation Description
ARDL Autoregressive distributed lag
CLRM Classical linear regression model
CGE Computable general equilibrium
CoLS Cost of load shedding
CoUE Cost of unserved energy
DCGE Dynamic computable general equilibrium
FBM Feeder balancing module
GE General equilibrium
GW Gigawatt
GWh Gigawatt hour
GDP Gross Domestic Product
GFCF Gross fixed capital formation
GVA Gross -value add
IPP Independent power producer
I-O Input-output
IoS Interruption of supply
kWh Kilowatt-hour
LFS Labour force survey
LSDV Least-squares dummy variable
MW Megawatt
MWh Megawatt-hour
NERSA National Energy Regulator of South Africa
OVB Omitted-variable bias
OCGT Open cycle gas turbine
OLS Ordinary least squares
PE Partial equilibrium
PALMS Post-Apartheid Labour Market Series
QLFS Quarterly labour force survey
q/q Quarter-on-quarter
Rm Rand million
SAM Social accounting matrix
SARB South African Reserve Bank
SIC Standard industrial classification
StatsSA Statistics South Africa
SNA System of National Accounts
UCT University of Cape Town
VoLL Value of lost load
vii
Executive summary
In late 2007, South Africans experienced the first of what would become a recurring series of
nationwide load shedding episodes. Load shedding refers to the deliberate shutdown of parts of
the electricity distribution network to avoid damaging the electricity grid and to safeguard against
a national blackout. It is usually implemented after alternative options to balance demand and
supply have been exhausted. Load shedding is implemented to reduce electricity demand, preserve
grid stability, and to prevent the collapse of the system. 1
The first load shedding episode in October 2007, marked the beginning of a national electricity
supply crisis that has persisted for over a decade. Weekly data from the Eskom’s system operator
show that load shedding occurred during 33 months between 2007 and 2019. There have been
three distinct periods of load shedding over the past 12 years – the first ran from 2007 to 2008, the
second from 2013 to 2015, and the most recent from 2018 to late 2020. While load shedding has
continued into 2020, we limited to our analysis to the 12 years ending in 2019.
Load shedding has caused significant disruption in the daily lives of South Africans and the national
economy. The purpose of this study was to provide Eskom with reliable and accurate estimates of
the economic cost of load shedding in South Africa.
We have estimated that Load shedding cost the South African economy nearly R35 billion in the 12
years between 2007 and 2019 2. Had all the load shedding experienced over the period taken place
in a single quarter in 2019, it would have resulted in a 5% contraction real q/q GDP growth. To put
this into perspective the total cost of load shedding at R35 billion is roughly equivalent to the impact
the 2008/9 financial crisis had on GDP growth (it also subtracted a cumulative five percentage
points from quarter-on-quarter GDP growth albeit over a much shorter period).
Our results suggest that the cost of load shedding (CoLS), expressed in rand per kilowatt-hour,
increased over the three main periods in which it occurred. During the first period (2007 to 2008)
the CoLS was R7.61/kWh 3, it rose to R8.80/kWh during the second period (2013 to 2015) to
R9.53/kWh in the third period (2018 to 2019). Our results of the total CoLS are similar to Eskom’s
previous estimate of R8.95/kWh 4, which was produced by Deloitte in 2009. 5
Another objective of this study was to give Eskom insight into the distribution of the CoLS across
different sectors of the economy. We were able to produce estimates of the CoLS for all sectors
1
Also referred to as rolling blackouts
2
Expressed in 2020 prices.
3
4
5
Deloitte, "Modelling the impacts of electricity disruptions, Chapter 3, Report on Eskom and the Electricity Sector," ed.
Landon Mc Millan (2009).
Novꭤ Economics viii

(defined at the 2-digit SIC level), except for mining and quarrying (SIC 2). 6 Our results show that
the CoLS is unevenly distributed - four of the nine industries, namely manufacturing (SIC 3),
transport and communication (SIC 6), wholesale and retail trade (SIC 5) and agriculture, hunting,
forestry and fishing (SIC 1), bore 80% of the total cost. . The manufacturing sector alone, shouldered
nearly 40% of the total cost.
We also normalised the CoLS to illustrate the impact of load shedding on each sector relative to its
size (contribution to total GDP). For example, we saw that while the agricultural sector was the
worst-affected, the manufacturing sector, which accounts for 13% of GDP carried the highest
proportion of the total CoLS. While the agricultural industry is a relatively small contributor to
national GDP (it accounts for 3.6% of total output) it lost 4.2 times more GDP per kWh of load
shedding than the average. Manufacturing (SIC 3) and utilities (SIC 4) lost three times more output
than the average. The output of the most service-oriented sectors was largely unaffected by load
shedding – this included financial and business (SIC 7) and community, social and personal services
industries (SIC 8).
The total cost of regular planned outages, as defined in the international literature, is a function of
the damages and costs incurred by a firm, its inherent resilience and ability to adapt. A firm may
incur costs related to direct and/or indirect damages (e.g. lost production or reduced productivity).
The inherent resilience of a particular firm or industry refers to its ability to shift production around
outages while the ability to adapt refers to the extent to which it can invest in alternative sources or
back-up generation.
There are several reasons, for example, why one would expect service-oriented industries to be
inherently more resilient to power outages and better able to adapt. By way of illustration, personnel
in the finance and business services industry can continue to work during power outages if the
electronic devices they rely on, such as laptops and IT systems, are fitted with back-up power
generation sources. Working hours in this industry tend to be more flexible so that people can shift
their working hours to better accommodate load shedding. Our estimate of the CoLS for the finance
industry (if the total cost to the economy is normalised to R1/kWh) is just three cents per kWh.
Our estimates of the CoLS are based on an econometric analysis of the historical relationship
between load shedding and GDP. Our estimates capture the CoLS, as reflected in the variation in
GDP growth (q/q) around its long-term trend. This includes direct and indirect damages (e.g. loss
of output) and the costs of adaptation and mitigation (e.g. higher costs such as investment in back-
up generation).
These estimates, however, do not include the longer-term costs of load-shedding (e.g. related to
reduced business and investor confidence). These longer-term costs would be reflected in a lower
long-run trend GDP growth, but it is not possible to estimate what the trend in growth might have
been if load shedding had not occurred. While many factors have contributed to the gradual decline
6
A reliable estimate of the impact of load shedding on mining GDP could not be obtained as it was not possible to
control for more than 10% of the variation in the highly volatile quarter-on-quarter growth in Mining GDP.
ix
in South Africa’s trend GDP growth over the past decade, there is little doubt that load shedding
and persistent electricity supply constraints were among them. From this perspective, our estimates
of the CoLS can be considered somewhat conservative because they exclude the longer-term
impact of recurring outages on business and investor confidence which cannot be easily identified
with econometric techniques.
Finally, it was envisaged that the estimates of the CoLS could inform future energy-sector policy
and strategic decision making. In particular, measures of the cost of outages are useful in assessing
the relative costs of interventions to mitigate against the risk of future load shedding, and so can
be used to make socially optimal investment decisions.
Eskom and its shareholder, the South African Government, face several choices when assessing
how best to mitigate against the risk of further load shedding. There are several potential options
to consider on both the demand and supply-side, but all these interventions are associated with
financial and/or broader economic costs.
The immediate options are limited but on the supply-side, they would include running emergency
generation or peaking plant (e.g. diesel-fired open cycle gas turbines) at higher load factors and
on the demand-side power buybacks from customers already under contract, voluntary curtailment
by top customers or as a last resort load shedding. There are more options in the short-to-medium
term, including returning moth-balled power plants to service, building or procuring utility-scale
renewable capacity, investing in large-scale energy-efficiency and demand-side management
programmes or entering into new interruptible supply agreements.
The socially optimal choice is the one that minimises the net cost to society. It is also important,
however, to consider who bears the costs. For example, if the system operator decides to run
Eskom’s peaking plant (OCGTs) at higher load factors in a bid to avoid load shedding, Eskom bears
the cost and must motivate to recover the costs from the consumer via a higher tariff.
While the cost of load shedding at R9.53/kWh is higher than the cost of running OCGTs estimated
by EPRI at R1.99/kWh (2015 prices), it is borne mainly by the most energy-intensive sectors of the
economy (e.g. manufacturing, mining, and transport, while Eskom itself only bears a small
proportion the cost (i.e. lost electricity sales).
x
1. Background and context
1.1. Purpose of the study and key objectives
The purpose of this study was to provide Eskom with reliable and accurate estimates the economic
cost of load shedding (CoLS) between 2007 to 2019. Eskom’s previous estimate of the CoLS was
produced by Deloitte in 2009. The main objective of this study was to update Eskom’s 2009 estimate
of the national CoLS. A further objective was to explore how the cost of load shedding differs and
is distributed across different sectors of the economy.
It is envisaged that estimates of the CoLS will be useful in informing energy sector policy. Measures
of the cost of outages are particularly useful in assessing the relative economic costs of interventions
to mitigate the future risk of load shedding, to make socially optimal investment decisions.
Our estimates of the CoLS were produced using econometric analysis of the historical relationship
between the magnitude and incidence of load shedding and GDP. Our estimates capture the CoLS,
as reflected in the variation in GDP growth (q/q) around its long-term trend. This would include
direct and indirect damages (e.g. loss of output) and the costs of adaptation and mitigation (e.g.
higher costs such as investment in back-up generation).
These estimates of the CoLS, however, can still be considered conservative, because they exclude
the longer-term impact of recurring outages on business and investor confidence which cannot be
captured using econometric techniques. These longer-term costs would be reflected in a lower
long-run trend GDP growth, but it is not possible to estimate what the trend in growth might have
been if load shedding had not occurred. While many factors have contributed to the gradual decline
in South Africa’s trend GDP growth over the past decade, there is little doubt that load shedding
and persistent electricity supply constraints were among them.
1.2. What were the events that precipitated the power crisis in South Africa?
In October 2007, the lights went out as South Africa experienced the first in a series of nationwide
load shedding episodes. Load shedding refers to the deliberate shutdown of parts of the electricity
distribution network to avoid damaging the electricity grid and to safeguard against a national
blackout.
When electricity demand exceeds the available supply, the electricity grid becomes unstable. This
could cause generation units to trip, further compromising the system. A loss of generation
capacity, in this context, increases the load on the remaining units and, in a worst-case scenario, a
Novꭤ Economics Background and context 1

cascading effect with multiple power station failures culminating in a national blackout. Such a
collapse of the electricity grid would leave the country without electricity for several days. 7
To avoid this, a utility or system operator will implement load shedding when national electricity
demand is threatening to exceed supply (when alternative interventions to increase supply in the
short-term have been exhausted). Rotational load shedding is usually implemented in blocks of two
to four hours – with various parts of the network affected at different times.
While several factors contributed to the emergence of South Africa’s electricity supply crisis in 2007,
chief among them was the failure by government to implement the ambitious electricity sector
reforms that had been outlined in the 1998 Energy Sector White Paper. The model of power-sector
reform laid out in the White Paper recommended the vertical and horizontal unbundling of Eskom
to separate the potentially competitive components of the industry (e.g. generation) from those
that are natural monopoly (i.e. transmission and distribution). The main objectives of the policy
paper included attracting private sector investment into the electricity generation sector, ensuring
a transition to cost-reflective electricity prices, expanding basic access to electricity and
consolidating the highly fragmented municipal distribution industry.
The South African Government, however, had not appreciated the extent to which Eskom’s highly
subsidised electricity price, would deter the private sector from investing in the sector. The price of
electricity had been implicitly subsidised by the government for over a decade and did not reflect
the true cost of generating, transmitting and distributing power. There was therefore little financial
incentive, for the private sector to invest.
At that stage, there was also no specific regulatory framework to facilitate the participation of
independent power producers (IPPs). Ian McRae, a former CEO of Eskom, noted in 2009 that
despite its intention to source power from IPPs since 1998, by 2009 it had failed to sign a single
power purchase agreement: “The government failed to recognise that IPPs would not rush into
South Africa to compete with Eskom’s large, low-cost, coal-fired stations”.
In 2004, Thulani Gcabashe, then CEO of Eskom, warned the parliamentary portfolio committee that
available generation capacity was “reaching its limit”. As a result, the government was forced to
reconsider the position adopted in 2001 that Eskom should be prohibited from building new
generation capacity and, in 2004, gave Eskom the green light to embark on a five-year capacity
expansion programme. The then Managing Director of Eskom Enterprises, Brian Dames, noted that
this would begin with the restoration of mothballed plants to service and the installation of two new
open-cycle gas turbines (OCGTs) to serve as peak generation capacity. He, however, noted that
Eskom’s decision on the commissioning of new baseload capacity would be deferred until 2010. 8
In 2007, Eskom warned that the system would become constrained within the next five or six years.
The utility and called for a collaborative effort from all stakeholders to minimise the likelihood of
7
Public Affairs Research Institute (PARI), Why the Lights Went Out: Reform in the South African Energy Sector. , Graduate
School of Development Policy and Practice, University of Cape Town (2013).
8
Public Affairs Research Institute (PARI), Why the Lights Went Out: Reform in the South African Energy Sector. .

power interruptions. Shortly thereafter, in October 2007, Eskom implemented the first round of
load shedding. Then-President Mbeki accepted responsibility for the oversight in planning. On the
12th of December 2007, he made a public apology, noting that “Eskom was right, the government
was wrong.” As a result, in 2007 the Eskom board was able to fast-track the approval of a massive
capacity expansion programme including the construction of two large coal-fired power stations -
Medupi and Kusile. 9
The failure by government and its regulatory authorities (over several decades) to transition to cost-
reflective electricity tariffs also contributed to the emergence of the power crisis. By 2007, the real
electricity price reached all-time lows. Eskom had neither the capital reserves nor the future revenue
stream to cover the cost of the new build programme. This prompted the National Energy
Regulator of South Africa (NERSA) to approve several sharp increases in annual tariffs. The
regulatory methodology applied by NERSA allows Eskom to recover all its prudently and efficiently
incurred costs including a return on capital invested in new generation assets. Between 2008 and
2013, electricity prices more than doubled in real terms (inflation-adjusted) rising by a cumulative
114%, while nominal prices rose by 191% over the same period. 10
The sharp increases in real electricity tariffs over this period provoked a public outcry, and NERSA
subsequently decided to limit the increase in real electricity tariff to ~2% per year between 2013
and 2018. This was much lower than the increase that Eskom required to reach cost-reflective tariff
(CPI plus ~10% per year). NERSA disallowed a substantial proportion (over R100 billion) of Eskom’s
budgeted costs between 2013 and 2018. This limited Eskom’s ability to mitigate the risk of further
load shedding, as there was limited funding for demand or supply-side initiatives.
A previous study, produced by researchers at the University of Cape Town, noted that the
implementation of an unsustainable maintenance strategy of ‘keeping the lights on at all costs’ and
poor coal planning, contracting and procurement also contributed to repeated load shedding. 11
The study suggests that Eskom’s ability to respond to the crisis was also hampered by a substantial
loss in skills and capabilities – the result of sector reform and transformation policies which
encouraged early retirement of many of the utility’s most experienced staff.
1.3. When and how often did load shedding occur?
Eskom’s system operator provided us with weekly data on the incidence and magnitude of load
shedding from 2007 to 2019. While load shedding continued into 2020, we limited to our analysis
to the period between 2007 and 2019. This was firstly, because comparable economic data for 2020
had not yet been released and secondly because it would have been difficult to accurately isolate
9
"ESKOM 2000 - 2008 - Our Recent Past -"Shift performance and grow sustainably","
https://www.eskom.co.za/sites/heritage/Pages/2000.aspx.
10
Deloitte, "The macroeconomic impacts of alternative scenarios to meet Eskom’s five-year revenue requirement," (16
June 2017 2017). http://www.eskom.co.za/Documents/EcoStudyMacroeconomicImpact2017.pdf.
11
Public Affairs Research Institute (PARI), Why the Lights Went Out: Reform in the South African Energy Sector.

the impact of load shedding from the tremendous economic shock precipitated by the COVID-19
pandemic during 2020.
Load shedding occurred in a total of 33 months between 2007 and 2019 (Figure 1). From the data
illustrated in Figure 1, it is clear that Between 2007 and 2019 load shedding occurred in three main
periods:
• Period 1: October 2007 to April 2008 (load shedding during 7 of 7 months)
• Period 2: November 2013 to October 2015 (load shedding during 16 of 24 months)

• Period 3: June 2018 to September 2020. The 2020 incidents were excluded from the scope
of our analysis – we focused on the period until the end of December 2019 where load
shedding occurred during 9 of 19 months.
According to estimates provided by the system operator, the average monthly magnitude of load
shedding was 122 GW. The single month in which the largest amount of load shedding occurred
was March 2019 when an estimated 420 GWh of electricity demand could not be met. The most
sustained period of load shedding was from 2013 to 2015 when load shedding occurred in 16
months.
Figure 1 Monthly data on the incidence of load shedding, 2007 until 2019
Period 1: Period 2: Period 3:

2007/10 to 2008/04 2013/11 to 2015/10 2018/06 to 2019/12
872 GWh 1 742 GWh 1 307 GWh
500
400
Load shedding (GWh)
300
200
100
0
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Source: Nova Economics analysis, based on data supplied by the Eskom system operator

1.4. Why does load shedding occur? The reserve margin in the context of long-
term electricity supply shortages
An electricity utility must continuously match the supply of electricity from its generation fleet with
the load demanded by its customers. However, producing an accurate forecast of electricity
demand, both in the short-term (i.e. day-ahead) and long-term, is a complicated task. To avoid
underestimating electricity demand and to account for other unforeseen factors, electricity utilities
operate at a slight oversupply. This surplus capacity is referred to as the reserve margin.
The utility’s system operator must balance the cost of surplus capacity with the risk of not meeting
demand. The reserve margin also provides some leeway for irregular spikes in demand and
unexpected failures in supply. Eskom’s targeted operating reserve margin is 15% (i.e. dispatchable
electricity supply exceeds forecast peak demand by 15%), which is consistent with optimal levels in
literature. In the literature, adequate reliability is defined as “the level of reserves that provide an
expectation of less than one event in 10 years due to generation deficiency. 12, 13
A narrow reserve margin poses a substantial risk to the stability of the national grid as a result of a
catastrophic cascading failure, while margins that are much above targeted optimal levels are
inefficient. Regular planned outages occur when the reserve margin regularly drops below the
target level over time. In this case, the margin does not provide an adequate buffer against
unexpected electricity supply or demand side deviations. 14 Once the reserve margin has been
eroded, the system operator may be forced to implement load shedding.
The reserve margin depicted represents the surplus of capacity over demand during peak periods
(Figure 2). While the initial period of load shedding came to an end in mid-2008, this was not
because the underlying electricity supply shortage had been resolved but rather because of the
negative impact of the 2008/9 global financial crisis on the SA economy (Figure 2). A sharp
contraction in economic activity led to a parallel drop in electricity demand, and Eskom’s reserve
margin increased to a comfortable 15 to 20%. However, with the recovery of the South African
economy and international commodity markets in general, electricity demand increased. This
economic recovery led to the gradual erosion of Eskom’s reserve margin, once again.
After the first period of load shedding, Eskom adopted a policy of ‘keeping the lights on’ which
meant postponing scheduled maintenance to prevent load shedding at all costs. At roughly the
same time, Eskom began to roll out an energy-efficiency and demand-side management
programme on behalf of the Department of Energy.
By mid-2013, however, it was clear that Eskom could no longer afford to postpone maintenance
because inadequate maintenance was adversely affecting plant reliability and the reserve margin
had fallen to just 3%. Eskom was forced to reintroduced rotational load shedding, from 2013 to the
12
Deloitte, "Modelling the impacts of electricity disruptions, Chapter 3, Report on Eskom and the Electricity Sector."
13
Kevin Carden and Nick Wintermantel, The economic ramifications of resource adequacy, Astrape Consulting (for EISPC
and NARUC) (2013), 45.
14

end of 2015, to “create room” to resume scheduled maintenance. The period with the most
sustained and significant load shedding was the final quarter of 2015. The reserve margin was
restored to comfortable levels in 2016 and 2017 as new generation capacity including unit 5 of
Medupi, Unit 1 of Kusile and Ingula pumped storage came online. A large proportion of new
renewable IPP generation capacity was also commissioned (almost 5 000 MW).
In late 2018/19 Eskom reintroduced load shedding in the context of severe financial challenges. The
National Energy Regulator of South Africa (NERSA) approved tariff increases that were far below
what Eskom required in terms of the approved methodology, and this further jeopardised the
utility’s financial sustainability as sales continued to stagnate. Eskom reported in its 2019 Integrated
Report 15 that it had faced severe operational challenges related to coal supply and quality issues,
deteriorating generating plant performance due partly due to a lack of funds to carry out planned
capital expenditure and maintenance. Eskom also faced uncertainty about restructuring and was in
the process of addressing previously reported incidences of irregular spending and was in the
process of restoring good governance.
15
Eskom, "Integrated Report 2019," (2019). https://www.eskom.co.za/IR2019/Pages/default.aspx.

