CN106446440A - Short-term photovoltaic generation power prediction method based on online sequential extreme learning machine - Google Patents

Abstract

The invention relates to a short-term photovoltaic generation power prediction method based on online sequential extreme learning machine; an online sequential extreme learning machine with forgetting mechanism is employed, and moment, temperature and illumination intensity are selected as inputs to a prediction model; the method comprises the steps of generating an initial hidden layer output matrix of the extreme learning machine according to input historical data, and calculating initial output weight; predicting photovoltaic generation power, and waiting for weather measurement data and power measurement data; after the arrival of the data awaited, calculating error evaluation indexes to obtain a prediction error, storing historical data, generating a new hidden layer output matrix, updating the output weight and photovoltaic generation predicted power, and continuing to read weather forecast data. The method of the invention can provide improved prediction precision.

Description

Short-term photovoltaic power generation power prediction method based on online sequential extreme learning machine

Technical Field

The invention belongs to the field of photovoltaic power generation power prediction.

Background

The research of the solar photovoltaic power generation system has great theoretical and practical significance for relieving energy problems and environmental problems, improving the energy consumption structure, improving the performance of the distributed power generation system and developing the photovoltaic power generation industry. While photovoltaic power generation has many advantages, it is affected by a number of factors, with its power output being fluctuating, random, and intermittent. When the photovoltaic power generation system is operated in a grid-connected mode, the fluctuation of the photovoltaic power generation system can impact the stability of the power system, the safety of a power grid is threatened, and the difficulty of power system scheduling is greatly increased.

From the viewpoint of prediction, there are physical methods and statistical methods. The physical method firstly predicts factors (such as radiation intensity and photovoltaic panel temperature) directly influencing power output, and then inputs the prediction result into a physical model of the photovoltaic system to obtain output power. The statistical method does not need to analyze a specific physical model, and establishes a statistical model through historical data to directly predict the output power. The statistical methods commonly used at present are: support vector machines, artificial neural networks, gray-markov chains, etc.

Disclosure of Invention

The invention aims to provide a short-term photovoltaic power generation power prediction method to reduce the impact of the short-term photovoltaic power generation power on a power grid, and is beneficial to a dispatching department to reasonably arrange a dispatching plan and timely adjust the operation mode of a power system. The technical scheme of the invention is as follows:

a short-term photovoltaic power generation power prediction method based on an online sequential extreme learning machine adopts the online sequential extreme learning machine with a forgetting mechanism, and is characterized in that time, temperature and illumination intensity are selected as input quantities of a prediction model. The method comprises the following steps:

step 1: generating an initial hidden layer output matrix of the extreme learning machine according to the input historical data, and calculating an initial output weight;

step 2: predicting the photovoltaic power generation power, and waiting for meteorological measurement data and power measurement data;

and step 3: after the waiting data arrive, calculating an error evaluation index, obtaining a prediction error, storing historical data, generating a new hidden layer output matrix H matrix, updating the output weight beta and the photovoltaic power generation prediction power P, and continuously reading weather forecast data;

and 4, step 4: go back to step 2.

The error evaluation index may be as follows:

(1) normalized mean square error nRMSE of

(2) Mean absolute percent error MAPE of

Wherein n is the number of samples in the power generation period of the photovoltaic power station, P_ratedAt rated power, P_piPredicted power for i period, P_miIs the actual power of the i period.

The short-term photovoltaic power prediction algorithm based on the on-line sequential extreme learning machine (FOS-ELM) with the forgetting mechanism continuously introduces new data, eliminates the influence of outdated data, and performs short-term prediction on photovoltaic power generation power through historical weather data, historical photovoltaic power data and weather forecast data. Simulation examples show that the method has the characteristics of high training speed and high prediction precision. The FOS-ELM-based short-term photovoltaic prediction algorithm is beneficial to a scheduling department to reasonably arrange a scheduling plan, provides support for space-time complementation and coordination control of various power supplies, and has important significance for ensuring the safety and stability of a system and promoting the optimal operation of a power grid.

Drawings

FIG. 1 extreme learning machine architecture

FIG. 2 FOS-ELM prediction model flow chart

FIG. 3 (a), (b) and (c) are the comparison of the predicted value and the actual value in sunny days of the three methods

FIG. 4 (a), (b) and (c) are the predicted values and actual values in rainy days for the three methods, respectively

Detailed Description

(1) Extreme learning machine

The network structure of the Extreme Learning Machine (ELM) is shown in fig. 1, and the basic algorithm is as follows.

Assuming that the feedforward neural network model has L hidden layer nodes, the activation function is G (-. cndot.) for N different learning samples (x, y), x ∈ R^d*N,y∈R^N,a_i∈R^1*d，b_i∈R，a_i、b_iAre both randomly generated matrixes and vectors, and the ELM expression is shown as the formula.

