Forecasting The Nature of Power Transformer Insulation Oil Using Chemical Properties of Dissolved Gases
Forecasting The Nature of Power Transformer Insulation Oil Using Chemical Properties of Dissolved Gases
Forecasting The Nature of Power Transformer Insulation Oil Using Chemical Properties of Dissolved Gases
ABSTRACT
The life expectancy of transformers at various operating conditions is not accurately known.
Deterioration of power transformer insulation is a function of time and temperature. Since in power
transformer, the temperature distribution is not uniform, the part which is operating at the highest
temperature will ordinarily undergo the greatest deterioration. Therefore, it is usual to consider the effects
produced by the highest temperature hottest spot. The hottest-spot winding temperature is the principal
factor in determining life due to loading. The temperature cannot be measured directly because of the
hazards in placing a temperature detector at the proper location because of voltage. In this paper a neural
network based regression analysis was done to forecast the nature of power transformer insulation oil and
life time of transformer including loading conditions.
INTRODUCTION
gassing from dissolved gas analysis results, the transformer is subjected to frequent testing
weekly or even daily to understand, what is happening inside the transformer and to prevent
a catastrophic failure.
The main difficulty in making use of dissolved gas analysis results is that it is not
easy to draw the line between normal and abnormal results. So in4 interpretation schemes
include a normal condition as one of the diagnostic outcomes, but have not been particularly
effective in reliably identifying a normal condition. There are three most important
discussions needed from user point of view in power transformer.
Int. J. Chem. Sci.: 13(4), 2015 1779
Tables 3 and 4 show the dissolved gas value in ppm with variable load for a 400/230
KV transformer. Figs. 1 and Fig. 2 show that variation of acetylene with respect to load and
variation of hydrogen with respect to load. The inference from the Fig. 1 is, when sudden
increase in acetylene gas level in transformer oil from 2 ppm to 10 ppm shows that there is
increase in load hence load, current. Also from Fig. 2, we conclude that as load increases,
due to thermal effect the hydrogen value in ppm also get increased10. As per the standard,
sudden increase in hydrogen leads thermal fault. Due to this thermal effect, cellulose
degradation will occur and acetylene gas level also get increased, which leads to arcing fault.
This data is used to train the neural network using Levenberg algorithm and the fault of that
transformer was predicted.
800
600
C 2H 2
400
Load (MW)
Current (A)
200
0
.11
.11
.11
13
13
13
12
12
12
12
.2.
.4.
.6.
.1.
.3.
.7.
.9.
.7
.9
.11
26
30
27
29
21
30
29
21
20
30
800
700
600
500
H2
400 Load (MW)
300 Current (A)
200
100
0
11
11
.11
13
13
13
12
12
12
12
.7.
.9.
.1.
.3.
.7.
.9.
.2.
.4.
.6.
.11
26
30
27
29
21
30
29
21
20
30
Levenberg’s main contribution to the method was the introduction of the damping
factor λ. This value is summed to every member of the approximate Hessian diagonal before
1782 K. I. Annapoorani: Forecasting the Nature of Power….
the system is solved for the gradient. Typically λ would start as a small value such as 0.1.
Then the Levenberg-Marquardt equation is solved commonly by using LU decomposition.
After the equation is solved, the weights w are updated using δ and network errors for each
entry in the training set are recalculated. If the new sum of squared errors has decreased, λ is
decreased and the iteration ends. If it has not, then the new weights are discarded and the
method is repeated with the higher value of λ.
DGA data
Initialize network
weight w and δ
Compute Jacobian J
Compute error
t
gradient g = J E
Solve (H+λI) δ = g
and find δ
Update weights
w and
δ
If
Sum squared error <
initial value
Stop
Variations of the algorithm may include different values for v, one for decreasing λ
and other for increasing it. Others may solve (H + λ diag (H)) δ = g instead of (H + λI) δ = g,
while others may select the initial λ according to the size of the elements on H, by setting
λo = tmax (diag (H)), where t is a value chosen by the user. We have chosen the identity
matrix equation because it is the same method implemented internally by the Neural
Network Tool Box in MATLAB.
105
104
103
0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
2 Epochs
The above performance plot shows, how the network performance improved during
training. It shows the value of the performance function verses the iteration number. The
network performance is measured in terms of mean squared error, shown in log scale. It
rapidly decreased as the network was trained. The performance plot doesn’t indicate any
major problems with training because the validation and test curves are very similar at both
ends. If the test curve increased, then it is possible that some over fitting might have
occurred. Performance is shown for each of the training, validation and test sets. The version
of the network that did best on the validation set was after training. The mean squared error
of the trained neural network can now be measured with respect to the training samples. This
will give us a sense of how will the networks will do, when data applied from real world.
The average squared error is the difference between the network outputs (a) and the target
outputs (t). It is defind as –
N
mse = 1 (Ci ) 2 = 1 ∑ (t i − a i ) 2 …(1)
N N i=1
1784 K. I. Annapoorani: Forecasting the Nature of Power….
The regression plot represents the training, validation, testing and all data. The
dashed line in each axis represents the perfect result. i.e. Output = Target. The solid line
represents the best fit linear regression line between outputs and targets. The R value is an
indication of the relationship between the outputs and targets. If R = 1, this indicates that
there is an exact linear relationship between outputs and targets. If R is close to zero, then
there is no linear relationship between outputs and targets. From the above regression, plot
training data indicates a good fit. The scatter plot is helpful in showing that certain data
points have poor fit. If it is, it would investigate this data point to determine, if it represents
extrapolation. If so, then it should be included in the training set and additional data should
be collected to be used in the test data.
Table 4 shows that when sudden increase is there in load, beyond the rating due to
utility, the thermal stress get increased, which lead degradation of insulation and possibility
Int. J. Chem. Sci.: 13(4), 2015 1785
of arcing fault. When the loading reaches 50% of nameplate, the hot metal gases ethylene,
ethane and methane starts increases. If it was unpredicted, it leads to thermal fault of 300oC
to 700oC with high and low energy discharges. When the temperature is greater than 700oC,
the breakdown of oil occurs and produces acetylene. It causes a sustained arcing, a more
serious operational issue that can lead transformer failure, if left unpredicted. So by
predicting the dissolved gas data by regression method using Neural Network, we can
deduce the health of the transformer by finding the type of fault will occur in future.
CONCLUSION
The important need for condition evaluation is that there is an aging problem of
transformers installed in industries and transmission and distribution of power system
network. Most of the transformers were installed 20 to 30 years ago, when large investments
were made in expanding electrical power system network. These transformers have been
exposed to various accumulative worsening stresses and these are in high risk of failure. So
there is increase in need of reburshment, repair and replacement. Based on this view, the
predictive analysis of transformer will enable the power system network to give reliable
availability of power supply.
REFERENCES