Figure 2 Relationship between the operating reserve margin and load shedding
Load shedding Operating reserve margin
Instituted policy of keeping Eskom could no longer Ingula, Medupi 5, Kusile NERSA MYPD4 decision
the lights on & postponing postpone maintenance, 1 and IPP renewable jeopardises financial
planned maintenance. Also dealt a financial blow plants commissioned. sustainability. Challenges related
introduced demand-side by inadequate tariff Plant availability to coal supply, quality & prices.
50 management increase. increased due improved Addressing governance issues.500
Load shedding (GWh)

maintenance
Reserve margin (%)
40 Slump in electricity demand 400

due to the impact of 2008/9
global financial crisis
30 300
20 200
10 100
0 0
-10 -100
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Source: Nova Economics analysis data provided by Eskom

2. Electricity supply interruptions – the key concepts
2.1. Introduction
In this section, we discuss some of the key concepts relevant to measuring the costs of electricity
supply interruptions. These include the distinction between planned and unplanned outages, the
measures that are typically used and the types of costs associated with outages.
2.2. Distinguishing between planned and unplanned supply interruptions
The literature on power outages distinguishes between two main types of interruptions - infrequent,
unplanned outages and regular, planned outages (Table 1). 16 Unplanned outages are defined as
infrequent, of short duration (usually lasting less than three hours). Unplanned outages take place
within even the most well-planned systems and often in the context of an adequate reserve margin.
Because they are infrequent, most consumers do not anticipate them and do not invest in back-up
supply or processes to mitigate against the risk of unplanned outages. 17
Planned outages by contrast, usually take place in the context of a persistent structural shortage of
electricity capacity, and an inadequate operating reserve margin. Because the outages are planned,
consumers have time to adapt 18 Consumers may adapt by investing in mitigating measures (e.g.
.
backup generation and storage) or by employing resilience tactics (e.g. reducing reliance on
electricity as a single source of energy or shifting production to avoid load shedding).
Table 1 The distinction between the two main types of electricity supply interruptions
Unplanned outages Electricity supply shortages

• Infrequent, unplanned occur suddenly • Regular, planned and unplanned outages
• Usually < 3 hours duration • Typically scheduled, but sometimes sudden
• Electricity demand temporarily exceeds supply • Due to a structural or long-term shortage of
• Stable reliable power supply, an adequate reserve electricity capacity, usually low and middle-income
margin countries (e.g. Zambia, Nigeria, South Africa, Pakistan)
• Very few customers have a back-up supply or invest • Power system reserve margin is inadequate
in other mitigation measures • Customers invest in back-up supply or other
• Occur in both developed and developing countries mitigation measures
• Also referred to as the cost of power shortages
Cost of unserved energy (CoUE)* Cost of load shedding (CoLS)**
16
Mohan. Munasinghe and A. Sanghui, "Reliability of Electricity Supply, Outage Cost and Value of Service: An Overview,"
The Energy Journal (1988), https://doi.org/10.5547/ISSN0195-6574-EJ-Vol9-NoSI2-1 Source: RePEc.
17
Eskom, "Cost of Unserved Energy Methodology," (2015).
18
We use the term consumers to collectively refer to the various types of electricity users, including households,
businesses, government institutions, etc.
Novꭤ Economics Electricity supply interruptions – the key concepts 8

• The value (in rand per kWh) that is placed on a unit of • The value (in rand per kWh) that is placed on a unit of
energy not supplied due to an unplanned outage of electricity not delivered due to frequent, recurring,
short duration and planned outages
• Also referred to as the Value of Lost Load (VoLL) • Accounts for the inherent resilience and adaptive
• Electricity system planners aim to balance the CoUE response of end-users (i.e. investing in mitigating
against the cost to supply the energy not serve measures or inherently more resilient e.g. due to
• In SA, a methodology was developed to estimate ability to substitute)
CoUE for use in transmission and distribution network • CoLS can be used to assess the relative costs of
planning (as required by the Distribution Network interventions to reduce the risk of power outages
Code) occurring in future
* Used synonymously in the electric industry and the literature with “Value of Customer Reliability,” “Value of Lost Load,”
“Cost of unplanned or unexpected interruptions,”
**Referred to in the literature as the cost of planned outages
Most consumers do not anticipate unplanned outages and do not invest in back-up supply or
processes to mitigate against the risk of unplanned outages. 19 Planned outages, by contrast, usually
take place in the context of a persistent structural shortage of electricity capacity, and an inadequate
operating reserve margin. Because the outages are planned, consumers have time to adapt.20
Consumers may adapt by investing in mitigating measures (e.g. backup generation and storage)
or by employing resilience tactics (e.g. reducing reliance on electricity as a single source of energy
or shifting production to avoid load shedding).
2.3. What measures have been used to estimate the cost of outages in South
Africa?
In South Africa, the cost of unserved energy (CoUE) is the measure used to provide an estimate of
the cost of unplanned outages. In the international literature, it is referred to alternately as the CoUE
or the value of lost load (VoLL). While there are many ways to estimate the CoUE, it is most simply
approximated as the gross-value add (GVA) produced for every unit of electricity consumed (e.g.
R/kWh). The CoUE, therefore, represents the electricity intensity of each type of economic activity
or region. The total CoUE for South Africa was estimated at R77.30 GVA/kWh in 2013 while the
estimate of the CoUE approved by NERSA for use in 2020 is R91.72 GVA/kWh. 21
Eskom’s previous estimate of the cost of planned outages (and specifically load shedding) was
produced by Deloitte in 2009. 22 The Deloitte report estimated the CoLS using a forward-looking
dynamic computable general equilibrium (DCGE) model. The report modelled various scenarios for
productivity decline due to load shedding (assuming that load shedding renders capital stock idle)
i.e. 10%, 5% and 2% decline, as well as a ‘realistic’ load shedding scenario where different sectors
19
Eskom, "Cost of Unserved Energy Methodology," (2015).
20
We use the term consumers to collectively refer to the various types of electricity users, including households,
businesses, government institutions, etc.
21
Eskom Sustainability Division, "Cost of Unserved Energy Update: 2019" (2019). http://nersa.org.za/regulator-decisions/.
22
Deloitte 2009

were affected to varying degrees. Their estimate of the CoLS, based on a simulation of a realistic
load shedding scenario, was R4.92/kWh in 2008 values which equates to R8.95/kWh in 2020 values.
As would be expected, previous estimates of the CoLS in South Africa are much lower than the
CoUE, because the cost of regular, planned outages takes into account the ability of consumers to
adapt, to substitute the factors of production, to invest in a back-up generation or to work around
the outages and limit losses in output. Methodologies to estimate the CoLS also typically take into
account the inherent resilience of some industries – for example, employees at a professional
services firm may be able to continue working when the power goes off or to simply shift their work
to a time later in the day.
According to De Nooij et al 23 CoUE measures are used internationally to make socially optimal
investment decisions and to determine which customers should be cut off in times of electricity
supply shortages. Internationally, the CoUE is often used to assess whether investments in
transmission and distribution networks are feasible or socially optimal. The CoUE has, however, also
been used in tariff design and generation planning in some countries (Table 2).
Minaar, Visser and Crafford 24 note that in the South African context the CoUE values are used
exclusively to inform socially optimal investment decisions for utility systems. Both the South African
Transmission Grid Code and the Distribution Network Code require the regulator – the National
Energy Regulator of South Africa (NERSA) – to approve a method of determining the CoUE as an
economic parameter for network investment criteria. The CoUE is also used in generation planning
in South Africa where it has been used as an input to the Integrated Resource Plan to quantify the
risk of economic damage (at macro-economic level) as a result of generation capacity inadequacy.
The CoUE is a proxy for the cost of unplanned, infrequent outages of short duration and is,
therefore, best used when weighing up the relative costs and benefits of interventions to reduce
unplanned outages or designing tariffs to compensate users for these interruptions, such as
interruptible service charges.
The CoLS can be used when assessing the costs and benefits or initiatives to reduce the likelihood
of load shedding, i.e. regular, planned outages. These include strategic decisions about whether to
employ emergency capacity to avoid load shedding, such as peaking plants. The CoLS can also be
used as an input to generation planning, to assess the relative costs of outages compared to
investments in new generation capacity or demand-side management initiatives, tariff design, or
compensation for voluntary curtailment (Table 2).
23
Michiel. de Nooij, Carl. Koopmans, and Carlijn. Bijvoet, "The value of supply security - the costs of power interruptionsL
Economic input for damage reduction and investment in networks.," Energy Economics 29 (2007).
24
UJ Minnaar, W Visser, and J Crafford, "An economic model for the cost of electricity service interruption in South
Africa," Utilities Policy 48 (2017).

Table 2 Potential applications of measures of planned and unplanned outages
Category Applications Measure
Generation planning • Optimising the reserve margin CoLS or CoUE

• Investment allocation criteria
• Cogeneration and IPP planning
• Return to service plant
• OCGT/peaking station usage planning
Transmission & • The economic justification for design requirements and CoUE
distribution investment spending and extension of the transmission
and distribution network.
Operations & • Generation plant service scheduling CoLS or CoUE

maintenance • Inventory management (coal stockpiling)
• Planned load shedding scheduling
• Emergency load shedding
• Service restoration sequence planning after outages
Tariff design • Time of use tariff pricing design CoUE or CoLS

• Design of curtailable and interruptible tariffs
Demand management • Assessing the relative costs of various measures to improve CoLS
and energy efficiency energy-efficiency – e.g. tariff design, demand-side
strategy management programmes.
Source: Nova Economics analysis adapted from Munasinghe and Sanghui (1988).
2.4. The costs associated with power outages
Consumers can be affected by outages both directly and indirectly. The overall impact differs based
on factors, such as the context, timing, duration, and frequency of the interruptions and consumers’
reliance on electricity (Table 3). Electricity consumers incur direct costs an outage directly impacts
activities and processes. 25
The most prevalent direct cost to businesses, for example, is lost production and the accompanying
opportunity cost of idle resources. Businesses may also incur additional shutdown and restart costs
because of outages or interrupted electricity supply could cause equipment failure, inventory
damage or spoilage, data loss or corruption, etc. 26, 27
25
Adam. Rose, Shu-Yi. Liao, and Gbadebo. Oladosu, "Business Interruption Impacts of a Terrorist Attach on the Electric
Power System of Los Angeles: Customer Resilience to a Total Blackout," Risk Analysis 27, no. 3 (2007),
https://doi.org/10.1111/j.1539-6924.2007.00912.x.
26
Munasinghe and Sanghui, "Reliability of Electricity Supply, Outage Cost and Value of Service: An Overview."
27
de Nooij, Koopmans, and Bijvoet, "The value of supply security - the costs of power interruptionsL Economic input for
damage reduction and investment in networks.."

Indirect costs are incurred as the initial impact of an outage permeates through the value chain
(e.g. delayed delivery of input materials, services, or products). For example, the shutdown of
Factory A may reduce its supplies to businesses B and C, which in turn may be forced to reduce
their production due to unavailability of necessary inputs. 28
Indirect costs also include the lagged impacts of an outage, or series of outages on the economy,
due to a loss of business confidence and investor sentiment over the long-term (Table 3). Our
estimates of the CoLS include direct and indirect costs incurred over the short-to-medium term.
Our estimates do not include the longer-term indirect cost to economic growth due to a loss of
investor confidence. These longer-term costs would decrease the long-run trend GDP growth, but
it is not possible to estimate what the trend in growth might have been if load shedding had not
occurred. The persistent electricity supply constraints and load shedding are among a host of
factors have contributed to the gradual decline in South Africa’s GDP growth over the past decade.
In the context of a persistent electricity supply shortage, consumers may experience a series of
recurring outages. In these circumstances, consumers are more likely to employ measures or tactics
to reduce the impact of outages. These measures are usually associated with costs, but consumers
are willing to bear the costs if the future stream of benefits will outweigh the cost – i.e. if a business
invests in a diesel generator, they expect the benefit of having a backup electricity supply to offset
the capital and operating costs of the generator.
Table 3 Factors that impact the cost of electricity outages
Damages and costs

Short-to-medium term Longer-term costs
Direct costs Indirect costs Indirect costs
• Lost production • Cost to customers or impacted • Negative impact on consumer,
• Opportunity cost of idle resources firms (e.g. delayed delivery of business, and investor sentiment.
• Shutdown and restart costs inputs, services, or final goods) • Loss of domestic and foreign
• Spoilage and damages • Cost to suppliers of impacted investment over the medium- to
• Inconvenience, nuisance, and stress firms (e.g. delayed orders or long-term
• Lost leisure purchases of inputs)
Resilience measures and tactics

Inherent Adaptive
• Energy conservation and efficiency of operations • Temporarily adopting alternative processes
• Ability to spread production over multiple facilities • Substituting inputs
• Recapturing lost production at a later stage • Shifting production to unaffected facilities
• Failsafe equipment that allows proper shutdown • Recapturing lost production once the electricity
procedures to take place supply is restored
• Back-up generation and storage
Source: Nova Economics analysis based on Munasinghe and Sanghui, 1988.
28
Kingsley Oladipo. Akpeji, "Cost of Electricity Interruption to Commercial and Industrial End-Users" (Master of Science in
Electrical Engineering University of Cape Town, 2019).

Rose et. al 29 distinguish between two different kinds of resilience – inherent and adaptive (Table 3).
They note that inherent resilience can be defined as the “ability under normal circumstances (e.g.,
the ability of individual firms to substitute other inputs for those curtailed by an external shock [such
as power outage], or the ability of markets to reallocate resources in response to price signals.”
A firm that is inherently resilient to power outages may perform processes or activities that are not
reliant on electricity, it may having the ability to switch to alternative fuel sources or for example to
shift production without affecting overall output. 30 Adaptive resilience is defined as the ability to
reduce the impact of outages due to ingenuity or extra effort (e.g., increasing input substitution
possibilities in individual business operations. A firm, for example, may temporarily adopt alternative
processes, substitute inputs, shift production to unaffected facilities, or recapture lost production
once electricity supply is restored. 31, 32
29
Rose, Liao, and Oladosu, "Business Interruption Impacts of a Terrorist Attach on the Electric Power System of Los
Angeles: Customer Resilience to a Total Blackout."
30
Ian Sue Wing and Adam Rose, "III. Economic Consequence Analysis of Electric Power Infrastructure Disruptions: An
Analytical General Equilibrium Approach," Frontiers in the Economics of Widespread, Long-Duration Power Interruptions:
Proceedings (2019).
31
Rose, Liao, and Oladosu, "Business Interruption Impacts of a Terrorist Attach on the Electric Power System of Los
Angeles: Customer Resilience to a Total Blackout."
32
Munasinghe and Sanghui, "Reliability of Electricity Supply, Outage Cost and Value of Service: An Overview."

3. Estimating the magnitude of load shedding
3.1. Introduction
The first step in estimating the cost of load shedding on the South African economy was to collect
reliable data on the frequency and magnitude of load shedding
Eskom has kept a detailed daily record of the date, time, and duration of load shedding events –
at both a national and at a substation level. Since the national power crisis emerged in 2007, there
have been three distinct periods of load shedding – October 2007 to April 2008, November 2013
to October 2015, and June 2018 to September 2020 (Figure 3).
While Eskom keeps a detailed record of when load shedding occurred, it is not possible to directly
observe how much electricity would have been consumed in the absence of outages. As a result,
while the frequency of load shedding is observable, the ‘magnitude of load shedding’ can only be
estimated.
Figure 3 Load shedding incidences by quarter, 2007-2020
Period 1 Period 2 Period 3
2007 2013 2015 2019

Q4 Q4 Q1 to Q4 Q1 & Q4
2008 2014 2018 2020

Q1 & Q2 Q1, Q2 & Q4 Q2 to Q4 Q1 to Q3
3.2. Two approaches to estimating the magnitude of load shedding
We have compared two estimates of the magnitude of load shedding. The first estimate was
provided by Eskom’s system operator (the “top-down estimate”) and was derived by taking the
difference between their national day-ahead forecast of electricity demand over 24 hours and
actual demand during the hours when load shedding occurred (Table 4).
We derived the second estimate (the “bottom-up estimate”) from data on the incidence and
duration of load shedding events by substation, and from data on curtailment by Eskom’s top
customers. We provide a brief explanation of the approach taken to deriving these estimates in
Section 3.2.1 and 3.2.2. The key differences between the two estimates are summarised in Table 4.
Novꭤ Economics Estimating the magnitude of load shedding 14

Table 4 Key differences in approach and data sources
Top-down estimate Bottom-up estimate

Data sources Eskom System-operator Distribution division
Weekly data provided by the • Monthly electricity sales by substation
system operator on the • Incidences of load shedding by substation
frequency and their estimates Eskom top-customer division
of the magnitude of load • Hourly sales to top customers
shedding. • Incidences of load curtailment by top customers
System operator
• Monthly data on Eskom power buybacks from top
customers
Summary of The system operator estimated • We derived the bottom-up estimate of loading by
approach to the magnitude of load aggregating monthly data on the incidence & duration of
shedding by calculating the load shedding events by substation.
estimating LS
difference between its one day- • We derived the magnitude of load lost during each
ahead forecasts of the national episode using monthly average electricity sales by
demand profile (without load substation.
shedding) and actual demand • We then added data on curtailment and power buybacks
during load shedding. by Eskom's top customers, based on data containing
incidences of curtailments to top customers and detailed
electricity sales data to these customers.
Source: Nova Economics analysis based on Eskom sales data
The Eskom system operator produces an estimate of the magnitude of load shedding by taking the
difference between its daily forecasts of the national electricity demand profile (by hour) and actual
electricity demand during the hours when load shedding occurred. The system operators’ top-
down estimate of load reduction is calculated as the difference between its forecast of national
residual demand and the actual demand on the days when load shedding took place. It is also
adjusted for the slight forecast error observed before and after load shedding took place. 33 Further
detail on the system operator’s approach to estimating the magnitude of load shedding is provided
in Appendix A.
We derived a second estimate (the “bottom-up estimate”) from data on the incidence and duration
of load shedding events by substation, and from data on curtailment by Eskom’s top customers.
33
Residual demand is defined by the Eskom system operator as “The portion of the demand that is supplied by
dispatchable resources”. This includes power sent out by Eskom onto the transmission (Tx) network; international imports;
generation from dispatchable independent power producers (IPPs); and interruption of supply (IoS). IoS in turn refers to all
contracted and mandatory demand reductions to maintain system frequency and security of supply including power
buybacks and curtailment. It excludes demand contracted from IPPs.

The bottom-up estimate is based on data drawn from two separate sources:
• Data on curtailment by Eskom’s top customers 34 (largest mining and industrial

customers) supplied by Eskom’s top customer division within distribution. This includes:
o Sales to top customers (in kWh) on an hourly basis from 2014 to 2019.
o Curtailment by top customers (including date, start time and end time of all
curtailment incidents)
o Total power buybacks (in kWh) from top customers for the months during which
load shedding occurred. This data was supplied by the system operator.
• Data on monthly electricity sales and the incidence and duration of load shedding
supplied by Eskom’s distribution division. 35
o Data indicating the duration (in minutes) and time (specified to the minute) of
electricity supply interruptions (i.e. not exclusively load shedding, but all types of
interruptions) across between 2007 – 2019 by overhead line.
o Monthly electricity sales data indicating average monthly sales data on by
substation from the Feeder Balancing Module (FBM) between 2012 -2019. There
were 1 363 substations in this dataset.
For Eskom’s regular customers (excl. top customers), we obtained data on the monthly incidence
and duration of electricity supply interruptions 36 at an overhead line level. We also obtained a
dataset which contained monthly electricity sales for each of Eskom’s 1 363 substations. Based on
the sales data, we calculated the average amount of electricity sold via each substation in the
absence of load shedding. After matching the overhead lines to substations, we estimated the
magnitude of load lost per substation by multiplying the duration of interruptions (at an overhead
line level) by the average electricity sales, at a substation level, for each month when load shedding
was experienced. This allowed us to derive the magnitude of load shedding for all Eskom’s regular
customers.
We then separately estimated the magnitude of load reduction by Eskom’s top customers who
have either entered power buyback agreements or are subject to voluntary load curtailment. We
followed a similar approach to that used for regular customers and estimated the magnitude of
34
Eskom defines top customers as customers who consume more than 100 GWh per annum. Given the energy intensive
nature of these customers’ operations, top customers often enter into electricity supply agreements with Eskom which
specify that top customers are not subject to regular load shedding, but rather electricity curtailments or power buybacks.
35
We understand that a single substation may be supplied by multiple transmission lines.
36
This dataset contained all incidences of power interruptions, which includes incidences of load shedding as well as other
supply interruptions, such as interruptions due to maintenance or cable theft, for example. We found that other supply
interruptions, in minutes, comprised less than 0.3% of the total duration of interruptions. Due to the trivial contribution of
other supply interruptions to total supply interruptions, we did not adjust the dataset to exclude these other interruptions.

curtailment by comparing sales on ‘an ordinary day’ for each day, month, and year to a very similar
day when curtailment had occurred (e.g. A typical Sunday or Monday in December 2015).
We used two data sets to estimate curtailment by top customers - the first was detailed hourly sales
data for top customers and the second was the incidence of curtailment including the date, start
and end times of each incident. Top customers who entered power buyback agreements with
Eskom are exempt from voluntary curtailment under stage 1 and 2 load shedding. So to complete
our bottom-up estimate of load shedding we added the magnitude of power buybacks from top
customers to the bottom-up estimate of curtailment on the on days where load shedding occurred
(Figure 4).
The composition of the bottom-up estimate of load shedding is presented in Figure 4. This
breakdown reveals that Eskom’s top customers (which account for approximately 40% of total
electricity consumption) contributed about 10% of the total national load reduction during periods
of load shedding. Eskom’s regular customers (which account for the remaining 60% of
consumption) carried a higher proportion of the total burden of load shedding (about 90% of total
load reduction).

Figure 4 Composition of the bottom-up estimate of load shedding
Load shedding (GWh)
Oct Nov Dec Jan Feb Mar Apr Nov Feb Mar Jun Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Jun Jul Nov Dec Feb Mar Oct Nov Dec Jan
2007 2008 2013 2014 2015 2018 2019 2020
Buybacks 0 0 0 0 0 0 0 98 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 53 25 0
Curtailment 2 1 3 21 1 3 17 0 0 4 0 0 0 7 29 1 33 29 3 13 2 0 2 6 0 6 4 18 3 0 6 0 0
Bottom-up estimate 27 17 51 351 18 45 296 0 0 27 7 85 107 41 233 39 193 296 171 159 41 4 0 24 2 34 141 165 432 83 11 381 161
Source: Nova Economics analysis

3.3. Comparison of the two estimates of load shedding
The two approaches to estimating the magnitude of load shedding yielded very similar results, both
in terms of magnitude and distribution of the incidents over time (Figure 5). The fact that there is
very little difference between the top-down and our bottom-up estimates gives us confidence
inaccuracy of the load shedding series. On aggregate, the two estimates differ by just one per cent
over the assessment period.
The series shows that load shedding occurred in a total of 33 months between 2007 and 2019. The
most severe period was from 2013 to 2015 when load shedding occurred in 16 months. The most
severe month of load shedding was in March 2019, when our estimates suggest that between
420 GWh and 435 GWh of electricity demand could not be met.
It was initially our intention to use both load shedding estimates in our analysis to produce a range
of costs. Be when we compared the explanatory power of each estimate in our regression analysis,
we found the difference in the estimated parameters was negligible. As such, we have based our
analysis of the economic costs of load shedding on the top-down, system operator’ estimate alone.
The top-down estimate is easier to derive and Eskom’s system operator calculates it regularly for
various applications.