G (-) is an activation function, is a weight vector connecting the ith node and the output node, is the number of hidden nodes, and G (-) can be any infinite differentiable function. As shown by Sigmoid function.

The formula (1) is rewritten into a matrix form as shown in the formula,

H·β＝Y \*MERGEFORMAT(3)

wherein,

β＝[β₁β₂... β_L]^TY＝[y₁y₂... y_N]^T

the least squares solution of equation (5) is shown.

The matrix H is a hidden layer output matrix of the ELM, the ith row H of the matrix is a hidden layer output vector relative to the input, and the output weight beta is the only parameter to be trained and determined.

In order to improve the stability and generalization capability of the result, a regularization parameter C is added, as shown in the following formula:

(2) online sequential extreme learning machine with forgetting mechanism

In many practical applications, the training data is not only a batch (fixed or variable batch size) or one arrival, but is also usually time-efficient, i.e. the data is valid only for a period of time, so that, in the learning process of the online sequence learning algorithm, outdated training data that has failed after a few units of time should be discarded, which is the idea of forgetting effect. The timeliness of online training data cannot be reflected by OS-ELM alone, and in this part, we add a forgetting mechanism and gradually exclude outdated data which may lead to wrong information. In the photovoltaic prediction system, because illumination and temperature change along with seasonal changes, and training data is only effective in one season, an Online sequential Extreme Learning machine (FOS-ELM) with a forgetting mechanism considers the timeliness of the data compared with the ELM, and is more suitable for the photovoltaic prediction system. The FOS-ELM algorithm is as follows.

First, initialization.

Step 1, use a small set of training dataAs initial data

Random generation of a_j、b_jj＝1,2...,L

A) Computing an initial hidden layer output matrix H₀

B) The estimated initial output weight is as shown.

β₀＝P₀H₀ ^TY₀\*MERGEFORMAT(6)

Wherein

C) Setting k to 0;

second, online learning with forgetting effect

Suppose that the k +1 th batch of data arrives

A) Computing a local hidden layer output matrix H_k+1

B) Calculate output weights β according to the formula^(k+1)、P_k+1

(3) Photovoltaic prediction model

The output power of the photovoltaic array per unit area is shown as the formula.

P_s＝ηSI[1-0.005(t₀-25)]\*MERGEFORMAT(11)

Wherein η is the conversion efficiency of the photovoltaic array, S is the array area, I is the illumination intensity, t₀Is at atmospheric temperature.

From the formula, the power output of the photovoltaic array is related to the conversion efficiency, the area, the illumination intensity and the atmospheric temperature of the photovoltaic array.

For a given photovoltaic array, the conversion efficiency and the area of the photovoltaic array are fixed, the numerical values of the photovoltaic array are hidden in historical data, and the solar illumination intensity is changed periodically along with time, so that the historical power of the moment, the temperature, the illumination intensity and the previous two moments is selected as the input quantity of a prediction model, and the obtained input vector is shown as the formula.

x_i＝[time tem I]^T\*MERGEFORMAT(12)

Wherein, time is a time value, if 06:00, the variable time should be 0600; if 06:15, the variable time should be 0625. tem is the atmospheric temperature in degrees celsius.

In the input vector shown in the formula, the quantity and the steel of each data are not completely the same, so normalization processing is required, and the method is shown in the formula.

Wherein x is_iFor inputting or outputting data, x_max、x_minThe maximum and minimum values of the data variation range are respectively.

The entire prediction algorithm proceeds according to the flow shown in fig. 2.

First, initial history data is input and initialized according to the formula writing mode to generate initial H₀Matrix, computing initial output weights β₀And P₀The method comprises the steps of inputting weather forecast data of the next moment including illumination intensity and temperature, calculating an output result, namely a prediction result, waiting for weather measurement data and power measurement data, calculating a prediction error according to a calculation method of a prediction error evaluation index after the data arrives, storing historical data, judging whether the time is over an hour, continuing to read the weather forecast data for prediction if the time is over the hour, and preprocessing the stored historical data within the hour if the time is over the hour to generate a new H matrix, updating output weights β and P, and continuing to read the weather forecast data for prediction.

To measure the prediction accuracy, the following error evaluation indexes were introduced.

Normalized Root Mean Square Error (nRMSE) is shown in equation.

The Mean Absolute Percent Error (MAPE) is shown in the formula.

Wherein n is photovoltaic powerNumber of samples, P, of station power generation period_ratedAt rated power, P_piPredicted power for i period, P_miIs the actual power of the i period.

A photovoltaic monitoring experiment of Oregon university in America is adopted to publish 5kW photovoltaic power data and weather data located in Ashland on a website of the photovoltaic monitoring experiment, and establishment and validity verification of a short-term power prediction model based on the FOS-ELM photovoltaic power generation system are carried out.