Figure 5 Comparison of the two estimates of the magnitude of load shedding
Bottom-up estimate Top-down estimate
400
Load shedding (GWh)
300
200
100
0
Oct Nov Dec Jan Feb Mar Apr Nov Feb Mar Jun Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Jun Jul Nov Dec Feb Mar Oct Nov Dec Jan
2007 2008 2013 2014 2015 2018 2019 2020
Source: Nova Economics analysis based on data provided by Eskom

3.4. Load shedding as a percentage of total electricity sales
Load shedding has caused significant disruption in the daily lives of South Africans and as such, we
were surprised to find that load shedding constituted a relatively small proportion of total electricity
sales (Figure 6). The estimates of the magnitude of load shedding never exceeded two per cent of
the total electricity consumed in any of the 52 quarters between 2007 and 2019. This implies that
Eskom was able to meet an average of 98.4% of total electricity demand during every quarter when
load shedding occurred.
Figure 6 Load shedding as a percentage of sales, quarterly
1.6%
Load shedding as %
1.2%
of electricty sales
1.0%
0.8%
0.7% 0.6%
0.5% 0.5%
0.4%
0.2%
0.1% 0.1% 0.0% 0.0% 0.0% 0.0%
Q4 Q1 Q2 Q4 Q1 Q2 Q4 Q1 Q2 Q3 Q4 Q2 Q3 Q4 Q1 Q4
2007 2008 2013 2014 2015 2018 2019
We also calculated the proportion of load shedding to total electricity sales every month; we found
that load shedding did not exceed three per cent of total electricity demanded in any of the 33
months in which load shedding occurred (Figure 7).
Figure 7 Load shedding as a percentage of sales, monthly
3%
Load shedding as %
of electricty sales
2%
1%
0%
Jun
Jun
Jun
Jul
Aug
Jul
Dec
Dec
Jan
Dec
Jan
Nov
Mar
Nov
Dec
Mar
Mar
Mar
May
Nov
Oct
Oct
Nov
Nov
Oct
Apr
Sep
Apr
Feb
Feb
Feb
Feb
2007 2008 2013 2014 2015 2018 2019

4. Methods used to estimate the cost of load shedding
4.1. Broad analytical approaches that can be taken to estimating the cost of
load shedding
The international literature on the cost of regular, planned outages is limited. Most of the economic
literature on electricity supply interruptions focuses on partial equilibrium (PE) analysis of the cost
of unplanned outages (CoUE). Wing and Rose 37 note that, apart from a few case studies of actual
events, the economy-wide losses that result from regular, planned load shedding (or even to some
extent unplanned outages) were not typically analysed until the 1990s. We did, however, identify 13
studies that had to some extent explored the economic costs of regular, planned outages. A
summary of the studies reviewed is provided in Appendix B.
Based on a review of this literature we noted that studies that estimate the cost of planned outages
will typically follow one of two broad analytical approaches which we have categorised as forward-
looking or retrospective (Figure 8).
Studies that adopt a forward-looking approach will typically simulate the potential impact of load
shedding on the economy under various hypothetical scenarios. The scenarios are usually simulated
using a class of empirical economic models known as computable general equilibrium (CGE)
models. CGE models fit economic data describing the microeconomic structure of an economy to
a set of non-linear simultaneous equations that describe the theoretical interaction between firms,
households, government, and the rest of the world.
The potential impact of load shedding on a specific sector or the national economy as a whole can
be simulated using a CGE by constructing and applying a shock – e.g. by simulating a once-off 10%
reduction in electricity supply – which in practice can be applied either by reducing the output of
the electricity sector directly or by rendering a portion of each industry’s capital stock idle.
Studies that follow a ‘retrospective’ approach, aim to ascertain the impact that load shedding has
historically had on economic growth in a particular sector, region, country, or group of countries.
These studies employ various methods to try and ascertain the impact that regular planned outages
have had on economic output during a particular event or series of events.
The retrospective studies were based on one of three methods – case studies of a particular event,
revealed preference surveys or the statistical/econometric analysis of historic data on outages.
Econometric models are statistical models that allow for the empirical verification of economic
37
Wing and Rose, "III. Economic Consequence Analysis of Electric Power Infrastructure Disruptions: An Analytical General
Equilibrium Approach."
Novꭤ Economics Methods used to estimate the cost of load shedding 22

theory or hypotheses. For example, examining whether the load shedding events have had a
negative impact on GDP growth when controlling for the influence of other factors. We have
provided a brief overview of what an econometric study entails in ‘Box 1. What is econometrics?’
Figure 8 Analytical approaches and empirical methods used to estimate the cost of load shedding
Analytical approaches to
estimating the CoLS
Forward-looking Retrospective
• Typically simulate the potential impact of load • Aim to ascertain the impact load shedding has
shedding under hypothetical scenarios. historically had on economic growth in a
• Some studies simulate adaptative responses particular sector, region, country or group of
including mitigation (e.g. back-up generator) & countries.
inherent resilience (e.g. delaying processes or • Attempt to isolate the marginal contribution of
finding substitutes). load shedding to the decline in economic output
during a particular event or series of events.
Empirical methods used Empirical methods used
• Scenarios typically generated using a class of • Retrospective studies are based on one of three
empirical economic models know as computable methods – case studies of a particular event,
general equilibrium (CGE) models. CGE models data collected in revealed preference surveys or
fit economic data describing the microeconomic statistical/econometric analysis of historical
structure of an economy to a set of non-linear data on outages.
simultaneous equations that describe the
theoretical interaction between firms, • Econometric models are statistical models that
households, government and the rest of the allow for the empirical verification of economic
world. theory or hypotheses – for example testing
whether load shedding events have had negative
• The potential impact of load shedding on a
impact on GDP growth. Some of the econometric
sector or national/regional economy can be
techniques that have been used in previous
simulated using a CGE by constructing and
studies to isolate the impact of load shedding on
applying a shock – e.g. by simulating a one-off
economic growth classical linear regression
10% reduction in electricity supply – either by
analysis, distributed-lag models, panel data
reducing the output of the electricity sector
regression analysis and vector auto-regression.
directly or making 10% of capital stock idle.

Box 1. What is econometrics?
Econometrics is the study concerned with the empirical verification of the economic theory.
Econometrics involves the application of a range of statistical techniques to economic data to
empirically verify a hypothesis – for example, to test the extent to which load shedding had a
negative impact on GDP growth.
In general, an econometric study will proceed along the following lines.
1. Statement of theory or hypothesis – e.g. load shedding subtracted from GDP growth.
2. Specification of the mathematical model of the theory – e.g. GDP can be expressed in terms
of an energy augmented Cobb-Douglas production function:
Equation 1 𝑌𝑌𝑖𝑖𝑖𝑖 = 𝐴𝐴𝑖𝑖𝑖𝑖 𝐾𝐾𝑖𝑖𝑖𝑖𝛼𝛼 𝐿𝐿𝛿𝛿𝑖𝑖𝑖𝑖 𝐸𝐸𝑖𝑖𝑖𝑖𝜃𝜃
3. Specification of the statistical/econometric model – while mathematical models assume an

exact or deterministic relationship between the variables and an econometric or statistical
model allows for an inexact relationship between variables and always includes an error term
or residual 𝜀𝜀𝑡𝑡 . The econometrician also transforms variables so that the parameters can be
estimated with the chosen statistical technique. For example, we take the natural logarithm
in the above production function equation so that we can estimate the equation using the
most well-used statistical estimation method – the classic linear regression model:
Equation 2 ln(𝑌𝑌𝑖𝑖𝑖𝑖 ) = 𝐴𝐴𝑖𝑖𝑖𝑖 + 𝛼𝛼 ln(𝐾𝐾𝑖𝑖𝑖𝑖 ) + 𝛿𝛿𝛿𝛿𝛿𝛿(𝐿𝐿𝑖𝑖𝑖𝑖 )+ 𝜃𝜃 ln(𝐸𝐸𝑖𝑖𝑖𝑖 ) + 𝜀𝜀𝑡𝑡 .
A simple two-variable linear regression between q/q GDP growth and load shedding as %
of sales is depicted in the scatterplot, which serves as a visual depiction of a simple linear
regression model, below. The line is fitted using an estimation procedure like OLS, which fits
the line by minimising the squared residuals 𝜀𝜀𝑡𝑡 .
4. Sourcing the data.
5. Estimation of the parameters – the statistical technique of regression analysis is the main tool
used to estimate the parameters. The model suggests that load shedding is negatively
correlated with GDP but that on its own, it only explains 23% of the variation, which means
that we should include other determinants of GDP growth in our model to improve the fit
and explanatory power of the model.
6. Hypothesis testing – run tests to see if the estimated parameter on load shedding (-0.7) is
statistically different from zero (i.e. load shedding does not influence GDP), with the t-test, f-
test, etc.
7. Application of the model and estimated parameters for insight, forecasting, or policy
analysis.

Figure 9 Scatterplot of economic growth (GDP q/q % change) against load shedding (as % of sales)
2.0
y = -0.5739x + 0.6534
R² = 0.1893
y: GDP change q/q (%)
1.0
0.0
-1.0
-2.0
0 0.5 1 1.5
x: Load shedding as % of elec. sales
Source: Nova Economics analysis based on Gujurati, Damodar N. Basic Econometrics
4.2. Previous estimates of the cost of load shedding in South Africa
We identified three previous studies that had attempted to estimate the broader costs of load
shedding in South Africa: Deloitte, 38 Goldberg 39 and Andersen and Dalgaard (Appendix B, App
Table 1). 40 The first two studies focused on South Africa while the latter (Andersen and Dalgaard)
assessed the impact of electricity shortages across 39 Sub-Saharan African countries, including
South Africa.
In 2009, Eskom commissioned Deloitte to model the economic cost and impacts of the regular
planned outages that occurred in 2007/8. This is one of the few studies that has attempted to
measure the cost of load shedding, rather than estimating the cost of infrequent, unplanned
outages (the CoUE) in South Africa. The key finding of the study was that load shedding was
estimated, under the most realistic scenarios, to cost the economy R4.90/kWh in 2008 prices which
would equate to ~R8.95/kWh in 2020 prices. This is much lower than the estimate produced that
year of the CoUE infrequent unplanned outages of R75/kWh (PB Power, 2008).
38
39
Ariel Goldberg, "The economic impact of load shedding: The case of South African retailers" (University of Pretoria,
2016).
40
Thomas Barnebeck Andersen and Carl-Johan Dalgaard, "Power outages and economic growth in Africa," Energy
Economics 38 (2013).

The Deloitte study adopted a forward-looking approach, the cost of
load shedding was simulated using a dynamic computable general
equilibrium model (DCGE). The authors considered three different
scenarios to model the potential impact of load shedding. In the most
realistic scenario, they cut the capital stock of each of the 40 sectors of
the varying amounts (2%, 5% or 10%) based on the electricity intensity
of the sector. This was based on the hypothesis that more electricity-
intensive sectors are likely to have more equipment or capital stock
rendered idle during a load shedding event with less ability to adapt or
find alternatives.
A limitation of the study is that they simply assume that the load
shedding event was equivalent to a load reduction of 3 000 MW and
that this would result in a decline of between 2% and 10% in the capital
stock of various sectors. This assumption was untested and was not
calibrated to an actual historical incident or incidents.
Goldberg 41 also attempted to estimate the cost of load shedding in

South Africa, but the study was limited to assessing the impact on a
sample of retailers in Gauteng. The study was based on semi-structured
interviews with retail managers and stated-preference and revealed-
preference survey techniques. The surveys made use of online and in-
store questionnaires to capture stated preference or willingness of retail
outlets to pay to avoid load shedding. The revealed preference
techniques were used on actual financial information that was collected
from retail head offices on the cost of providing back-up generation.
These three methods enabled the author to derive different estimates
on the cost of load shedding for retailers for different times of the day
based on a customer damage function approach.
The biggest limitation of this study is the small sample size. The semi-
structured interview insights were based on a sample of only eight
interviews, while the sample sizes for the state-preference and
revealed-preference methods were also limited (106 and 42 surveys
respectively).
Andersen and Dalgaard follow a retrospective econometric approach

to estimating the cost of power outages across 39 Sub-Saharan African
countries including South Africa. The authors estimate the total effect
of power outages on economic growth in Sub-Saharan Africa over the
period 1995-2007.
41
Goldberg, "The economic impact of load shedding: The case of South African retailers".

While the authors attempt to pay close attention both to potential errors of measurement of African
economic growth and to the endogeneity of outages, the study suffers from some limitations.
Firstly, it makes use of a very parsimonious model of outages on economic growth – the authors
explain economic growth only in terms of the GDP growth per capita, an intercept, the log of
outages and an error term. In trying to limit the risk of multicollinearity and endogeneity, their
model is likely to suffer from omitted variable bias. As a result, too much of the reported variation
in GDP per capita was likely attributed to outages.
4.3. Previous econometric studies of the cost of load shedding
Some of the specific econometric techniques that have been used in previous studies to isolate the
impact of load shedding on economic growth are classical linear regression analysis, autoregressive
distributed lag (ARDL) models, panel data regression analysis and vector auto-regression (VAR)
models.
Of the 13 studies we reviewed, four were based on an econometric analysis of the cost of historical
electricity supply shortages. Fisher-Vanden, Mansur, and Wang, 42 and Allcott et al. 43 estimated the
impact of electricity shortages based on panel data for manufacturing firms in China and India,
respectively. Andersen and Dalgaard, 44 as discussed earlier, estimated the impact of electricity
shortages on GDP per capita across Sub-Saharan Africa using time series techniques, while Ellahi 45
estimated the impact of electricity supply constraints on the development of the industrial sector in
Pakistan using an ARDL model. A more comprehensive review and summary of the results,
advantages and limitation of these studies can be found in Appendix B, Section B.3.
One of the main advantages of studies based on econometric analysis of historical time series or
panel data is that, in contrast to highly stylised and theoretical GE models, the results are based on
the empirical analysis of real-world outage events. Studies based on econometric analysis typically
aim to isolate the impact of a particular load shedding event (or series of events) on economic
growth or industry-level production. In this sense, they can be used to produce estimates of the
actual historical economic costs (direct and indirect and net of adaptive response) of regular or
persistent power shortages. For example, Andersen and Dalgaard 46 estimated across a sample of
39 Sub-Saharan African countries that a one per cent increase in the number of outages reduced
long-run GDP per capita by 2.86 per cent during the period 1997 to 2007.
42
Karen Fisher-Vanden, Erin T Mansur, and Qiong Juliana Wang, "Electricity shortages and firm productivity: evidence
from China's industrial firms," Journal of Development Economics 114 (2015).
43
Hunt Allcott, Allan Collard-Wexler, and Stephen D O'Connell, "How do electricity shortages affect industry? Evidence
from India," American Economic Review 106, no. 3 (2016).
44
Andersen and Dalgaard, "Power outages and economic growth in Africa."
45
Nazima Ellahi, "Testing the relationship between electricity supply, development of industrial sector and economic
growth: An empirical analysis using time series data for Pakistan," International Journal of Management Science and
Engineering Management 6, no. 4 (2011).
46

Another advantage of retrospective econometric studies it that is possible to estimate the net
impact of power shortages on different groups or sectors – to understand which firms or sectors
are more resilient (able to avoid or cushion the impact) than others. For example, Allcott et al. 47
found that in 2005 when Indian manufacturing firms faced outages, 7.1% of the time, that the output
of firms that were able to self-generate (i.e. had back-up generators) lost only 0.7% of their output.
Firms that did not have back-up generation lost 10.3% of their output. The authors concluded that
electricity shortages were a substantial drag on Indian manufacturing from 1992 to 2010, reducing
manufacturing output by an average of about five per cent over the period.
4.4. Our approach to estimating the CoLS
We estimated the historical impact of load shedding on GDP growth from 2007-2019. We have
produced three alternative sets of estimates of the CoLS, based on three different econometric
techniques – a classical linear regression, an auto-regressive distributed lag model and a panel data
regression. The first two estimates are based on an expenditure-side model of the determinants of
GDP while the third was based on an energy-augmented Cobb-Douglas production function
(Figure 10).
Figure 10 Approach to estimating CoLS
Our econometric approach - estimating the CoLS
1. Expenditure-side model of GDP 2. Energy-augmented production function

Theoretical basis
• The expenditure-side models of the determinants of • We also estimated the impact of load shedding on GDP
GDP are based on the national income identity where using a Cobb-Douglas production function. In terms of
GDP (Y) is function of consumption expenditure (C) by this standard theory of economic production, industries
firms and households, government spending (G), combine inputs including capital (K), labour (L) and
investment (I) , plus exports (X) less imports (M) . Hicks-neutral technology (A) to produce output (Y) or
gross value-added (GVA).
1a. Linear Regression 1b. ARDL 2a. A Panel data regression, OLS with fixed effects
Econometric methods
We produced the first set For the second set of We produced the third and final set of estimates of the
of estimates of the CoLS estimates of the CoLS CoLS by estimating the energy-augmented Cobb-Douglas
using a classical linear based on the expenditure- production function in logarithmic form. We estimate the
regression to estimate the side model of GDP we use parameters as a system of equations using a panel
expenditure-side model of an Auto-regressive regression technique known as the Fixed Effect Least-
GDP. distributed-lag technique. Squares Dummy Variable (LSDV) Model.
47
Allcott, Collard-Wexler, and O'Connell, "How do electricity shortages affect industry? Evidence from India."

4.5. Estimation of the CoLS based on the expenditure-side models of GDP
The expenditure-side econometric model of the determinants of GDP is based on the national
income identity presented in Equation 3. GDP (Y) is a function of consumption expenditure (C) by
firms and households, government spending (G), investment (I), plus exports (X) less imports (M).
To isolate the impact of load shedding on GDP we add a variable to capture the frequency and
magnitude of load shedding (LS) to the standard identity so that:
Equation 3 𝑌𝑌𝑡𝑡 = 𝛽𝛽0 + 𝛽𝛽1 C𝑡𝑡 + 𝛽𝛽2 G𝑡𝑡 + 𝛽𝛽3 I𝑡𝑡 +(𝛽𝛽4 X𝑡𝑡 − 𝛽𝛽5 𝑀𝑀)+𝛽𝛽6 LS𝑡𝑡
where t represents time, 𝛽𝛽0 to 𝛽𝛽6 are the estimated parameters, and 𝛽𝛽6 specifically captures the
impact of load shedding on GDP.
This is the theoretical basis of the econometric models used to produce our first two estimates of
the CoLS.
National GDP aggregates
The expenditure-side econometric models of the determinants of GDP are both based on a
quarterly series of the expenditure on gross domestic product and its sub-components, sourced
from the South Africa Reserve Bank (Table 5). These are expressed in constant 2010 rand and are
seasonally adjusted and annualised values.
Industry-level series on gross-fixed capital formation
To determine the cost of load shedding on GDP at a sector level, we replaced aggregate gross
fixed capital formation (GFCF) with an industry-specific measure of GFCF, but only where it
improved the model fit. Until 2016 the SARB published a quarterly series of GFCF for six sectors
including mining and manufacturing (Table 5). We sourced the discontinued series from the SARB
and interpolated annual series on GFCF (which is still published) to generate quarterly data for the
remainder of the period (2016 to 2019).
Load shedding variable
We based our analysis on the top-down estimate of the magnitude of load shedding derived by
Eskom’s system operator. For the regression analysis, we tested the explanatory power of the load
shedding variable in various forms. The load shedding variable had the most explanatory power
when expressed as a percentage of total quarterly electricity sales. This format shows how
significant the amount of load shedding was in each quarter relative to overall electricity demand
– which is an input to production. Monthly data on total domestic electricity sales was also
supplied by Eskom and aggregated into quarters.

Table 5 Data and variables used for the expenditure-side model of GDP
Variable Source and data code Description

GDP (Y) South African Reserve Expenditure on domestic product (including
Bank (SARB) (KBP6006D) residual), seasonally adjusted and annualised,
Rm, market prices, constant 2010 values.
Quarterly data
Consumption (C) SARB (KBP6007D) Final consumption expenditure by households
Expenditure side GDP aggregates
General Government (G) SARB (KBP6008D) Final consumption expenditure by general

government
Gross Fixed Capital Formation SARB (KBP6009D) Gross fixed capital formation
(I)
Imports (M) SARB (KBP6014D) Imports of goods and services,
Exports (X) SARB (KBP6013D) Exports of goods and services,
Inventories (S) SARB (KBP6010D) Change in inventories
Gross fixed capital formation SARB GFCF manufacturing Quarterly data on GFCF by type of economic
for selected industries (I) (NRI6082D), GFCF activity seasonally adjusted and annualised
Electricity and water rates, Rm, market prices, constant 2010 values.
(NRI6085D). Annual series Values between 2016 and 2020 were
(6082Y) and (6085Y) interpolated from annual series.
Load shedding as a % of total Derived from Eskom data Monthly data aggregated into quarters
sales (LS) (see below)
Load shedding data
Load shedding Top-down, Eskom system Monthly data on the national estimate of the
operator magnitude of load shedding in GWh
aggregated into quarters.
Total electricity sales Eskom: SAP Total Monthly national sales data in kWh (excl.
Consumption (kWh) international sales) aggregated into quarters
from 2006 until 2019
Dummy variables (D) 2008/9 financial crisis, Dummy variables were included for drought in
drought, rains, credit boom 2Q06, 3Q06, 1Q18, 2Q18, rainy season 2017,
and oil-price shock global financial crisis 2008/9 and oil-price
shock 2007.
Our initial hypothesis was that load shedding was likely to have had a noticeable negative impact
on GDP growth in the 16 quarters when it historically occurred. Since the impact of load shedding
is likely to be fairly immediate, we felt it would probably be more evident in the quarter-on-quarter
growth in GDP (as opposed to in level or year-on-year change).
To test this hypothesis, we began with a simple visual inspection of the relationship between
historical load shedding episodes and quarter-on-quarter GDP growth (Figure 11). We note while
the latter two periods of load shedding were associated with negative GDP growth, quarter-on-
quarter growth remained positive in the first period (2007/8). Viewed against the linear trend in
quarter-on-quarter GDP growth, growth was consistently below trend when load shedding

occurred. It also appears that more severe episodes of load shedding were associated with a bigger
decline in growth and that the impact occurs within the same quarter – it is most evident in the
quarters when load shedding occurred. In other words, the impact is not significantly delayed.
Figure 11 Change in GDP against load shedding as a percentage of total electricity sales
2 2
Load shedding as a % of sales

1 1
GDP growth q/q (%)
0 0
-1 -1
-2 -2
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Load shedding as % of sales GDP growth Linear (GDP growth)
This suggested that we should find a negative correlation between historical load shedding
episodes and GDP growth if we estimated the relationship econometrically. It also appears, based
on visual inspection that load shedding had a noticeable impact on GDP growth in the 16 quarters
when it occurred.
Transformation of variables for the classic linear regression model (CLRM)
To adhere to the assumptions of the classic linear regression model, the GDP aggregates which all
trend with time need to be transformed into a stationary series that are mean reverting. As such,
we express all the GDP aggregates in quarter-on-quarter changes, which as explained above is
likely to produce a better result than year-on-year changes which are influence by base effects. The
load shedding variable was expressed as a percentage of total electricity sales.
We estimated 11 separate single equation regression, one for total GDP and then a regression for
each of the ten individual sectors of the economy, as reported in the system of national accounts.
For each regression, we only include the subcomponents of GDP that were statistically significant
and improved the overall model fit.

The reserve bank does not publish the sub-components of expenditure on GDP at an industry level,
except for GFCF, so most industry-level equations include national consumption aggregates. The
specification of each equation is summarised in Appendix D, App Table 6. The final regression
equation to estimate the impact of load shedding on total GDP growth was specified as follows:
Equation 4 𝑌𝑌𝑡𝑡 = 𝛽𝛽0 + 𝛽𝛽1 C𝑡𝑡 + 𝛽𝛽2 G𝑡𝑡 + 𝛽𝛽3 I𝑡𝑡 +𝛽𝛽4 LS𝑡𝑡 +𝛽𝛽5 D0,𝑡𝑡 +𝜀𝜀𝑡𝑡
where:
𝑌𝑌𝑡𝑡 = Quarter-on-quarter % change in GDP
𝛽𝛽𝑡𝑡 = Constant
C𝑡𝑡 = Quarter-on-quarter % change in final consumption expenditure
I = Quarter-on-quarter % change in gross fixed capital formation
LS = Load shedding in GWh as a percentage of total electricity sales
D0 𝑡𝑡 = Dummy variable for the 2008/09 financial crisis dummy
𝜀𝜀𝑡𝑡 = Error term (residual)
Where the explanatory variables in a regression model are non-stationary (i.e. trending over time),
and the interest is in understanding the long-run relationship between them, it more appropriate
to use specialist time series techniques, such as the ARDL to estimate the coefficients than a classic
linear regression. The reason for this is that in order to be able to estimate them with a CLRM, one
has to difference the series to remove the trend (to avoid the problem of spurious regression
between trending series) but in doing so one loses information about the long-term relationship
between the trending series.
Figure 12 Trend in real GDP and load shedding in GWh, 2007 to 2019
3.2 900
Load shedding, quarterly (GWh)
3.1 800
700
GDP (R trillion)
3.0
600
2.9 500
2.8 400
300
2.7
200
2.6 100
2.5 0
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Load shedding (right axis) Real GDP
Source: Nova Economics analysis based on data from SA Reserve Bank and Eskom

Since we are more interested in the immediate short-term relationship between GDP growth and
load shedding and are less concerned about accurately estimating the long-run relationship
between GDP and its subcomponents, we felt the classic linear regression model (CLRM) would
give us the best results but we have presented the results of the ARDL estimation as an alternative.
Based on simple visual inspection of the relationship between historical load shedding episodes
and the trend in real GDP, it is clear that as one would expect, load shedding incidents had little
direct influence on the overall trend in GDP but would rather have had some influence in the
variation in GDP around its long-term trend (Figure 12).
Transformation of variables for the autoregressive distributed lags (ARDL) approach
For the ARDL model, we include GDP and its sub-aggregates which all trend with time and need
to be transformed into stationary series that are mean reverting. As such, we express all the GDP
aggregates in quarter-on-quarter changes, which as explained above is likely to produce a better
result than year-on-year changes which are influence by base effects. For the ARDL model, we
distinguish between dynamic regressors (those that trend with time) and fixed regression including
the dummy variables and load shedding series. We expressed all the dynamic regressors in a
logarithmic form (by taking the natural logarithms of the underlying series) to make the estimated
coefficients easier to interpret.
We expressed load shedding variable as a percentage of total electricity sales and estimated 11
single equation regression - one for each of the ten individual sectors of the economy and another
for total GDP as reported in the system of national accounts. For each regression, we only include
the subcomponents of GDP that were statistically significant and improved the overall model fit.
The reserve bank does not publish the sub-components of expenditure on GDP at an industry level,
except for GFCF, so most industry-level equations include national consumption aggregates. The
specification of each equation is summarised in Appendix D. The final regression equation for total
GDP was estimated using an ARDL model as follows:
Equation 5 ln (𝑌𝑌𝑡𝑡 ) = ln (𝛽𝛽0 ) + 𝛽𝛽1 ln (C𝑡𝑡 )+ 𝛽𝛽2 ln (G𝑡𝑡 ) + 𝛽𝛽3 ln (I𝑡𝑡 ) +𝛽𝛽4 ln (LS𝑡𝑡 ) +𝛽𝛽5 ln (D0 𝑡𝑡 ) +𝜀𝜀𝑡𝑡
where:
𝑌𝑌𝑡𝑡 = Real GDP, Rm, seasonally adjusted and annualised
𝛽𝛽𝑡𝑡 = Constant
C𝑡𝑡 = Final consumption expenditure by households
𝐺𝐺𝑡𝑡 = Final consumption expenditure by government
I = Gross fixed capital formation
LS = Load shedding in GWh as a percentage of total electricity sales
D0 𝑡𝑡 = Dummy variable for the 2008/09 financial crisis dummy
εt = Error term (residual)

4.6. Estimation of the CoLS based on an energy-augmented Cobb-Douglas
production function
Our third estimate of the CoLS was estimated using a panel data regression model based on the
Cobb-Douglas production function (Equation 6). In terms of this standard theory of economic
production, industries combine inputs including capital (K), labour (L), and Hicks-neutral technology
(A) to produce output (Y) or gross domestic product (GDP) (Equation 6). This standard production
function can be augmented to include total national electricity sales (E) since power is also an input
to production.
Equation 6 𝑌𝑌𝑖𝑖𝑖𝑖 = 𝐴𝐴𝑖𝑖𝑖𝑖 𝐾𝐾𝑖𝑖𝑖𝑖𝛼𝛼 𝐿𝐿𝛿𝛿𝑖𝑖𝑖𝑖 𝐸𝐸𝑖𝑖𝑖𝑖𝜃𝜃
where:
𝑌𝑌𝑖𝑖𝑖𝑖 = Real GDP, Rm, seasonally adjusted and annualised
𝐴𝐴𝑖𝑖𝑖𝑖 = Hicks neutral technology
K 𝑖𝑖𝑖𝑖 = Fixed capital stock
𝐿𝐿𝑖𝑖𝑖𝑖 = Labour (formally employed)
𝐸𝐸𝑖𝑖𝑖𝑖 = Total domestic electricity sales (in GWh)
and i represents the different sectors of the economy and t represents time and there are
constant returns to scale so that [α = 1-δ].
We took the natural logarithm of the variables in Equation 7 so that we could estimate the equation
using a panel regression technique known as the Fixed Effect Least-Squares Dummy Variable
(LSDV) Model:
Equation 7 𝑙𝑙𝑙𝑙(𝑌𝑌𝑖𝑖𝑖𝑖 ) = 𝐴𝐴𝑖𝑖𝑖𝑖 + 𝛼𝛼 𝑙𝑙𝑙𝑙(𝐾𝐾𝑖𝑖𝑖𝑖 ) + 𝛿𝛿𝛿𝛿𝛿𝛿(𝐿𝐿𝑖𝑖𝑖𝑖 )+ 𝜃𝜃 𝑙𝑙𝑙𝑙(𝐸𝐸𝑖𝑖𝑖𝑖 ) + 𝜀𝜀𝑡𝑡
Following the approach used by Burger et al., 48 we estimate the Cobb-Douglas production function
as a system and isolate the contribution of load shedding to the change in GDP, rather than
individually estimating single-equations for GDP and each of the ten sectors of the economy. The
systems approach has some advantages over the single equation approach as it allows the
simultaneous identification of factor-augmenting technical change and the contribution of
electricity consumption, capital, and labour to GDP growth across all sectors.
We initially tried to isolate the influence of load shedding on GDP by augmenting the Cobb-Douglas
production function with a single series representing the incidence and magnitude of load at a
national level. We experimented by including it in various forms (in levels, as a dummy variable and
expressed as a percentage of sales), but it was not possible to isolate the immediate and short-
term effect of load shedding on GDP. This is probably because all the other variables included in
the panel were slow-moving aggregates that are expressed in levels and overall load shedding has
48
Friedrich Kreuser, Rulof Burger, and Neil Rankin, "The elasticity of substitution and labour-displacing technical change in
post-apartheid South Africa," UNI-WIDER, 2015/101 (2015).

little bearing on the long-term trend in GDP when compared to the influence of growth in the
capital stock and employment.
We then approached the estimation of load shedding using the panel approach from a different
perspective - and augmented the production function with total electricity sales. Since total
electricity consumption is an input to production and is a slow-moving aggregate, it is possible to
isolate the contribution of electricity to GDP growth within the production function. We then use
the long-term relationship between electricity consumption and GDP to estimate the proportion of
output that is being lost when load shedding occurs.
We illustrate the relationship between the trend in the normalised capital stock, labour, electricity
consumption and GDP series for the Manufacturing industry in Figure 13. If we had attempted to
estimate a single-equation production for the Manufacturing industry based on these series, the
estimated parameters would only be based on the time-series variation between these four series.
Figure 13 Manufacturing industry production function variables, normalised (2003=100)
Labour_Manu Capital_Manu GDP_Manu Elec sales from 2003
135
125
Index (2003 = 100)
115
105
95
85
2000 2002 2004 2006 2008 2010 2012 2014 2016 2018
The trends in the series for this industry suggest that while all three of the explicit inputs to
manufacturing sector GDP has been in trend decline over the past decade, manufacturing GDP
remained relatively stable, suggesting that total factor productivity increased (for example through
gains in energy efficiency and/or labour productivity). By pooling the data and estimating the
coefficients on capital, labour, electricity and productivity across all ten industries, estimated
parameters are based on the much richer variation between industries and across time and are
therefore likely to be far more efficient and reliable.

A summary of the data sources and variables used to estimate the Cobb-Douglas production
function is provided in Table 6.
Table 6 Data sources and variables- Cobb-Douglas production function
Variable Source and data code Description
GDP (Y) South African Reserve Bank • Expenditure on domestic product (including residual),
(SARB) KBP6006D seasonally adjusted and annualised rates, (Rm, market
prices, constant 2010 values. Quarterly data.)
Capital (K): SARB: KBP6140Y-KBP6148Y • Fixed capital stock: seasonally adjusted and
Fixed Capital Stock SARB: KBP6080Y-KBP6088Y annualized, (Rm, market prices, constant 2010 values.)
Gross Fixed Capital SARB: NRI6081/2/5/8D, • Quarterly data on GFCF by type of economic activity
Formation NRI6091/4D seasonally adjusted and annualised rates, (Rm, market
Quarterly gross prices, constant 2010 values).
fixed capital
formation (GFCF)
Labour (L) Post-Apartheid Labour Market • Number of labourers formally employed in each
Series (PALMS) industry.
StatsSA: QLFS and LFS
Total electricity • Eskom distribution monthly sales Derived from a combination of the following data sets:
sales by industry and revenue data reported for • Disaggregated electricity sales and revenue data
each region at 5-digit SIC level (classified in most cases at the 5-digit SIC code) for
or higher, 2006 to 2019 each of Eskom’s regions and its top customers,
• Eskom distribution, historical extracted from Eskom’s billing system for the years
monthly sales and revenue 2006 to 2019.
series, national, 2003 to 2008 • Disaggregated national sales and revenue data
• Aggregate monthly sales data (classified in most cases at the 5-digit SIC code)
reported by major Eskom sourced from Eskom distribution for the years 2003 to
customer category 2006 to 2019. 2008.
• States of SA cities report 2006, • Total monthly national sales data by customer
2011 and 2014/15, breakdown of category, extracted from SAP system. Categories
municipal electricity sales by include IPPs, Agri, Redistributors, Commercial,
industry. Industrial, Int. Sales, Mining, Prepayment, Public lights,
Residential, Traction, external and internal sales.
• Municipal sales by industry as reported in SA Cities
report and by CoUE report for 13 large municipalities
for 2006, 2011, 2014/15.

Gross domestic product
The output series used to estimate the production function is the SARB’s quarterly expenditure
series on real gross domestic product (including residual) at seasonally adjusted and annualised
values.
Capital stock
Following the method used by Burger et al., 49 we derived a quarterly series of capital stock for the
ten industries used in the national accounts based on data sourced from SARB. Further detail on
the method used to derive the series can be found in Appendix C.1. This is included the annual real
fixed capital stock by industry (seasonally adjusted and annualised) and the quarterly gross fixed
capital formation series, by industry where available. The SARB used to publish a quarterly gross
fixed capital formation series for six of the ten industries defined in the national accounts but this
was discontinued in 2015. For the industries and the years where quarterly gross fixed capital
formation (GFCF) was not available, quarterly figures for fixed capital stock were created by
interpolating the values in the annual series using cubic spline and moving average techniques.
In Figure 14 we present the accumulation of capital by industry. As seen, most of the growth in
capital stock over the past decade has been in the electricity, transport, personal and financial
services industries. Capital stock in the mining industry, by contrast, has declined.
Figure 14 Trend in quarterly fixed capital stock, by industry, 2000 to 2019
2.5
Capital stock (Rand, trillion)
1.5
0.5
Fin. Govt. Pers. Transp. Util.
Mining Manuf. Retail Agri. Constrn.
Source: Nova Economics analysis based on data from SA Reserve Bank
49
Kreuser, Burger, and Rankin, "The elasticity of substitution and labour-displacing technical change in post-apartheid
South Africa."

Labour
The labour data used in the study is taken from the Post-Apartheid Labour Market Series (PALMS)
1994-2019Q2 compiled by DataFirst (UCT) from Statistics South Africa’s (StatsSA) Labour Force
Survey (LFS) and Quarterly Labour Force Surveys (QLFS) till Q2 2019, whereafter the QLFS from
StatsSA were used for Q3 and Q4 of 2019.
The LFS is a biannual series that ran from 2000 to 2007. We interpolated the biannual series to
create a quarterly series. From 2008 StatsSA ran a quarterly labour force survey and we use the
employment series from this publication until the end of 2019.
Many papers have investigated the problems in comparing the StatsSA household surveys and
particularly the effect that modifications in questionnaire design and sampling methodology may
have had on the comparability of the household surveys over time. The most serious comparability
problems occur for the informal sector or self-employed workers so the effect of these
inconsistencies can be limited by omitting these workers from the sample and restricting our dataset
to formal sector employees only. 50
After omitting all the unemployed, economically inactive, and self-employed workers from the
dataset, aggregate industry employment per period is calculated from the individual responses to
the questions regarding the industry of employment and the survey weights. We look at this data
at an industry level, using the SIC one-digit categories to group the data into ten industries, and
many of the measurement issues are ameliorated by aggregation.
The ten industries are therefore used as the cross-sectional units of observation for our production
function model. Many papers have reported issues in comparing labour force data over time due
to changes in survey design and sampling methodology. The most serious compatibility issues are
in the series on the informal sector and self-employed workers, so most studies recommend
omitting these categories of workers from the sample so that employment trends reflect only formal
sector employment. 51 We, therefore, omit all the unemployed, economically inactive, and self-
employed workers from the dataset.
We estimated total formal sector employment by industry by aggregating individual responses to

the survey and by applying the relevant survey weights. Employment data was classified into
industries using SIC codes at the 2-digit level, to be comparable to industries as defined in the
South African system of national accounts.
50
South Africa."
51
South Africa."

Electricity sales
To augment the production function with electricity sales, we had to derive electricity sales by
industry (classified according to SIC code at the 2-digit level) so that we could regress this against
GDP (also reported by industry at the 2-digit SIC code level). To derive an electricity sales series by
industry, we made use of the following three sets of data provided by Eskom distribution, extracted
from the Eskom billing system:
1. Highly disaggregated electricity sales and revenue data classified in most cases at the 5-
digit SIC code for each of Eskom’s 6 regions and its top customers, extracted from the
billing system at Eskom distribution for the years 2006 to 2019.
2. Highly disaggregated national sales and revenue data (classified in most cases at the 5-
digit SIC code) sourced from Eskom distribution for the years 2003 to 2008.
3. Aggregated total monthly national sales data by major Eskom customer category,
extracted from SAP system at Eskom distribution from 2006 to 2019.
A breakdown of 13 major municipal redistributors’ electricity sales, as reported in various editions

of the SA State of Cities report, 52 was also used to allocate the ~40% of Eskom’s direct sales that
are sold to municipal distributors’ industries. An explanation of the step-by-step process we
followed to derive electricity sales by industry is presented in Appendix D.
Figure 15 Electricity sales by industry, seasonally adjusted
45
40 Gov.
Pers.
35
Transp.
30
Retail
Electricity sales (TWh)
25
Const.
20 Util.
15 Agri
10 Fin.
5 Mining
Manu.
0
52
Sustainable Energy Africa, State of Energy in South African Cities (2006, 2011, 2015). 3

5. Results
5.1. Introduction
. In the CLRM specification,
isolate the short-to-medium term impact that load shedding

had on GDP. We also, however, discuss the alternative estimates of the CoLS, which based on ARDL
and panel regression methods.
5.2. National estimates of the cost of load shedding
We estimated that the overall cost of load shedding is between R7.61/kWh and R9.53/kWh (in 2020
prices). Our estimates of the CoLS increased steadily over time – from R7.61/kWh during the first
major episode of load shedding (2007-2008) to R9.53/kWh during the third period (2018 – 2019)
(Table 7).
Table 7 Primary estimate of the cost of load shedding, 2007/08 to 2018/19
2007-2008 2013-2015 2018-2019
Cost of load shedding (R/kWh) 7.61 8.80 9.53
Total load reduction (GWh) 872 1 724 1 307
Number of months in which load shedding

7 15 9
occurred (frequency)
Number of months in which load shedding
2 1 2
exceeded 300 GWh (severity)
This was contrary to our expectation that in the context of recurring electricity supply shortages the
cost of load shedding (in kWh) would decrease, as electricity consumers have the opportunity and
incentive to employ measures and tactics to reduce the cost of these disruptions. The increase in
the CoLS over the three periods may be related to the increase in the frequency and/or severity of
load shedding. Eskom estimates that it carried out a total of ~870 GWh of load shedding in the
2008-9 period as compared to ~1 700 GWh during 2013-15 and ~1 307 GWh in 2018-19. The month
in which load shedding was most severe was March 2019.
Novꭤ Economics 40
A detailed analysis of our primary estimates of the CoLS (based on the CLRM technique) show that
load shedding subtracted up to 0.6 percentage points from quarter-on-quarter GDP growth in
South Africa in the quarters when it occurred (Figure 16 and Figure 18). In the third quarter of 2015,
the South African economy came close to entering a technical economic recession, when q/q GDP
(at basic prices) did not expand, after contracting by 0.5% q/q in the second quarter. In the absence
of load shedding, GDP growth would have remained positive on a q/q basis throughout 2015
(Figure 16).
Figure 16 The impact of load shedding on quarterly GDP growth, 2007 to 2019
1
GDP growth q/q (%)
0
-0.1 -0.2 -0.2
-0.3
-0.3 -0.3-0.2 -0.4
-1 -0.6
-2
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP growth GDP growth in the absence of load shedding
We estimate that load shedding cost the South African economy a total of R34.5 billion (in 2020
values) between 2007 and the end of 2019 (Figure 17). In the second quarter of 2015, Eskom
implemented an estimated 809 GWh of load shedding, making it the quarter with the most load
shedding to date (Figure 18 and Figure 19). During the second quarter of 2015, 809 GWh of load
shedding was imposed, which subtracted 0.6 percentage points from q/q GDP growth and cost the
economy R7 billion in GDP.
Figure 17 Impact of load shedding on GDP (rand billion, constant 2020) by period
(R bn, constant 2020
-6.66
prices)
-12.59
-15.25
2007-2008 2013-2015 2018-2019
Novꭤ Economics 41
Figure 18 Impact of load shedding on GDP growth q/q
0.00
-0.03 -0.04
-0.01 -0.02 0.00
Change in q/q GDP attributed to load shedding
-0.08
-0.16
-0.18
-0.20
-0.27 -0.26
(percentage points)
-0.32
-0.42
-0.49
-0.63
2007 2008 2013 2014 2015 2018 2019
Figure 19 Impact of load shedding on GDP (rand billion, constant 2020)
-0.01
-0.30 -0.44
-0.06 -0.21 -0.03
-0.78
(R bn, constant 2020 prices)
-1.82
-2.08
-2.28
-2.68
-2.92
-3.20
GDP
-4.84
-5.69
-7.16
2007 2008 2013 2014 2015 2018 2019
Novꭤ Economics 42
As mentioned, our primary estimates of the CoLS were calculated using a classic linear regression
model (CLRM). We expressed all model variables (except for load shedding and dummy variables)
in first differences (i.e. in quarter-on-quarter percentage change).
We also produced alternative estimates of the CoLS using two different econometric techniques -
an ARDL model of the expenditure-side GDP model and a panel regression where we estimate the
relationship between electricity consumption and GDP across 10 different industries using the fixed
effects, least-square dummy variable (LSDV) technique.
If the underlying variables in a regression model are non-stationary (i.e. trend over time) the goal
is usually to understand whether there is a long-run relationship between these variables. In such
cases, it is more appropriate to use specialist time series techniques, such as the ARDL to estimate
the coefficients, as opposed to a classic linear regression model (CLRM). Since we were interested
in isolating the impact of load shedding (a stationary variable) on GDP growth in the short-term,
we favour the CLRM results.
The estimated CoLS, based on our ARDL results, varied between R5.26/kWh and R6.59/kWh (Table
8). Our specification of the ARDL was based on the same theoretical model as the CLRM. The
resulting estimate of CoLS, however, was ~30% lower than our primary estimates (based on the
CLRM). We believe the ARDL technique underestimates the impact of load shedding on GDP,
perhaps because it is better suited to estimating the long-run relationship between slow-moving
non-stationary time series variables. We have provided a detailed description of the ARDL
regressions in Appendix D.
Table 8 Cost of load shedding (R/kWh, 2020 constant prices)
Primary estimate Alternative estimates
Period CLRM ARDL Panel

2007-08 7.61 5.26 2.86
2013-15 8.80 6.09 3.31
2018-19 9.53 6.59 3.58
We also estimated the CoLS based on the Cobb-Douglas production function, following the
approach used by Burger et al., 53 using a panel regression. This approach is more suited to
determining the long-term relationship between electricity consumption and GDP growth across
all ten industries. We used the estimated long-term relationship between electricity consumption
and GDP to infer what proportion of GDP would have been lost due to load shedding.
53
Kreuser, Burger, and Rankin
Novꭤ Economics 43
This systems approach has some advantages over the other single equation methods (i.e. CLRM
and ARDL) as it uses both the time-series variation within industries and the variation across
industries to simultaneously identify factor-augmenting technical changes and the contribution of
electricity consumption, capital, and labour to GDP growth across all sectors.
The estimates of the total national CoLS based on the panel regression were lower than our primary
estimates ranging between R2.86/kWh and R3.58/kWh (Table 8). As was the case with the estimates
based on the ARDL technique, we believe our estimates of the CoLS inferred from the panel
regression are underestimates. This is likely because the influence of load shedding on GDP had to
be inferred systems, as it could not be estimated directly using the systems approach. We believe
our estimates of the relationship between electricity, labour, capital and GDP growth are robust.
We are, however, less confident in the inferred amount of output that would be lost during load
shedding (based on the long-term direct relationship between electricity consumption and
economic activity). We have provided a detailed description of the panel technique in Appendix C.
Comparing the three estimates of CoLS (Figure 20), we find that the CLRM estimate is consistently
the largest, which suggests that the ARDL and panel techniques may understate the impact of load
shedding. All three estimates of the CoLS increased over time.
Figure 20 Cost of load shedding by periods and estimation method (R/kWh, 2020 constant prices)
9.53
8.80
7.61
CoLS (R/kWh)
6.59
6.09
5.26
3.58
3.31
2.86
2007-2008 2013-2015 2018-2019
CLRM (Primary estimate) ARDL (Secondary estimate) Panel (Secondary estimate)
Novꭤ Economics 44
We also compare our primary estimate of the CoLS to the estimate produced by Deloitte in 2009
(Table 9). Our primary estimate of the CoLS in 2007 - 2008 at R7.61/kWh is similar to the forward-
looking estimate produced by Deloitte in their ‘realistic load shedding scenario’ which was
R8.95/kWh (inflated to 2020 values). 54 We compare the values for the 2007-08 period since the
Deloitte study was conducted in 2009.
Table 9 Comparison of our primary estimate of the CoLS with Deloitte 2009
Deloitte estimates (R/kWh) Our estimates (R/kWh)
Estimate 2008 values 2020 values Estimate 2008 values 2020 values
Primary estimate,
Realistic scenario 4.92 8.95 4.18 7.61
(2007-08)
5.3. The impact of load shedding by industry
The impact of load shedding is not uniformly distributed across the ten sectors of the economy that
are defined in the national accounts. In Figure 21 we have illustrated the proportion of the total
national cost of load shedding (in R/kWh) attributed to each of the sectors of the economy. We
were able to estimate the cost of load shedding for every sector except for the mining industry. 55
The results of the single-equation regression estimates of the CoLS for each industry are
summarised in Appendix D.
Our results show that four of the nine sectors, namely manufacturing, transport and
communication, retail trade and agriculture bore just over 80% of the total national cost of load
shedding (expressed in rand of GDP lost for every kilowatt-hour of load shedding, i.e. R/kWh). The
manufacturing sector alone bore nearly 40% of the total cost of load shedding.
The extent to which an individual industry bears the cost is a function of its size, the electricity
intensity of the sector and the ability of firms in the sector to adapt to or mitigate against electricity
supply interruptions. For example, during the most recent period of load shedding, the
manufacturing sector lost R3.85 worth of output for every kWh of nationwide rotational load
shedding that occurred (~40% of the overall CoLS). By contrast, the financial and business services
54
55
Novꭤ Economics 45
sector lost only R0.07/kWh (~1% of the overall CoLS) while community and personal services sector
did not suffer any loss. 56
Figure 21 Contribution of each sector to the total cost of load shedding (R/kWh)
0.07
0.06 0.53 Mining
0.05 0.49 0.60
0.58 0.65 Pers.
0.38
0.45 0.64
0.99 Fin.
0.64 0.98
0.81 1.14 Govt.
1.05
CoLS (R/kWh)
0.89
Const.
1.93
1.49 1.76
Util.
Agri.
3.55 3.65 3.85 Retail

Transp.
Manuf.
2007-2008 2013-2015 2018-2019
We provided a summary of the cost of load shedding by industry, normalised by each industry’s
respective contribution to total GDP (Figure 22 and Figure 23).
Weighting the contribution of each sector to the total CoLS by its share in GDP (normalised CoLS)
shows that the agriculture industry, which is a relatively small contributor to total GDP (3.6%), was
the sector most adversely affected by load shedding. The normalised CoLS for the agriculture
industry was 4.2 times that of the average (R1/kWh). Manufacturing and utilities lost three times
more output than the average industry while output in the financial and business and community,
social and personal services industries emerged largely unscathed. This suggests that it was the
more energy-intensive primary and secondary industries that were the most adversely affected by
load shedding.
56
Our primary regression analysis indicated that the relationship between load shedding and GDP growth was not
statistically significant from zero
Novꭤ Economics 46
Figure 22 Contribution to total GDP by sector (%)
Fin.
Govt. 2.2 2.4
Retail
Manuf.
Transp.
23.1 17.1 15.1 13.3 9.3 7.9 6.0
Mining
Pers.
Util.
Constrn.
3.6
Agri.
Figure 23 CoLS: actual and normalised values (R/kWh, 2020 values)
4.01
3.85
2.85 2.86
CoLS (R/kWh)
1.93 2.03
1.62
1.14
0.99
0.75
0.65 0.60 0.53
0.32
N/A 0.07 0.03 0.00 0.00
Agri. Mining Manuf. Util. Const. Retail Transp. Fin. Govt. Pers.
CoLS Normalised CoLS
Note: A reliable estimate of the impact of load shedding on the mining sector GDP was not obtained.
In Figure 24 we have illustrated the impact of load shedding GDP growth by industry. Our estimates
suggest that the Agricultural sector, which was the worst affected, lost over 5% of its output because
of load shedding between 2013 and 2015. Load shedding was not the only factor responsible for
the contraction in agricultural output over that period but it was a significant contributor (Appendix
E: App Figure 8).
Novꭤ Economics 47
Loss of GDP attributed Figure 24 Impact of load shedding GDP growth by industry, 2007 to 2019
(percentage points)
to load shedding
2007-2008 2013-2015 2018-2019

Agri. -2.77 -5.35 -4.28
Util. -1.94 -3.88 -3.11
Manuf. -1.96 -3.85 -3.12
Transp. -1.40 -2.81 -2.25
Constrn. -1.13 -2.26 -1.78
Retail -0.53 -1.06 -0.85
Govt. -0.22 -0.45 -0.36
Fin. -0.02 -0.04 -0.04
Pers. 0.00 0.00 0.00
Mining A reliable estimate of the impact of load shedding on mining GDP growth was not obtained.
The other sectors that experienced a significant contraction in output due to load shedding (more
than 2% over the 2013 to 2015 period) are Manufacturing, Electricity & Water and Transport and
Communications. Further figures illustrating the impact of load shedding on the trend in quarterly
GDP growth for these and other sectors can be found in Appendix E.
While agriculture suffered the biggest losses due to load shedding relative to its contribution to
GDP, the manufacturing industry lost the largest amount in absolute terms – a total of R11 billion in
GDP between 2007 and 2019.
Novꭤ Economics 48
Figure 25 Decrease in GDP attributed to load shedding, per period and by industry (2020 prices)
Mining Pers. Fin. Govt. Constrn.

Util. Agri. Retail Transp. Manuf.
-0.09
-0.71
-0.79
2018-2019
-0.85
-1.29
-1.50
-2.55
-5.06
-0.11
-0.85
-1.01
2013-2015
-1.10
-1.69
-1.83
-3.06
-6.25
-0.05
-0.34
-0.39
2007-2008
-0.55
-0.72
-0.78
-1.30
-3.11
GDP (R bn, 2020 prices)
The cost of unserved energy (CoUE) is widely used internationally as a proxy for the cost of
infrequent unplanned outages. The CoUE is, in essence, a measure of each sector’s electricity
intensity- the rand of output the sector produces for every kWh of electricity consumed. The
estimates of the CoUE however, are not directly comparable with the cost of regular planned
outages (e.g. cost of load shedding) because they don’t account for differences in an industry or
firms inherent resilience planned outages or their ability to adapt to and mitigate against the
damages.
The CoUE suggests that it is the least electricity-intensive industries that would lose the most output
per kWh of electricity lost during a power outage. Our estimates of the CoLS suggest that the
opposite is true(Figure 26 and Figure 27). Energy-intensive sectors like mining, manufacturing and
agriculture are highly dependent on electricity for production, have fewer alternatives and are likely
to suffer much greater losses than more services-oriented industries like finance and business
services.
Novꭤ Economics 49
Figure 26 Estimate of CoUE by sector
385 390
333
CoUE (R/kWh)
178
126 131
67 67
44
29
Agri. Mining Manuf. Util. Constrn. Retail Transp. Fin. Govt. Pers.
Source: Eskom Cost of Unserved Energy Update: 2019
Figure 27 Estimate of CoLS by sector
4.01
Normalised CoLS (R/kWh)
2.85 2.86
2.03
1.62
0.75
0.32
N/A 0.03 0.00
Agri. Mining Manuf. Util. Const. Retail Transp. Fin. Govt. Pers.
Note: A reliable estimate of the impact of load shedding on mining GDP could not be estimated as it was not possible to
control for more than 10% of the variation in the highly volatile q/q growth in Mining GDP.
The difference between the two estimates is the most evident at the extreme: the least energy-
intensive industry, finance and business services, and most electricity-intensive industries such as
manufacturing. The CoUE for finance and business services is R390/kWh, this means that for every
unit of electricity the sector consumes it generates R390 worth of output. The assumption is that if
an unplanned outage were to occur it would lose R390 for every kWh of electricity that was not
provided. By contrast, the CoUE suggests that manufacturing would lose roughly a sixth of the
output of the finance sector for every kWh of load shedding occurred (R67/kWh).
Novꭤ Economics 50
This is not likely to be the case as the nature of the finance industry is that it is inherently quite
resilient to power outages. For example, many finance and business professionals would be able to
continue working on their battery-powered laptops, key IT systems would have back-up power
generation installed and some office buildings have back-up generation. Alternatively, people will
simply shift their workload to a later time.
Our estimate of the CoLS for the finance and business services sector (normalised), which is based
on the historical variance in sectoral GDP and therefore include differences in the inherent ability
of industries to adapt and mitigate costs, suggests that the normalised CoLS for the manufacturing
sector is R2.85/kWh and is 95 times higher than the normalised CoLS for the finance industry which
is just 3 cents per kWh (R0.03/kWh).
Load shedding cost the South African economy nearly R35 billion in the 12 years between 2007 and
2019. Had all the load shedding experienced over the period taken place in a single quarter in 2019,
it would have resulted in a 5% contraction real GDP growth (q/q%). To put this figure into
perspective the total cost of load shedding at R35 billion is roughly equivalent to the impact the
2008/9 financial crisis had on GDP growth (it also subtracted a cumulative five percentage points
from quarter-on-quarter GDP growth but over a much shorter period (Figure 28)).
Figure 28 Impact of load shedding on GDP growth, compared to Covid-19 and the financial crisis
4 5
GDP growth q/q (%)
0 4
-4 3
-8 2
-12 1
-16 0
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Load shedding Financial crisis Covid-19 pandemic GDP
Source: Nova Economics analysis based on seasonally adjusted and annualised time series GDP data from StatsSA.
The Covid-19 pandemic, however, resulted in a larger contraction in GDP in just one quarter - the
second quarter of 2020, when GDP fell by over 15% q/q than the combined impact of load shedding
on GDP over the past 12 years – which is equivalent to roughly a 6% q/q drop in GDP if it occurred
in a single quarter. The contraction in South African GDP in the second quarter of 2020 which was
precipitated by the Covid-19 pandemic was, however, the most severe contraction the SA economy
has experienced since the Second World War.
Novꭤ Economics 51
5.4. Caveats to the results
Our estimates of the CoLS capture the impact of load shedding as reflected in the variation in GDP
growth around its long-term trend. This includes the direct and indirect damages (e.g. loss of
output) and the costs of adaptation and mitigation (e.g. higher input costs such as investment in
back-up generation). However, they can still be considered conservative estimates because they
exclude the longer-term impact of recurring outages on business and investor confidence.
In an announcement on the 20th of March 2020, credit rating agency Moody’s noted that
“Unreliable electricity supply … continue[s] to constrain South Africa's economic growth” 57. Similarly,
in April 2020, Standard and Poor’s noted “In the second half of 2019, the economy shrank, due
partly to a set of severe rolling power blackouts,” and this was a factor in their confirmation of South
Africa’s sub-investment grade sovereign credit rating. 58
As this would be reflected in lower longer-term GDP growth, not in the variation around the trend,
it is not possible to identify this potential loss of output econometrically. While many factors have
contributed to the gradual decline in GDP growth in South Africa over the past decade, there is
little doubt that load shedding and persistent electricity supply constraints are among them.
While we were able to produce reliable estimates of the historical CoLS for nine of the ten sectors
reported in South Africa’s national accounts, we cannot comment on the differences within these
sectors – for example, the impact on transport vs. communication or the impact on tourism-related
activities which are distributed across different sectors. Our estimates are based on the variation in
macroeconomic aggregates and the data is not sufficiently disaggregated to comment on impact
at a sub-industry or firm level.
57
Moody’s, Moody’s downgrades South Africa’s ratings to Ba1, maintains negative outlook (2020),
www.moodys.com/research/Moodys-downgrades-South-Africas-ratings-to-Ba1-maintains-negative-outlook--PR_420630.
58
Standard and Poor's, "South Africa Ratings Lowered To 'BB-' From 'BB' As COVID-19 Further Impairs Fiscal And Growth
Prospects; Outlook Stable," (2020). https://www.standardandpoors.com/en_US/web/guest/article/-
/view/sourceId/11471713.
Novꭤ Economics 52
6. Key insights and recommendations
6.1. The cost of load shedding – key insights
The purpose of this study was to provide Eskom with reliable and accurate estimates of the
economic cost of load shedding. There were three distinct periods of load shedding: 1) October
2007 to April 2008, 2) November 2013 to October 2015, and 3) June 2018 to September 2020.
Our estimates suggest the CoLS increased steadily over time – from R7.61/kWh during the first
major episode of load shedding (2007-2008) to R9.53/kWh during the third period (2018 – 2019).
59
This was contrary to our expectation that in the context of recurring electricity supply shortages
the cost of load shedding (in kWh) would decrease, as electricity consumers adapted and invested
in back-up generation. The increase in the CoLS over the three periods may be related to the
increase in the frequency and/or severity of load shedding. Eskom estimates that it undertook a
total of ~870 GWh of load shedding in the 2008-9 period as compared to ~1 700 GWh during
2013-15 and ~1 307 GWh in 2018-19. The month in which load shedding was most severe was March
2019. Our estimates are similar to Eskom’s previous estimate of R8.95/kWh 60 produced by Deloitte
in 2009. 61
We estimate that Load shedding cost the South African economy a total of R34.5 billion 62 since
2007. Had all the load shedding experienced over these 12 years taken place in a single quarter in
2019, it would have resulted in a 5% contraction real q/q GDP growth. The put this into perspective,
the impact of load shedding is more or less equivalent to the total impact of the 2008/9 financial
crisis on the economy ( which subtracted roughly 5 percentage points from trend q/q GDP growth
over five quarters).
We also set out to assess how the cost of load shedding is distributed across different sectors of
the economy. We produced estimates of the cost of load shedding for all sectors of the economy,
except for the mining industry. 63 As anticipated, we found that the cost of load shedding is unevenly
distributed. During the third period of load shedding, four of the nine industries, worst affected by
load shedding bore 80% of the total cost (measured in R/kWh) - namely manufacturing (SIC 3),
transport and communication (SIC 6), wholesale and retail trade (SIC 5) and agriculture, hunting,
forestry and fishing (SIC 1). The manufacturing sector alone shouldered nearly 40% of the total cost
of load shedding.
The extent to which an individual industry bears the cost is a function of its size (i.e. contribution to
total GDP), the electricity intensity of the sector, and the ability of firms in the sector to adapt to or
59
60
61
62
Expressed in 2020 prices and for the period from 2007 until 2019.
63
Novꭤ Economics 53
mitigate against unplanned electricity supply interruptions. We found that the less electricity-
intensive and more service-oriented industries emerged relatively unscathed. For example, the
financial and business services sector, which is the largest contributor to GDP but is not very
electricity-intensive, lost just R0.07 per kWh of national load shedding during the most recent period
(2018-19). This was less than 1% of the total cost of load shedding over the period. The community
and personal services sector did not suffer any loss of GDP due to load shedding (the estimated
CoLS was not statistically significant from zero).
There are several reasons why service-oriented and less electricity-intensive industries are inherently
more resilient and better able to adapt to power outages. For example, professionals in the finance
and business services industry can continue to work during power outages – the electronic
equipment they rely on such as laptops central IT systems are fitted with back-up power generation
sources. Working hours in this industry also tend to be more flexible so that people can shift their
working hours to better accommodate load shedding.
We also normalised the cost of load shedding by each industry’s respective contribution to total
GDP. The normalised estimates reveal which sectors were the most adversely affected by load
shedding relative to their size. From this perspective, the agricultural sector was the most adversely
affected by load shedding. While the agricultural industry is a relatively small contributor to national
GDP (it accounts for 3.6% of total output) the normalised estimates of the CoLS show that it lost
4.2 times more output per kWh of load shedding than other industries, on average. Manufacturing
(SIC 3) and utilities (SIC 4) lost three times more output than the average, while output in the
financial and business (SIC 7) and community, social and personal services industries (SIC 8) were
largely unaffected.
6.2. The potential uses of the CoLS estimates
It was envisaged that the updated estimates of the CoLS may provide insights about the distribution
and magnitude of the cost of outages that could inform energy sector policy. Measures of the cost
of outages are useful in assessing the relative costs of interventions to mitigate against the risk of
load shedding, and so can be used to make socially optimal investment decisions.
Eskom and its shareholder face several choices when assessing how to mitigate against the risk of
further load shedding (examples are provided in Table 10). There are several potential options to
consider on both the demand and supply-side, but all of these interventions are associated with
financial and/or broader economic costs.
On the supply-side, the immediate options are limited but would include running emergency
generation or peaking plant (e.g. diesel-fired open cycle gas turbines) at higher load factors. On
the demand-side immediate options may include power buybacks from customers under contract,
voluntary curtailment by top customers or as a last resort emergency load shedding.
Over the short-to-medium term, several interventions can be considered. These may include
decisions to return moth-balled plants to service, building or procuring modular utility-scale
renewable capacity, investing in large-scale energy-efficiency and demand-side management
Novꭤ Economics 54
programmes or entering into new interruptible supply agreements. Other options that have been
mooted in the past include procuring power from mobile power generating ships and entering co-
generation agreements with domestic industry.
Table 10 Types of supply-side and demand-side interventions given persistent outages
Category Immediate 12-36 months >36 months

Supply-side • Running OCGTs at high load • Return-to-service of a • Grid extension by way of
interventions factors. mothballed plant additional generation
• Building modular utility-scale plants.
renewables capacity • Building modular utility-
• Power ships scale renewables capacity.
• Co-generation agreements
• Postponing scheduled
maintenance of plants.
Demand- • Power buybacks • Energy-efficiency demand- • Energy-efficiency

side • Reducing demand from top side management demand-side
customers under • Entering into interruptible management
interventions
interruptible and curtailable load supply agreements
load supply agreements
• Emergency load shedding.
Source: Adapted from Deloitte, 2009
It is also important to consider who bears the costs. For example, if the system operator decides to
run Eskom’s peaking plant (OCGTs) at higher load factors in a bid to avoid load shedding, Eskom
bears the cost and must motivate to recover the costs from the consumer via a higher tariff.
While the cost of load shedding at R9.53/kWh is higher than the cost of running OCGTs estimated
by EPRI 64 at R1.99/kWh (2015 prices), it is borne mainly by the most energy-intensive sectors of the
economy (e.g. manufacturing, mining, and transport, while Eskom itself only bears a small
proportion of the cost (i.e. lost electricity sales).
De Nooij et al 65 note that measures of the cost of outages are sometimes used to minimise the
economic costs of outages by informing which customers should be disconnected from the grid.
According to the representatives of Eskom, this is not practical in South Africa. Load shedding is
implemented at a substation level, making it difficult to isolate specific sectors. Although top
customers often have dedicated substations, most substations supply a large number of customers
spanning a broad range of sectors.
64
ELECTRIC POWER RESEARCH INSTITUTE, "Power Generation Technology Data
for Integrated Resource Plan of South Africa," (2015). http://www.energy.gov.za/IRP/2016/IRP-AnnexureA-EPRI-Report-
Power-Generation-Technology-Data-for-IRP-of-SA.pdf.
65
de Nooij, Koopmans, and Bijvoet, "The value of supply security - the costs of power interruptionsL Economic input for
damage reduction and investment in networks.."
Novꭤ Economics 55
References
Africa, Sustainable Energy. State of Energy in South African Cities. (2006, 2011, 2015).
Akpeji, Kingsley Oladipo. "Cost of Electricity Interruption to Commercial and Industrial End-Users."
Master of Science in Electrical Engineering, University of Cape Town, 2019.
Allcott, Hunt, Allan Collard-Wexler, and Stephen D O'Connell. "How Do Electricity Shortages
Affect Industry? Evidence from India." American Economic Review 106, no. 3 (2016): 587-624.
Andersen, Thomas Barnebeck, and Carl-Johan Dalgaard. "Power Outages and Economic Growth
in Africa." Energy Economics 38 (2013): 19-23.
Botelho, Vinícius. "Estimating the Economic Impacts of Power Supply Interruptions." Energy
Economics 80 (2019): 983-94.
Carden, Kevin, and Nick Wintermantel. The Economic Ramifications of Resource Adequacy. Astrape
Consulting (for EISPC and NARUC) (2013).
de Nooij, Michiel., Carl. Koopmans, and Carlijn. Bijvoet. "The Value of Supply Security - the Costs
of Power Interruptionsl Economic Input for Damage Reduction and Investment in Networks.".
Energy Economics 29 (2007): 277-95.
Deloitte. "Modelling the Impacts of Electricity Disruptions, Chapter 3, Report on Eskom and the
Electricity Sector." edited by Landon Mc Millan, 2009.
———. "The Macroeconomic Impacts of Alternative Scenarios to Meet Eskom’s Five-Year

Revenue Requirement." 16 June 2017 2017.
http://www.eskom.co.za/Documents/EcoStudyMacroeconomicImpact2017.pdf.
Division, Eskom Sustainability. "Cost of Unserved Energy Update: 2019 ", 2019.
http://nersa.org.za/regulator-decisions/.
Ellahi, Nazima. "Testing the Relationship between Electricity Supply, Development of Industrial
Sector and Economic Growth: An Empirical Analysis Using Time Series Data for Pakistan."
International Journal of Management Science and Engineering Management 6, no. 4 (2011): 272-
77.
Eskom. "Cost of Unserved Energy Methodology." 2015.
"Eskom 2000 - 2008 - Our Recent Past -"Shift Performance and Grow Sustainably"."
https://www.eskom.co.za/sites/heritage/Pages/2000.aspx.
———. "Integrated Report 2019." 2019. https://www.eskom.co.za/IR2019/Pages/default.aspx.
Novꭤ Economics 56
Fisher-Vanden, Karen, Erin T Mansur, and Qiong Juliana Wang. "Electricity Shortages and Firm
Productivity: Evidence from China's Industrial Firms." Journal of Development Economics 114 (2015):
172-88.
Goldberg, Ariel. "The Economic Impact of Load Shedding: The Case of South African Retailers."
University of Pretoria, 2016.
Gujurati, Damodar N. Basic Econometrics. McGraw-Hill Higher Ed., 2003.
Kreuser, Friedrich, Rulof Burger, and Neil Rankin. "The Elasticity of Substitution and Labour-
Displacing Technical Change in Post-Apartheid South Africa." UNI-WIDER, 2015/101 (2015).
Major, Klára, and Luca Flóra Drucker. Macroeconomic Impact of Electric Power Outage: Simulation
Results from a Cge Modelling Experiment for Hungary. EcoMod (2016).
Minnaar, UJ, W Visser, and J Crafford. "An Economic Model for the Cost of Electricity Service
Interruption in South Africa." Utilities Policy 48 (2017): 41-50.
Moody’s. Moody’s downgrades South Africa’s Ratings to Ba1, Maintains Negative Outlook. (2020).
www.moodys.com/research/Moodys-downgrades-South-Africas-ratings-to-Ba1-maintains-
negative-outlook--PR_420630.
Munasinghe, Mohan, and A Sanghvi. "Reliability of Electricity Supply, Outage Costs and Value of
Service: An Overview." The Energy Journal 9, no. Special Issue 2 (1988).
Munasinghe, Mohan., and A. Sanghui. "Reliability of Electricity Supply, Outage Cost and Value of
Service: An Overview." The Energy Journal (1988). https://doi.org/10.5547/ISSN0195-6574-EJ-Vol9-
NoSI2-1 · Source: RePEc.
Ou, Peng, Ruting Huang, and Xin Yao. "Economic Impacts of Power Shortage." Sustainability 8,
no. 7 (2016): 687.
Poor's, Standard and. "South Africa ratings Lowered to 'Bb-' from 'Bb' as Covid-19 Further Impairs
Fiscal And growth Prospects; Outlook Stable." 2020.
https://www.standardandpoors.com/en_US/web/guest/article/-/view/sourceId/11471713.
Public Affairs Research Institute (PARI). Why the Lights Went Out: Reform in the South
African energy Sector. . Graduate School of Development Policy and Practice, University of Cape
Town (2013).
Rose, Adam., Shu-Yi. Liao, and Gbadebo. Oladosu. "Business Interruption Impacts of a Terrorist
Attach on the Electric Power System of Los Angeles: Customer Resilience to a Total Blackout." Risk
Analysis 27, no. 3 (2007): 513-31. https://doi.org/10.1111/j.1539-6924.2007.00912.x.
Wang, Xiaoshu, and Yu Fu. "Some Characteristics of the Cobb-Douglas and Ces Production
Functions in Microeconomics." Abstract and Applied Analysis (2013).
Novꭤ Economics 57
Wing, Ian Sue, and Adam Rose. "Iii. Economic Consequence Analysis of Electric Power
Infrastructure Disruptions: An Analytical General Equilibrium Approach." Frontiers in the Economics
of Widespread, Long-Duration Power Interruptions: Proceedings (2019): 71.
Novꭤ Economics 58
Estimating the magnitude of load
shedding.
Detailed approach to producing the top-down estimate
The top-down s estimate of the magnitude of load shedding, supplied by Eskom’s system operator,
is based on the difference between day-ahead forecasts of the national load profile and the actual
demand profile over the days when load shedding took place (App Figure 1).
App Figure 1 Process taken by the system operator to estimate the magnitude of load shedding
Historic data used to calculate residual demand forecast
Actual demand different to forecast, yielding forecast error
Load shedding implemented
Forecast error for period of load shedding estimated, based on forecast error
observed before and after load shedding
Forecast demand less estimated forecast error during LS calculated
Magnitude of load shedding estimated as forecast demand less error less actual
consumption
Eskom’s system operator produces forecasts of the residual electricity demand profile o over the
next 24-hour period, daily Actual observed demand is likely to be slightly lower or higher than the
forecast- the difference between forecasted demand and actual demand on a given day is referred
to as the forecast error. The difference between forecast and actual demand is depicted for an
example day, 13 February 2019, with no load shedding (App Figure 2).
Novꭤ Economics Appendix A 59

App Figure 2 Forecast demand versus actual demand, 13 Feb 2019
32
30
Electricity demand (MWh)
28
26
24
22
20
18
01:00
02:00
03:00
04:00
05:00
06:00
07:00
08:00
09:00
10:00
11:00
12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
23:00
00:00
Residual demand forecast Actual demand
Adapted from a presentation by Eskom at the NERSA Public Hearing, 24 February 2020.
Since it is not possible to directly observe load shedding, the Eskom system operator accepts that
it must be estimated. When load shedding takes place, the amount of “load shed” in GWh is
estimated by calculating the difference between forecast demand and actual demand during load
shedding. However, the system operator also adjusts this estimate for the forecast error before and
after load shedding took place.
The estimation of the forecast error is represented in App Figure 3, where it is assumed that load
shedding took place between 9 am and 11 pm, for example. The actual forecast error was observed
as 161 MWh at 8 am, and 543 MWh at midnight, based as the difference between forecast demand
and actual consumption at 8 am, and midnight for this example day. The system operator derives
a linear trend line between these two data points and estimates the forecast error for the hours
where load shedding is experienced, represented in teal.

App Figure 3 Forecast error estimation
Actual forecast error Estimated forecast error
600 543
Forecast error (MWh)
500
400
300
200 161
100
The system operator estimates load shedding by taking the difference between the actual demand
and forecast demand during a load shedding event and accounting for the forecast error (App
Figure 4).
Based on discussions with the team responsible for producing the forecasts at the system operator’,
we understand that day ahead forecast of residual demand is produced based on a statistical
analysis of historic average trends in sales and variables that influence demand as well short-term
actual consumption data from the recent past (last week/month/quarter). This means that the
recent trends in sales including a decline in demand due to load shedding would be factored into
the system operator estimate of future demand forecasts.
As a result, the magnitude of load shedding might be underestimated by the system operator. To
validate the system operator’ estimate of the magnitude of load shedding, we calculated a second
measure of the magnitude of load shedding, based on substation level data on the incidence of
load shedding.

App Figure 4 System operator top-down estimate of load shedding
Residual demand forecast less estimated forecast error

Actual demand during load shedding
Residual demand forecast
Actual demand before and after load shedding
32 Magnitude of load shedding 32000
30 30000
Electricity demand (GWh)
28 28000
26 26000
24 24000
22 22000
20 20000
01:00
02:00
03:00
04:00
05:00
06:00
07:00
08:00
09:00
10:00
11:00
12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
23:00
00:00
Bottom-up estimate
We calculated the bottom-up estimate of load shedding to independently estimate the magnitude
of load shed in GWh and validate the system operator’ estimates. Our bottom-up estimate is
calculated using historic electricity sales at a granular level and considers consumption patterns of
various types of customers.
Eskom’s top customer segment consists of energy-intensive firms, mainly mining and industrial
operations, which consume more than 100 GWh per annum. Stable and consistently reliable
electricity supply is critical for Eskom’s top customers, given the energy-intensive nature of these
customers operations. Top customers, therefore, enter into agreements with Eskom which regulates
the quantity and stability of their electricity supply. These agreements often contain provisions that
top customers will not be subject to load shedding, but rather agree to load curtailments and power
buybacks in the event the national grid is under pressure. Our bottom-up approach separately
estimates load curtailment and power buybacks from top customers, and the magnitude of
scheduled rotational load shedding to all other customers.

Estimate of the magnitude of load shedding for customers without load curtailment
agreements
We obtained substation level monthly sales data from 2007 to 2019 (in GWh) from the distribution
performance division. This data excluded sales to the top customer segment. A robust estimate of
the magnitude of load shedding should look at what consumers would demand in the absence of
load shedding. Demand modifications by customers solely as a result of load shedding being
implemented should not factor into the estimate of the magnitude of load shedding. To account
for potential demand modifications by consumers in periods with high incidences of load shedding,
we attempted to estimate average consumption per month per substation based on two reference
years without load shedding, (2016 and 2017). This gave us a baseline indication of what electricity
consumption could be at a substation level in the absence of load shedding.
We then obtained a second dataset detailing all the incidences of load shedding per substation.
This dataset contained the time, date, and duration of every load shedding incident since 2007, at
a substation level. Based on this dataset, we were able to derive an average availability factor of
electricity per month at a substation level for every year since 2007. This availability factor was
calculated as the average duration of load shedding in minutes experienced by a particular
substation divided by the total number of minutes in a month. We were then able to estimate the
magnitude of load shedding in GWh experienced by a substation per month by multiplying the
availability factor calculated for a given month by the average electricity sales from that substation
for the same month in a non-load shedding year.
Magnitude of load curtailments and power buybacks for top customers not subject to
rotational load shedding
Top customers are subject to curtailment and power buybacks as opposed to load shedding, as
experienced by other customer segments. The process we followed to estimate the magnitude of
these two components of load reduction is illustrated in App Figure 5.
App Figure 5 Process taken to estimate the magnitude of curtailments by top customers
Obtained dataset containing incidences of all curtailment

events to top customers
Obtained electricity sales data which detailed sales to top customers

on a half hourly basis
Added indicators to sales data including hour, week, month and whether curtailment occurred
Estimated average sales to top customers on an hourly basis

without curtailment
Curtailment magnitude estimated as average sales less actual sales during curtailment events

The starting point for estimating the magnitude of load curtailments was to obtain sales data for all
sales to top customers. This sales data was broken down into kWh sold to all top customers
nationally on a half-hourly basis for the period 2011-2019 (48 separate data points for each day in
the year = 24 hours x 2). Based on this data, we determined that electricity sales to top customers
is volatile and fluctuates significantly during the day, with peaks and troughs at certain times.
App Figure 6 illustrates the fluctuation in hourly sales for an average day based on this sales data.
Pricing agreements signed with top customers specify time-of-use tariffs, where electricity sales
outside of business hours and during off-peak are substantially cheaper than during peak demand
times. Sales in MWh to top customers peak during the night and early hours of the morning, and
are lowest during times of peak residential demand, such as between 6 am to 10 am, and
5 pm to 9 pm.
App Figure 6 Average daily electricity demand profile for top customers (2012-2019)
4,600 1
4,550 0.9
4,500 0.8
0.7
4,450
Hourly sales (MWh)
0.6
4,400
0.5
4,350
0.4
4,300
0.3
4,250 0.2
4,200 0.1
4,150 0
00:00
01:00
02:00
03:00
04:00
05:00
06:00
07:00
08:00
09:00
10:00
11:00
12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
23:00
Peak national demand Hourly sales Average daily sales
Based on descriptive statistics and graphic analysis of the sales data, we found that electricity
demand to top customers fluctuates based on the day of the week such that consumption patterns
and total demand on a weekday are different to weekends, for example. We discovered that, even
for weekdays, there are inter-day fluctuations; on average (i.e. the consumption pattern for an
average Monday differs from that of the average Tuesday). Likewise, monthly electricity demand
fluctuates. Electricity sales in an average January and an average June are different. Based on these
observed variations, we added indicator variables such as day of the week, the hour of the day and
month of the year to the sales dataset that allowed us to analyse and segment the data more
closely.
After adding indicator variables to the dataset, we obtained another dataset which contained the
incidences of all load curtailment events since 2014. This dataset contained the time, date, duration,
and stage of all curtailment events. Curtailment agreements signed with key customers specify that

under stage one and two of curtailment, up to 10% of the normal electricity demand should be
curtailed. Under level three, this rises to 15%, while stage 4 requires a 20% reduction in demand.
Our estimate of the magnitude of load curtailment considers the hourly, daily, and monthly variation
in sales to key customers. We calculated an average consumption for each type of indicator monthly
when no curtailment events took place. For example, we calculated the average consumption from
8 am to 9 am on an average Monday in September 2015. We calculated unmet electrical load per
incidence of load curtailment as the difference between average demand of key customers in the
absence of curtailment (using the detailed averages calculated taking into account the time and
date of curtailment and hour of curtailment) and actual electricity sales observed during curtailment.

Review of previous studies on the cost of power outages
Summary of previous studies on the cost of load shedding in South Africa
Overview and critique of previous studies on the cost of load shedding in South Africa
App Table 1 Summary of the literature on the economic impacts of power shortages in South Africa and the Sub-Saharan African region
Author and title Approach Description Critique

Deloitte (2009). Forward-looking • Study simulate the potential economic impacts of • The authors assumed that scheduled load shedding of 3 000 MW,
Modelling the General equilibrium electricity supply disruptions using a dynamic 1 500 MW and 750 MW would be associated with a resultant loss in
Impacts of model computable general equilibrium (DCGE) model detailing the capital stock of 10%, 5% and 2%. But the assumption about the
Electricity 40 economic sectors. amount of capital stock that would be rendered idle by load
Disruptions, • In the first scenario, the authors simulate an outage by shedding does not appear to have been calibrated to any historical
Chapter 3, Report cutting the capital stock of the electricity sector by 10%. data and is unsubstantiated.
on Eskom and the The impact is transmitted through an increase in • The model does not allow for adaptive resilience or mitigation of
Electricity Sector." electricity prices (which reflect increased scarcity). The the impact of outages –only inherent resilience of different sectors
authors note however that this assumption is unrealistic to load-shedding incidences. The model does not allow for the
as Eskom’s tariffs are regulated. possibility of adaptive resilience or mitigation of the impact of
• In the second set of scenarios, the capital stock of all 40 outages – it only simulates the inherent resilience of different
sectors is cut by 10% to simulate the impact of load sectors to load-shedding.
shedding on GDP. In the final scenario, the capital stock • The authors try to contrast their CoLS estimates to the cost of
was cut by varying amounts (2%, 5% or 10%) based on running open-cycle gas turbines, but this analysis was flawed since
the electricity intensity of each sector. they were unable to relate any of the scenarios modelled to the
• The study found that load shedding cost the economy actual magnitude or frequency of load shedding that would have
R4.90/kWh. This was much lower than estimates of the taken place.
CoUE for infrequent unplanned outages of R75/kWh (PB
Power 2008).
Novꭤ Economics Appendix B 66

Goldberg (2015) Retrospective • Studied the impact of LS on the retail sector through • The semi-structured interviews insights were based on a small
The economic Stated-preference three alternative methods, 1) semi-structured interviews sample (eight interviews), while the sample sizes for the state-
impact of load survey with retail managers, 2) administering a state-preference preference and revealed preference methods were also fairly limited
shedding: The case Revealed- or willingness-to-pay survey of retail outlets online and (106 and 42 respectively).
of South African preference survey in-store, and 3) collecting financial data from retail head • The estimates of the CoLS based on the revealed-preference survey
retailers Case study based offices on the cost of providing back-up generation were limited to the cost of running back-up generators and at a
on semi-structured (marginal cost method under revealed preference survey total of R716 million and did not include the cost of damages (direct
interviews method). The authors derived estimates of the CoLS for or indirect) but also did not subtract the cost of grid-supplied
retailers for different times of the day based on a power that would have been purchased.
customer damage function approach. • The stated preference estimates of R13bn for 1H15 likely
overestimated the impact on retailers.
Andersen, and Retrospective • Estimates the total effect of power outages on economic • Makes use of a very parsimonious model of outages on economic
Dalgaard. (2013) Econometric growth in Sub-Saharan Africa over the period 1995-2007. growth – they explain economic growth only in terms of the g (GDP
Power Outages and analysis of time Considers both potential errors of measurement of growth per capita), an intercept, the log of outages and an error
Economic Growth series data. African economic growth and the endogeneity of term.
in Africa. outages. • In trying to limit the risk of endogeneity by limiting explanatory
• Used Penn World Tables GDP data with satellite-based variables they introduce omitted-variable bias (OVB) – as a result
data on nightlights to arrive at a more accurate measure probably attributes too much of the variance in GDP to outages.
of economic growth. Data on outages obtained from
World Bank’s Enterprise Surveys 2011. Lightning density
was used as an instrument for power outages. Results
suggest weak power infrastructure is a substantial drag
on growth.

Review of forward-looking studies focusing on those based on general
equilibrium models
Introduction
Wing and Rose note that except for a few case studies of actual events, economy-wide losses
associated with both planned and unplanned power outages were typically not analysed until the
1990s. 66 Vinícius Botelho notes that decisions about whether to invest in infrastructure or initiatives
to increase power system resilience have been measured mainly in terms of the cost of unplanned
outages (either CoUE or VoLL) or output lost in values per kWh of electricity not supplied. 67
However, the empirical methods most often used to estimate these parameters (partial equilibrium
(PE) models, such as input-output tables or social accounting matrices) suffer from several
limitations. PE models are limited by their inherent linearity, lack of behavioural content and do not
allow for prices to adjust to clear markets. 68
More specifically, measures of the cost of outages based on partial equilibrium models (e.g. CoUE)
rely on the convenient but unjustified assumption that a sector’s output is directly proportional to
its electricity consumption and that resources (e.g. capital and labour) are unconstrained. PE models
also assume fixed prices (no change in prices or wages in response to demand or supply shocks).
Analyses based on PE models cannot account for the fact that interruption costs are time-
dependent, although some have proposed adaptations to account for this by applying weighting
factors. 69 Lastly, and perhaps most importantly, they do not account for the fact that some
industries are more resilient to power outages than others and can adapt – this is particularly
relevant when there is a structural electricity supply shortage. 70
Most of the recent advances in modelling the economic consequences of electricity disruptions are
concerned with trying to more accurately reflect the varying ability of consumers to mitigate the
losses incurred. As discussed in Section 2.1.3, there are many different tactics that firms employ to
cushion themselves against the impact of power outages – some are inherently more resilient (are
not very reliant on electricity or can easily outsource the parts of their processes that are) while
others can adapt (for example by procuring backup power solutions).
66
67
Vinícius Botelho, "Estimating the economic impacts of power supply interruptions," Energy Economics 80 (2019).
68
69
Mohan Munasinghe and A Sanghvi, "Reliability of electricity supply, outage costs and value of service: an overview," The
Energy Journal 9, no. Special Issue 2 (1988).
70

Key features of general equilibrium models
CGE models, as noted by Rose and Wing, 71 maintain the best features of PE models – the high
level of sectoral detail and ability to trace linkages between industries – while overcoming many of
the limitations of PE models (e.g. fixed prices and unconstrained resources). CGE models are
particularly useful for analysing the potential impact of regular planned outages as it is possible to
simulate some of the resilience tactics firms are likely to employ. For example, power conservation
strategies can be simulated by changing the productivity parameter of a production function, while
the inherent resilience of a sector (e.g. ability to substitute inputs and switch to imports) are built
into the dynamic equations in the form of substitution elasticities and Armington elasticities (i.e. the
elasticity of substitution between products of different countries, a standard assumption of
international CGE models), respectively.
Computable general equilibrium (CGE) or dynamic computable general equilibrium models (DCGE)
build-up to the macro-economy from microeconomic foundations. CGE models are complete
numerical representations of economies in the form of systems of non-linear algebraic equations
or related mathematical structures. In contrast to I-O or other PE models, the input-output relations
among industries are nonlinear and to a degree flexible, a function of technology assumptions,
prices, and other factors. A CGE model consists of:
• A system of nonlinear algebraic equations describing the interaction between micro and
macroeconomic variables; and
• A database which takes the form of a social accounting matrix (SAM) and a set of elasticities
describing how product substitution takes place.
A SAM model describes the structure of the economy and interlinkages between industries, sectors,
households, government, and the rest of the world at a detailed level. The theoretical structure of
the typical CGE models are based on neo-classical economic theory: consumers are assumed to
maximise utility and firms maximise profits subject to resource constraints.
Advantages and limitations of estimating the impact of electricity supply shortages within
a GE framework
Of the 13 studies we found on the economic consequences of persistent electricity shortages, six
analyse the potential impacts within a general equilibrium framework. The main advantage of using
a static CGE model over the PE models used to estimate the CoUE is that the inherent resilience of
a sector - ability to substitute inputs and switch to imports - is built into the dynamic equations of
all CGE models in the form of substitution elasticities and Armington elasticities, respectively.
As Wing et. al. note, CGE models overcome the very limiting assumption that economic actors do
not engage in substitution (as is the case for measures based on PE models). The analysis by Wing
71

et al. revealed that when producers and consumers can substitute (e.g. use alternative materials or
labour for electricity or imports for domestically produced intermediate inputs) the impact of supply
interruptions remain unambiguously negative for the economy but the magnitude of the impact is
much smaller.
Of the studies we reviewed, there were four that use a CGE to simulate the impact of outages while
allowing for inherent resilience across sectors - these included studies on the economic impacts of
power shortages by Major and Drucker 72 in Hungary, by Ou et al. 73 in China, Deloitte 74 in South
Africa and Botelho in Brazil. 75
All these studies share limitations. In all cases, the electricity shortages were initially simulated by
moving the electricity supply curve left - which reduces the quantity of electricity supplied and
increases its relative price. However, in all three country cases studied (China, Hungary and South
Africa) prices are administered or regulated and there is no perfectly competitive market or ‘efficient
market mechanism’ that would allow prices to rise to reallocate power to its most valuable use. In
this regard, analyses based on a CGE framework is likely to underestimate the impact of shortages.
All these studies share limitations. In all cases, the electricity shortages were initially simulated by
moving the electricity supply curve left - which reduces the quantity of electricity supplied and
increases its relative price. However, in all three country cases studied (China, Hungary and South
Africa) prices are administered or regulated and there is no perfectly competitive market or ‘efficient
market mechanism’ that would allow prices to rise to reallocate power to its most valuable use. In
this regard, analyses based on a CGE framework is likely to underestimate the impact of shortages.
Most of the CGE models used were based on simple Cobb-Douglas production function which has
a high degree of substitutability between key inputs to production and consequently also
underestimated the true impact of supply shortages. Botelho noted that the general equilibrium
properties of rationing policy, other production functions should be tested.
None of these studies attempted to model the ability of firms and households to employ adaptive
measures to cushion the impact of outages by either reducing their reliance on electricity or by
installing a back-up power supply or generation.
CGE models are theoretical and highly stylised (a model which only tries to reproduce a very
specialized time series or economic phenomena) and are therefore quite abstracted from reality.
For example, in the Deloitte study, the authors assumed that scheduled load shedding of 3 000
MW, 1 500 MW and 750 MW would be associated with a resultant loss in the capital stock of 10%,
5% and 2%. But the assumption about the amount of capital stock that would be rendered idle by
72
Klára Major and Luca Flóra Drucker, Macroeconomic impact of electric power outage: simulation results from a CGE
modelling experiment for Hungary, EcoMod (2016).
73
Peng Ou, Ruting Huang, and Xin Yao, "Economic impacts of power shortage," Sustainability 8, no. 7 (2016).
74
75
Botelho, "Estimating the economic impacts of power supply interruptions."

a given amount of load shedding does not appear to have been calibrated to any historical data
or outage event.
Because CGEs are abstract, theoretical, and comparative they cannot be used to assess the actual
historical impact of load shedding on the economy or a particular sector. However, as Botelho
notes, GE models do provide a sound theoretical and empirical basis to evaluate the relative
economic impacts of different power rationing policies – concerning both intensity and design.
Some of the more recent studies including Rose et al. and Wing et al. have shown how CGE models
can be extended to evaluate the ability of firms to mitigate against or adapt to regular electricity
shortages. In the most ‘realistic’ of the three scenarios modelled, Wing et al., extend their CGE
models to allow firms to make for deliberate investments in mitigation (back-up generation) to
further dampen the consequent price and quantity changes, and ultimate welfare losses. However,
they found that because of the complexity of the algebraic equations the results were hard to
interpret, and they had to rely on the fairly crude assumption that all industries employ a single
monolithic backup technology.
In conclusion, GE models, because they are highly disaggregated (often including detail on between
20 and 100 different sectors in the economy and the linkages between them) provide a useful
theoretical framework to assess the potential impact of various power rationing policies or events
on different sectors.
They can also be used to provide more accurate estimates of the cost of electricity shortages
(particularly regular planned outages) than traditional measures based on PE models (such as CoUE
and VoLL). This is because they overcome many of the limiting or unrealistic assumptions that PE
models are based on – they allow for prices to adjust, place a constraint on resources and allow for
substitution between inputs (e.g. capital and labour) and local production and imports.

App Table 2 Summary studies based on forward-looking approaches to simulating the impact of electricity shortages

Major, Klára, and Luca General • Presents the results of a CGE that is used to assess how an electricity • The CGE assumes perfectly competitive markets so that the adjustment is
Flóra Drucker. (2016) equilibrium supply shock might influence the Hungarian economy. driven by mainly by rising electricity prices.
Macroeconomic Impact model • The electricity outages are modelled by the decrease in the supply of • In the base case, when prices are flexible and there are no limits to
of Electric Power energy. The capital stock in the energy industry is shocked, which leads to adjustments, agents with the highest motivation to do so react more: either
Outage: Simulation a decrease in the supply of energy. because their price-elasticity is higher or because their energy intensity is
Results from a CGE • In the base scenario, a ~2% decline in the supply of energy leads to a higher or because they still need to compete with foreign competitors.
Modelling Experiment 0.53% decline in the GDP. • Therefore, this estimation must be considered as a lower bound on actual
for Hungary. costs of electricity outages: if there are limitations to adjustments of any
kind, the GDP costs of the outage can be even higher.
Ou, Peng, Ruting General • A static CGE model is used to simulate the economic impacts of hard • The power shortages are simulated by moving the electricity supply curve
Huang, and Xin Yao equilibrium power shortage and soft power shortage. Simulation results show that left (reducing the quantity of electricity supplied and increasing the price).
(2016). Economic model the negative effects of power shortage on economic development are In reality, electricity prices in China are administered and cannot adjust so
Impacts of Power very significant, and the effects vary across sectors. The study simulates the rising prices as an ‘efficient market mechanism’ to reallocate power is
Shortage. in China the impact of four shortage scenarios - a: 3%, 7%, 11% and 15% reduction an unlikely scenario and in reality, the impact of shortages would likely be
in the power supply. greater than reported.
• The results suggest that in the scenario of hard power shortage, the • On the other hand, the study does not appear to have incorporated the
industrial sector suffers most. The economic cost of power shortage is possibility of adaptative responses such as mitigation (e.g. back-up power).
considerable, and the study notes the main reason for it is the specific
administrative pricing system that applies to electricity in China. There is
little doubt that the cost of avoiding the shortage is less than the cost of
the power shortages.

Botelho, Vinícius. General • This paper notes that the costs of power system rationing are usually • One of the limitations of the study is that CGE is based on a Cobb-Douglas
(2019). Estimating the equilibrium estimated using reduced form linear models or I-O analysis that are ill- production function which high-degree of substitutability between inputs
Economic Impacts of model suited to understanding how consumers respond to shortages or probably underestimated the true rationing impact values. The general
Power Supply rationing policies. This paper reviews alternative approaches to equilibrium properties of rationing policy, other production functions must
Interruptions. Energy estimating the effects of power system rationing and concludes that GE be tested.
Economics. models are best suited to evaluate the effects of different power-
rationing strategies. The paper concludes that general equilibrium
models provide theoretical and empirical bases to estimate the economic
impacts of rationing policies – both rationing intensity and design.
Rose, Liao, Oladosu. General • The study estimates economic losses from electricity interruptions for • The study is limited to assessing the economic impacts of an electric power
2007. Business equilibrium businesses from a terrorist attack in Los Angeles. Indirect effects and outage on businesses.
Interruption Impact of a model resilience are also included in the losses estimate. Indirect effects • The study states that several considerations are omitted, such as the value
Terrorist Attack on the (multiplier/ GE effects) and resilience are included in the losses estimate. of any lives lost, increased crime, psychological trauma, some infrastructure
Electric Power System They find moderate indirect effects, but strong resilience, that pushes in costs, and property damage.
of Los Angeles: the opposite direction, indicating that customers can mute the potential • The authors acknowledge that many of the resilience factors are rough
Customer Resilience to shock to their business operations by as much as 86%. estimates and that more empirical work is needed to refine them.
a Total Blackout • They find that the most prominent resilience measure is the rescheduling
(recapture) of production after electric service is restored. This, together
with inherent aspects of the electricity-economy relationship (e.g. inter-
fuel substitution) and adaptive responses (e.g. conservation, on-site
generation) can reduce the potential disruption impacts significantly.
Wing, Sue, and Rose. General • The authors developed a simple analytical general equilibrium model of • The study is limited by the stylised, highly simplified nature of GE models. It
(2019) Economic equilibrium the economy-wide impacts of electricity infrastructure disruptions. The is not consistent with the physical reality of the power system and
Consequence Analysis model authors modelled three scenarios – in the first they consider the extreme additional research would be required to align the model with the reality of
of Electric Power case where economic actors do not engage in substitution (as is the case the power sector.
Infrastructure in studies based on PE models). • This was particularly relevant for the scenario where they allowed
Disruptions: An substitution and back-up because they relied on a rather crude assumption
Analytical General that all industries employ a single monolithic backup technology.
Equilibrium Approach.

• In a second more realistic scenario where producers and consumers do • The authors note that their model was insufficiently detailed in terms of the
engage in substitution, the results change dramatically impact on the number of electricity-using sectors represented, and the model’s static
power supply is unambiguously negative, as before, but the impact is nature meant it is not well-suited to analysing the longer-term impacts of
much is smaller in magnitude. In the third scenario, they allow for firms to power disruption events. The authors recommend developing a dynamic
make for deliberate investments in mitigation (back-up generation) to multi-sectoral CGE simulation which could be coupled with techno-
further dampen the consequent price and quantity changes, and ultimate economic power system models to address these limitations.
welfare losses.
Greenberg, Mantell, Structural • The economic impacts of potential terrorist attacks on the New Jersey • The authors state that a weakness of the model is that the set of important
Lahr, Felder, econometric electric power system are examined using a regional econometric model. relationships among sectors, as well as their magnitudes and directions, are
Zimmerman. (2007) time series • The study finds that the magnitude and duration of the effects vary by fixed. It is unknown how relationships among sectors change when one or
Short and intermediate model and I-O type of business and income measure. more of them suffers a large, unexpected shock.
economic impacts of a model • The suggested policy implication is that the cost and benefits of making • The authors also acknowledge that the model cannot directly capture the
terrorist-initiated loss of the electric power system more resilient to plausible attacks should be immediate reactions to economic shocks of business, government, and
electric power: A case weighed and that the restorative capacity of the system should be consumers.
study of New Jersey strengthened. • Most importantly, the authors state that no simulation model can perfectly
forecast the implications of a shock.
• The accuracy of the results is subject to limited information and data.

Review of retrospective studies with a focus on econometric methods
Introduction
Of the 13 studies we reviewed, four were based on an econometric analysis of the cost of historical
electricity supply shortages. Fisher-Vanden, Mansur and Wang 76 and Allcot et al. 77 estimated the
impact of electricity shortages based on panel data for manufacturing firms in China and India
respectively (App Table 3). Andersen, and Dalgaard 78, as discussed earlier, estimated the impact of
electricity shortages across Sub-Saharan Africa using time series techniques, while Ellahi 79 estimates
the impact of electricity supply constraints on the development of the industrial sector in Pakistan
using an ARDL model (App Table 3).
Key features of econometric models
Econometrics can be defined as the social science in which the tools of economic theory,
mathematics, and statistical inference are applied to the analysis of economic phenomena. 80 The
econometric analysis of the cost of electricity supply interruptions is usually based on one of two
types of data – times series data or pooled and panel data. A time series is a set of observations
on the values that a variable such as GDP takes at different points in time – usually collected at
regular intervals (monthly, quarterly, annually). Pooled data sets are multi-dimensional in that they
contain values for different attributes of several firms or countries collected over time. Panel data
sets are simply a special type of pooled data in which the same cross-sectional unit (e.g. firm) is
surveyed over time.
Advantages and limitations of the econometric approach to estimating the historical

impact of electricity supply shortages
One of the main advantages of studies based on econometric analysis of historical time series or
panel data is that, in contrast to highly stylised and theoretical GE models, the results are based on
the empirical analysis of real-world outage events. Studies based on econometric analysis typically
aim to isolate the marginal impact of a particular load shedding event (or series of events) on
economic growth or industry-level production. In this sense, they can be used to produce estimates
of the actual historical economic costs (direct and indirect and net of adaptive response) of regular
or persistent power shortages. For example, Andersen and Dalgaard 81 estimated across a sample
of 39 Sub-Saharan African countries that a one per cent increase in the number of outages reduced
long-run GDP per capita by 2.86 per cent during the period 1997 to 2007. It is possible however
76
Fisher-Vanden, Mansur, and Wang, "Electricity shortages and firm productivity: evidence from China's industrial firms."
77
78
79
Ellahi, "Testing the relationship between electricity supply, development of industrial sector and economic growth: An
empirical analysis using time series data for Pakistan."
80
Gujurati Damodar N Gujurati, Basic econometrics (McGraw-Hill Higher Ed., 2003).
81

that they overestimated the impact of outages on GDP per capita, by failing to adequately control
for other influences.
One of the limitations of econometric analyses is that it can be difficult to accurately isolate the
impact of load shedding (particularly if it is a one-off event) on economic growth from other
influences. One of the challenges in this regard, endogeneity. Power outages are likely correlated
with a number of the other determinants of economic growth (e.g. growth in gross fixed capital
formation) and are subject to reverse-causal influence (economic growth also causes shortages).
To address the potential endogeneity of shortages a proper identification strategy is required.
Anderson and Dalgaard 82 instrument for shortages by using lightning density as an exogenous
determinant of power disturbances. Lightning damage accounts for about 65% of all over-voltage
damage to electrical distribution networks in South Africa. Allcot et al. 83 instrument with changes
in electricity production from dams, which are driven by changes in the amount of water flowing
into reservoirs.
Econometric analysis is also frequently limited by the relatively low frequency of macroeconomic
data (quarterly or annual), quality of data on electricity outages and the length of the economic
time series – the introduction of several lags of each explanatory variable for example in a Vector
Autoregressive (VAR) or Vector Error Correction Model (VECM) model can consume a lot of
degrees of freedom.
Another advantage of retrospective econometric studies it that is possible to estimate the net
impact of power shortages on different groups or sectors - to understand which firms or sectors
are more resilient (able to avoid or cushion the impact) than others. For example, Allcott et al. 84
found that in 2005 when Indian manufacturing firms faced outages, 7.1% of the time, that the output
of firms that were able to self-generate (i.e. had back-up generators) lost only 0.7% of their output.
Firms that did not have back-up generation lost 10.3% of their output. The authors concluded that
electricity shortages were a substantial drag on Indian manufacturing from 1992 to 2010, reducing
manufacturing output by an average of about five per cent over the period.
82
83
84

App Table 3 Summary of international studies that estimate the historical impact of power shortages (retrospective approach)

Fisher-Vanden, Mansur, Econometric • Based on an analysis of a panel of 23 000 energy-intensive Chinese firms • The study was limited to studying the effects of electricity
Wang (2015) analysis of panel from 1999 to 2004. Examined how firms responded to severe power scarcity on energy-intensive manufacturing firms.
Electricity shortages data shortages in the early 2000s. • The study does not examine the effect of power
and firm productivity: • The study found that in response to electricity scarcity, Chinese firms shortages on investment and employment by industrial
Evidence from China’s substitute materials for energy – buying rather than making intermediate firms.
industrial firms inputs to production (i.e. outsourcing). This enabled firms to minimise
productivity losses.
• Even the less energy-intensive industries, like food and machinery,
increased materials share and reduced shares of electricity. The study did
not find evidence that electricity shortages led to a marked increase in self-
generation – there was some evidence in mining, food, and petroleum, but
even so, less than 20% of firms opted to self-generate. As a result of the
increase in electricity scarcity unit production costs rose by 8%.
Oseni, Musiliu O, and Revealed • Using cross-sectional data from 6854 firms currently operating in 12 African • The study (like most others) does not illustrate the
Michael G Pollitt. (2013) preference countries. Based results on applying 3 different methods to the survey data broader and longer-term impacts of power shortages on
The Economic Costs of survey - marginal cost, incomplete backup (revealed preference techniques) and investment and employment.
Unsupplied Electricity: (retrospective) subjective evaluation techniques (stated preference). • It also did not estimate the potential environmental costs
Evidence from Backup Stated • The study found the cost of an unsupplied kWh of electricity is significantly associated with back-up power (e.g. noise and emissions
Generation among preference higher than the cost of electricity from the public grid. Lastly, they found from diesel back-up generators).
African Firms. survey (forward- that the cost of mitigating a kWh of electricity is significantly higher than a • Finally, the conversion of the reported electricity
University of looking) cost-reflective tariff would have been. expenditure by firms to obtain their corresponding
Cambridge, Faculty of • Firms engaging in exports, and those using the Internet for their operation electricity demand by dividing it by an approximate price
Economics. suffered higher unmitigated outage costs despite having a higher may not be accurate.
propensity of investing in a back-up generation. Unmitigated costs or
damages still account for the larger proportion of the total outage costs
despite the high prevalence of backup ownership among the firms.

Hunt, Collard-Wexler, Econometric • Developed a hybrid Leontief and Cobb-Douglas production function • The results are based on a static model, focusing on the
and O'Connell (2016) analysis of model that characterises how input shortages affect firms (2013). In the effects of annual variation in shortages with fixed capital
How Do Electricity pooled data. case study, they analysed how ‘power holidays’ affect daily production at stock. However, because most of the policies available to
Shortages Affect Case study large Indian textile plants, using data from Bloom et al. address shortages would be unlikely to fully eliminate
Industry? Evidence • Studied the short-run effects of electricity shortages on all Indian shortages for many years, a model with annual variation
from India. manufacturing plants between 1992 and 2010, using archival data on may identify the most policy-relevant effects.
shortages, panel data, and an instrument for shortages based on variation
in hydro reservoir inflows. Found that electricity shortages were a
substantial drag on Indian manufacturing, reducing output by about five
per cent. Found that because of economies of scale in self-generation,
shortages impose much greater losses on small plants,
Ellahi (2013) Econometric • The empirical analysis is based on Romer’s endogenous growth theory and • The regression model used in this study is subject to a
Testing the relationship times series uses an Auto-Regressive Distributed Lag (ARDL) model, using data for number of shortcomings. Firstly, it only includes a simple
between electricity analysis Pakistan from 1980-2009. binary dummy variable for the periods with and without
supply, development of • The author finds that electricity shortages in Pakistan have a significant electricity shortages, without capturing any information
industrial sector and negative short-run and long-run effect on the performance of the on the frequency or magnitude of the outages. Further,
economic growth: An industrial sector. They find that electricity is a significant contributing factor the regression includes variables for both outages and
empirical analysis using to long-run economic growth. electricity consumption, two variables that almost
time series data for certainly correlated.
Pakistan • Furthermore, it appears the model was poorly specified
and has low explanatory power as none of the estimated
coefficients in either the error-correction or long-run
from of the ARDL model (on outages or other
determinants of growth) were statistically significant –
the p-values were all very high.
• The study also produces a seemingly counterintuitive
negative relationship between labour and growth, which
further detracts from confidence in the results.

Energy-augmented production function
In this section, we describe the panel regression which provided the alternative estimate on the cost
of load shedding and provides technical detail on the data inputs and methods used. Our panel
regression is based on the Cobb-Douglas production function, a positive nonconstant function that
expresses economic output as a function of capital, labour and total factor productivity. The
production function is a key concept in mainstream neoclassical economics. 85 Almost all economic
theories presuppose a production function, either on the firm level or the aggregate level.
The panel regression was estimated pooled classic linear regression model with fixed effects, where
the fixed effects control for the different industries. We structure the components of the production
function as matrices (by sector, at one-digit SIC level). The agriculture and utilities industries are,
however, excluded from the panel production function regression, as the Cobb-Douglas
specification is not a good predictor for their GDP output.
The coefficients on the variables yielded sensible results (App Figure 7 and App Table 4). When we
included electricity sales as an input variable, the coefficient on capital reduced. This was expected,
as electricity consumption is positively correlated with capital stock formation - the more
equipment, buildings, machinery, the more electricity is consumed.
We estimated how much growth in electricity sales have contributed to growth in GDP over the last
~17 years (2003 to 2019) and found that a 1% increase in electricity sales was associated with a
0.15% increase in GDP after controlling for the influence of capital, labour, and technology.
Thereafter, we estimated the percentage point change in GDP growth that could have been
attributed load shedding in each of the quarters when load shedding occurred. This was done by
multiplying the coefficient of 0.15 by the percentage of electricity sales lost due to load shedding.
We inferred amount of GDP lost during load shedding by multiplying by the amount of GDP
generated in each of the 16 quarters and dividing by the magnitude of load shedding in GWh and
inflated these costs to 2020 prices.
85
Wang and Fu, “Some Characteristics of the Cobb-Douglas and CES Production Functions in Microeconomics.”
Novꭤ Economics Appendix C 79

App Figure 7 Relationship between GDP, capital and labour for each industry
Legend: 1 = Agri; 2 = Mining; 3 = Manufacturing; 4= Utilities; 5 = Construction; 6 = Trade; 7 = Transport; 8 = Business; 9

= Personal; 10 = Government

App Table 4 Panel regression output
Dependent Variable: LOG(GDP_EXCL_AGRI_UTIL)

Method: Panel Least Squares
Date: 08/03/20 Time: 13:56
Sample (adjusted): 2003Q1 2019Q4
Periods included: 68
Cross-sections included: 8
Total panel (balanced) observations: 544
Variable Coefficient Std. Error t-Statistic Prob.
C -0.239637 0.229806 -1.042780 0.2975

LOG(OPENING_CAPITAL_SMOOTHED) 0.235433 0.023198 10.14879 0.0000
LOG(LABOUR_PALMS) 0.469995 0.021258 22.10881 0.0000
LOG(SALES) 0.149549 0.012205 12.25307 0.0000
Effects Specification
Cross-section fixed (dummy variables)
Root MSE 0.053745 R-squared 0.991130

Mean dependent var 12.46813 Adjusted R-squared 0.990964
S.D. dependent var 0.571191 S.E. of regression 0.054297
Akaike info criterion -2.968682 Sum squared resid 1.571372
Schwarz criterion -2.881755 Log likelihood 818.4815
Hannan-Quinn criter. -2.934696 F-statistic 5955.817
Durbin-Watson stat 0.187126 Prob(F-statistic) 0.000000

The method used to derive the quarterly capital stock series
Annual fixed capital stock by industry is published by the SARB. Until 2016, the SARB also published
a quarterly series of gross fixed capital formation for six sectors including mining and manufacturing.
We sourced the discontinued series from the SARB and interpolated annual series to generate
quarterly data for the remainder of the period (2016 to 2019). Following the method used by Burger
et al. 86 the quarterly fixed capital formation by industry is used to calculate the quarterly capital
stock by simultaneously solving the equations below for capital stock 𝑘𝑘𝑡𝑡,𝑞𝑞 in quarter 𝑞𝑞 of year 𝑡𝑡. 𝐾𝐾𝑡𝑡
and 𝐾𝐾𝑡𝑡−1 are the gross fixed capital stock values reported at the end of year 𝑡𝑡, while 𝐼𝐼𝑡𝑡,𝑞𝑞 is the gross
fixed capital formation for the quarter 𝑞𝑞 in year 𝑡𝑡. The solution of these equations also allows for
quarterly depreciation, 𝑑𝑑, to be identified per industry. The depreciation rates can then be used to
construct a quarterly capital series.
𝑘𝑘𝑡𝑡,1 = (1 − 𝑑𝑑)𝐾𝐾𝑡𝑡−1 + 𝐼𝐼𝑡𝑡,1

𝑘𝑘𝑡𝑡,2 = (1 − 𝑑𝑑)𝑘𝑘𝑡𝑡,1 + 𝐼𝐼𝑡𝑡,2
𝑘𝑘𝑡𝑡,3 = (1 − 𝑑𝑑)𝑘𝑘𝑡𝑡,2 + 𝐼𝐼𝑡𝑡,3
𝐾𝐾𝑡𝑡 = (1 − 𝑑𝑑)𝑘𝑘𝑡𝑡,3 + 𝐼𝐼𝑡𝑡,4
The method used to derive the quarterly series on electricity sales by

sector
Data on electricity sales were sourced from Eskom. Three different sources of sales data were used
to derive monthly series of electricity sales for each of the ten industries reported by Standard
Industrial Classification (SIC) code at the 2-digit level in the national accounts, these include:
8. Highly disaggregated electricity sales and revenue data classified in most cases at the 5-
digit SIC code) for each of Eskom’s 6 regions and its top customers, extracted from the
billing system at Eskom distribution for the years 2006 to 2019.
9. Highly disaggregated national sales and revenue data (classified in most cases at the 5-
digit SIC code) sourced from Eskom distribution for the years 2003 to 2008.
10. Aggregated total monthly national sales data by major Eskom customer category,
extracted from SAP system at Eskom distribution from 2006 to 2019.
Eskom does not report electricity sales and revenue by at the SIC 2 -digit level, which is how GDP
data is aggregated in the South African System of National Accounts. Eskom distribution does
classify its direct sales to commercial and industrial customers by 5-digit SIC codes in the billing
system but these are then aggregated into the following 13 Eskom customer categories which do
not map in most cases to the SNA (System of National Accounts):
11. Independent Power Producers
86
,Frierich Kreuser, Rulof Burger and Neil Rankin, The elasticity of substitution and labour-displacing technical change in
post-apartheid South Africa.l

12. Agricultural
13. Re-distributors
14. Commercial
15. Industrial
16. International Sales
17. Mining
18. Prepayment
19. Public lights
20. Residential
21. Traction
22. External Sales
23. Internal Sales
The process that we followed to categorise Eskom’s highly disaggregated electricity sales into
industries at the 2-digit SIC level is as follows:
We began by mapping the highly disaggregated monthly sales and revenue, recorded by
region and in some cases by 5-digit SIC code to the 10 industries at 2-digit SIC level, as
reported in the national accounts using the categorisation presented in App Table 5.
Some categories of sales could not be mapped to the SIC industries. These included various
residential sales, international and bulk sales, and sales to re-distributors (Local
municipalities and major metropolitan areas). In doing so we had to account for missing
values and negative sales values (which were included) because they are corrections of prior
billing errors.
Eskom records these sales by “Eskom region”, six are defined by geographic boundaries
and the seventh consists of the utility’s top customers – the regions are Central, Eastern,
Northern, North West, Southern, Western, and top customers. The regional data classified
by the ten industries and four additional customer categories were summed to create a 14
separate national series of monthly sales and revenue.
We compared these series where possible with Eskom’s aggregates and found that there
was an issue with the residential electricity sales series in Eskom’s billing system and so we
replaced this series with data from Eskom’s SAP system.
The Eskom sales and revenue data captured in the current billing and SAP systems only
dates back to 2006 but for the econometric analysis, we needed a longer time series.
Fortunately, we were able to obtain a previous set of data on national monthly electricity
sales and revenue at a similar level of disaggregation (classified in most cases at the 5-digit
SIC code) from Eskom distribution for the years 2003 to 2008.
We repeated the mapping exercise for 2003 to 2008 sales data using the classification in
Appendix F and then combined these to create national time series of electricity sales for
ten industries in the SNA and four additional sectors spanning the period 2003 to 2019.

The final challenge in the process was how to allocate the almost 40% of Eskom’s total
direct electricity sales that are sold to municipal redistributors to the various industries that
are the end-consumers but buy Eskom power via municipalities.
To allocate municipal sales to the relevant industries that consumed them we used the
breakdown of municipal electricity sales by industry provided in the State of Cities report
for 2006, 2011 and 2014/15 and additional data provided by the authors of Eskom’s CoUE
report for the 2014/15 year. The reports cover sales by 13 of South Africa’s largest
metropolitan and local municipalities.
Since we only had a breakdown of municipal sales by the industry for three of the 17 years
between 2003 and 2019, we interpolated the trend in the composition of sales for the
intervening periods.
We then summed Eskom’s direct sales and reallocated municipal sales by industry to create
ten series reflecting Eskom’s monthly sales by industry (at the 2-digit level) but this excludes
international sales and residential sales.
App Table 5 Mapping Eskom sales to 2-digit SIC classification of industries
SIC category (high-level) Allocation by Eskom codes New SIC codes

Agriculture, Forestry and Fishing 1 01 to 03
Mining and Quarrying 2 05 to 09
Manufacturing 3 10 to 33
Electricity, Gas and Water 412, 413, 420, 41111, 41113, 4SPU 35 to 39
Construction 5 41 to 43
Wholesale and Retail Trade, Hotels and 6 45 to 47
Restaurants
Transport, Storage and Communication 7 49 to 63
Finance, Real Estate and Business Services 8, 5SPU Commercial 64 to 82
General Government Services 91, 4publiclighting 84
Personal Services 92 to 99 85 to 99
Redistributors 41112
Residential 0SPU Residential, 0SPU
Prepayment, 01010-Urban
Domestic, 01020-Rural Domestic,
01040-Peri-Urban Domestic
International 41114
Bulk supplies 41116

Expenditure side regression output
App Table 6 Classic linear regression model estimates (2003Q1-2019Q4)
Manu. Retail sales Transport Utilities Personal Agri. Mining Constrn. Finance Govt. Total GDP
Constant 0.521269* 0.051437- 0.494876*** -0.063304- 0.104773- -0.017698- 0.394786- 0.337557- 0.481112*** 0.43392*** 0.235226**
(0.224299) (0.113183) (0.117793) (0.197326) (0.114995) (0.332945) (0.639073) (0.261984) (0.107877) (0.09286) (0.094041)
LS_GWH_PC_sales -0.189244 -0.314774 -0.8458*** -0.201381** -0.0 -1.671265 -0.933103 -0.677822 -0.013 -0.131501 -0.388425
(Load shedding as a % of P. 0.1384 P. 0.1683 (0.2333) (0.060718) P. 0.945 P. 0.0622 - P 0.1976 P 0.9509 P. 0.4778 P. 0.0248
sales) (0.597257) (0.225958) (0.203605) (0.880134) (1.139481) (0.520646) (0.211459) (0.184134) (0.169019)
@PC(FCE) (Final Consumption 0.857712*** 0.443443***
Expenditure) (0.107671) (0.098685)
@PC(KBP6009D_GFCF) (Gross 0.085613* 0.055798- 0.094237** 0.465339*** 0.187551** 0.139543*** 0.090173*** 0.099911***
Fixed Capital Formation) (0.079620) (0.065512) (0.031362) (0.114435) (0.068372) (0.028296) (0.024856) (0.023029)
@PC(KBP6008D_FCEG) (FCE 0.15908- 0.166686- -1.06692* 0.837519** 0.308122** 0.186421*
Govt.) (0.115928) (0.09639) (0.519746) (0.247345) (0.100562) (0.088686)
@PC(KBP6007D_FCEH) (FCE 0.500258* 0.396386*** 0.692995-
Households) (0.060718) (0.097992) (0.429689)
@PC(KBP6010D_Inventories) -0.002141*
(0.000919)
@PC(KBP6014D_Imports) -0.056416**
(0.020525)
Dummy 2008/09 -4.16851*** -0.62303* -1.849984 * 0.481256* -0.595136**
(Financial crisis dummy) (0.841064) (0.314098) (0.807396) (0.247164) (0.22963)
Dummy Oil Price 1.141546***
(0.314508)
Dummy Rains 8.819864***
(1.141744)
Dummy Drought -8.414656***
(1.159585)
Dummy Credit Boom 1.568994***
(0.279152)
R2 0.4091 0.546975 0.53939 0.395272 0.440478 0.718645 0.100717 0.307327 0.581595 0.302363 0.67734
N 68 68 68 68 68 68 68 68 68 68 68
Note: *** p<0.001, **p<0.01, *p<0.05, - p>0.05 (not significant.); standard errors in parentheses
Novꭤ Economics Appendix D 85

App Table 7 ARDL regression estimates (2003Q1-2019Q4)
Manu. Retail sales Transport Utilities Personal Agri. Mining Constrn. Finance Govt. Total GDP
Selected model LOG(GDP): 1 LOG(GDP)L 1 LOG(GDP): 2 LOG(GDP): 3 LOG(GDP): 2 LOG(GDP): 2 LOG(GDP): 1 LOG(GDP): 1 LOG(GDP): 1 LOG(GDP): 1 LOG(GDP): 1
(lags in dynamic lag lag lags lags lags; lags lag lag lag lag lag
regressors) LOG(GGCF): LOG(FCE): 1 LOG(FCEG): 2 LOG(Capital LOG(GFCF) 0 LOG(GFCF): 1 LOG(FCEG): 4 LOG(FCEG): 0 LOG(GFCF): 0 LOG(FCEG): 0 LOG(GGCF): 2
4 lags lag lags formation): 4 lags;LOG(FC lag lags lags lags lags lags
LOG(FCE): 2 LOG(GFCF): 1 lags’ EG): 0 lags LOG(FCEH): 3 LOG(GFCF): 2 LOG(FCEH): 0 LOG(GCF): 2 LOG(FCE): 1
lags lag LOG(DCEH): 1 ;LOG(FCEH): lags lags lags lags lag
lag 4
lags;LOG(IM
PORTS): 4
lags
Select variables of interests:
LOG(GDP(-1)) 0.534063*** 0.833977*** 1.161302*** 0.76037*** 1.198837*** 1.142966*** 0.32944** 0.845908*** 0.828885*** 0.862783*** 0.786639***
(0.092118) (0.071677) (0.11666) (0.107013) (0.108366) (0.068232) (0.116073) (0.03508) (0.029611) (0.033092) (0.075328)
LOG(GDP(-2)) -0.24494 ** 0.015982 - -0.496085 -0.262934 **
(0.111716) (0.140204) *** (0.066821)
(0.096726)
LOG(GDP(-3)) 0.172375
-
(0.111795)
Constant 1.7186 *** -0.074063 - 0.277113 ** 0.551607 - 0.177988 - 0.39239 - 9.986889*** 0.114948 - -0.379026 ** 0.035769 - 0.859376 **
(0.402874) (0.111396) (0.112042) (0.416624) (0.17506) (0.278261) (1.685167) (0.281963) (0.171887) (0.071121) (0.230412)
Fixed regressors:
LS_GWH_PC_sales -0.01044 -0.004061 -0.00722 -0.010603 -0.001178 -0.016654 0.001672 -0.00944 0.00057 -0.0000108 -0.002572
(Load shedding as * P 0.0729 ** P. 0.0155 P. 0.4452 P. 0.0438 P. 0.8649 P. 0.0046 P. 0.7883 P. 0.9951 P. 0.1201
a % of sales) (0.005275) (0.002226) (0.002522) (0.004244) (0.001532) (0.008087) (0.009786) (0.003206) (0.002114) (0.001743) (0.00163)
Dummy 2008/09 -0.051031*** -0.007693 ** -0.005452 - -0.016361 ** -0.01638 ** -0.000771 - 0.00438 - -0.009724 **
(Financial crisis) (0.009423) (0.003386) (0.004025) (0.007703) (0.00534) (0.00363) (0.002786) (0.002857)
Dummy Drought -0.081022***
(0.010529)
Dummy Rains 0.068087***
(0.011159)
Dummy Oil Price 0.010459**
(0.003766)
R2 0.982626 0.998624 0.998802 0.960811 0.999222 0.975834 0.755778 0.99911 0.999187 0.99933 0.999156
N 68 68 68 68 68 68 68 68 68 68 68
Note: *** p<0.001, **p<0.01, - p>0.05 (not significant); standard errors in parentheses
Novꭤ Economics Appendix D 86

Impact of load shedding on GDP growth
by sector
App Figure 8 Impact of load shedding on agriculture sector GDP growth
15
10
GDP growth q/q (%)
-5
-10
-15
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Agri Agri GDP growth in the absence of LS
App Figure 9 Impact of load shedding on manufacturing sector GDP growth
4
GDP growth q/q (%)
-2
-4
-6
-8
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Manu. Manu. GDP growth in the absence of LS
Novꭤ Economics Appendix E 87

App Figure 10 Impact of load shedding on utilities sector GDP growth
2
GDP growth q/q (%)
-2
-4
-6
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Util. Util. GDP growth in the absence of LS
App Figure 11 Impact of load shedding on construction sector GDP growth
6
GDP growth q/q (%)
-2
-4
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Constr. Constr. GDP growth in the absence of LS

App Figure 12 Impact of load shedding on retail sector GDP growth
2
GDP growth q/q (%)
-1
-2
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Retail Retail GDP growth in the absence of LS
App Figure 13 Impact of load shedding on transport sector GDP growth
2
GDP growth q/q (%)
-1
-2
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Transp. Transp. GDP growth in the absence of LS

App Figure 14 Impact of load shedding on finance sector GDP growth
3
GDP growth q/q (%)
-1
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Fin. Fin. GDP growth in the absence of LS
App Figure 15 Impact of load shedding on government sector GDP growth
2
GDP growth q/q (%)
-1
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Govern. Gov. GDP growth in the absence of LS

App Figure 16 Impact of load shedding on personal services sector GDP growth
2
GDP growth q/q (%)
-1
-2
2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Decline in GDP attributed to LS GDP Pers. Pers. GDP growth in the absence of LS

Industry classification by SIC code
App Table 8 Industry classification by SIC code
SIC Short name Description

SIC 1 Agri. Agriculture, hunting, forestry, and fishing
SIC 2 Mining Mining and quarrying
SIC 3 Manu. Manufacturing
SIC 4 Utilities Electricity, gas, and water (Utilities)
SIC 5 Constr. Construction
SIC 6 Retail Wholesale and retail trade, hotels, and restaurants
SIC 7 Transp. Transport, storage, and communication
SIC 8 Fin. Finance, insurance, real estate, and business services
SIC 92-96, 99 Pers. Community, social and personal services
SIC 91 Govt. General government
Novꭤ Economics Appendix F 92

www.novaeconomics.co.za

Appendix C - Estimating The Economic Cost of Load Shedding in South Africa Report

Uploaded by

Copyright:

Available Formats

Appendix C - Estimating The Economic Cost of Load Shedding in South Africa Report

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Appendix C - Estimating The Economic Cost of Load Shedding in South Africa Report

Uploaded by

Copyright:

Available Formats

Estimating the economic cost

of load shedding in South Africa

Kay Walsh, MCom (economics).

Ahmed Seedat BBus Sci (finance) CA (SA).

© Nova Economics – economic and strategy consulting

Novꭤ Economics viii

Novꭤ Economics Background and context 1

Novꭤ Economics Background and context 2

1.3. When and how often did load shedding occur?

Novꭤ Economics Background and context 3

• Period 1: October 2007 to April 2008 (load shedding during 7 of 7 months)

• Period 2: November 2013 to October 2015 (load shedding during 16 of 24 months)

Period 1: Period 2: Period 3:

Novꭤ Economics Background and context 4

Novꭤ Economics Background and context 5

Novꭤ Economics Background and context 6

Load shedding Operating reserve margin

Load shedding (GWh)

40 Slump in electricity demand 400

Source: Nova Economics analysis data provided by Eskom

Novꭤ Economics Background and context 7

2.2. Distinguishing between planned and unplanned supply interruptions

Unplanned outages Electricity supply shortages

Cost of unserved energy (CoUE)* Cost of load shedding (CoLS)**

Novꭤ Economics Electricity supply interruptions – the key concepts 8

Novꭤ Economics Electricity supply interruptions – the key concepts 9

Novꭤ Economics Electricity supply interruptions – the key concepts 10

Category Applications Measure

Generation planning • Optimising the reserve margin CoLS or CoUE

Operations & • Generation plant service scheduling CoLS or CoUE

Tariff design • Time of use tariff pricing design CoUE or CoLS

2.4. The costs associated with power outages

Novꭤ Economics Electricity supply interruptions – the key concepts 11

Table 3 Factors that impact the cost of electricity outages

Damages and costs

Resilience measures and tactics

Source: Nova Economics analysis based on Munasinghe and Sanghui, 1988.

Novꭤ Economics Electricity supply interruptions – the key concepts 12

Novꭤ Economics Electricity supply interruptions – the key concepts 13

Figure 3 Load shedding incidences by quarter, 2007-2020

Period 1 Period 2 Period 3

2007 2013 2015 2019

2008 2014 2018 2020

3.2. Two approaches to estimating the magnitude of load shedding

Novꭤ Economics Estimating the magnitude of load shedding 14

Top-down estimate Bottom-up estimate

Source: Nova Economics analysis based on Eskom sales data

Novꭤ Economics Estimating the magnitude of load shedding 15

• Data on curtailment by Eskom’s top customers 34 (largest mining and industrial

Novꭤ Economics Estimating the magnitude of load shedding 16

Novꭤ Economics Estimating the magnitude of load shedding 17

Load shedding (GWh)

Source: Nova Economics analysis

Novꭤ Economics Estimating the magnitude of load shedding 18

Novꭤ Economics Estimating the magnitude of load shedding 19

Bottom-up estimate Top-down estimate

2007 2008 2013 2014 2015 2018 2019 2020

Source: Nova Economics analysis based on data provided by Eskom

Novꭤ Economics Estimating the magnitude of load shedding 20