Model 1(FOS-ELM model): and (3) selecting a sigmond function as an activation function, using data which is 24 hours (only data of 6:00-18: 00) before the predicted moment and is every 15 minutes as training data, predicting the photovoltaic output at the moment, updating the training data once every hour, and eliminating the data before 24 hours, namely performing online learning calculation with a forgetting effect.

Model 2(OS-ELM model): and (3) selecting a sigmoid function as an activation function, using data every 15 minutes 24 hours (only data of 6:00-18: 00) before the predicted moment as training data, predicting the photovoltaic output at the moment, and updating the training data once every hour, namely performing calculation of online learning.

Model 3(ELM model): and (3) selecting a sigmoid function as an activation function, and taking data of the last 24 hours (only data of 6:00-18: 00) of the month of the predicted day as training data every 15 minutes to predict the photovoltaic output of each moment of the month. I.e., the ELM is retrained every other month.

The regularization parameter in the model is taken to be 1000 and the number of hidden layer nodes is taken to be 200.

The predicted results of three models on a certain sunny day in 1 month in 2015 are taken, and the predicted power and the actual power are shown in fig. 3.

Where the horizontal axis represents time and 600 represents 6: 00.

As can be seen from fig. 3, in a sunny day, the prediction accuracy is high, the error is small, and the prediction effect is good. Calculating the predicted nRMSE of the model 1 in the day to be 0.023 and the predicted MAPE to be 9.707; model 2 predicted nRMSE of 0.035, MAPE of 10.893; model 3 predicted nRMSE to be 0.054 and MAPE 12.706.

The predicted power and the actual power are shown in fig. 4, which takes the prediction results of three models in 1 month in 2015 on a certain rainy day.

As can be seen from fig. 4, in rainy days, the cloud amount of the sky is large, which brings more uncertainty, but the prediction model can still make a more accurate prediction. Calculating the predicted nRMSE of model 1 in the day to be 0.067 and the predicted MAPE to be 13.833; model 2 predicted nRMSE to be 0.074 and MAPE 14.303; the predicted RMSE for model 3 was 0.082 and the MAPE was 15.112.

Further analyzing the accuracy of the algorithm, three algorithm tests were performed by taking 2015 data of 4 months (spring), 7 months (summer), 10 months (autumn), 1 month (winter) and four months, and comparing nRMSE, RMSE and MAPE comparison results are shown in table 1.

TABLE 1 comparison of prediction accuracy for three models

As can be seen from table 1, from the index nRMSE, the prediction accuracy in winter and summer is higher than that in spring and autumn because the weather changes more severely in spring and autumn and the weather conditions in winter and summer are more stable; model 1 is more accurate than model 2 than model 3. In view of the index MAPE, the accuracy in summer is higher than that in winter because the power generation power in winter is lower than that in summer; model 1 is more accurate than model 2 than model 3. As a whole, the prediction precision of FOS-ELM is higher than that of an OS-ELM model, and the prediction precision of OS-ELM is higher than that of ELM.

The run time for each step of the three models was calculated in MATLAB, with the results: the time required by initialization of the model 1 is about 0.095s, and the time of online learning each time is about 0.052 s; the time required for initializing the model 2 is the same as that of the model 1, and the time for online learning each time is about 0.049 s; the time required for model 3 to train once is approximately 0.076 s. FOS-ELM online learning saves about 30% of the time per retrain, and OS-ELM trains about 20% of the time per retrain.

The invention provides a short-term photovoltaic power prediction algorithm based on FOS-ELM, and compared with classical ELM and OS-ELM. Theoretical analysis and example simulation proves that the FOS-ELM added with the regularization parameters has the advantages of high training speed, strong generalization capability and high precision no matter the month average data of sunny days, cloudy days or different seasons.

Claims (2)

1. A short-term photovoltaic power generation power prediction method based on an online sequential extreme learning machine adopts the online sequential extreme learning machine with a forgetting mechanism, and is characterized in that time, temperature and illumination intensity are selected as input quantities of a prediction model. The method comprises the following steps:

step 1: generating an initial hidden layer output matrix of the extreme learning machine according to the input historical data, and calculating an initial output weight;

step 2: predicting the photovoltaic power generation power, and waiting for meteorological measurement data and power measurement data;

and step 3: after the waiting data arrive, calculating an error evaluation index, obtaining a prediction error, storing historical data, generating a new hidden layer output matrix H matrix, updating the output weight beta and the photovoltaic power generation prediction power P, and continuously reading weather forecast data;

and 4, step 4: go back to step 2.

2. The short-term photovoltaic power generation power prediction method according to claim 1, characterized in that the error evaluation index is as follows:

(1) normalized mean square error nRMSE of

$nRMSE=\frac{1}{P_{rated}}\sqrt{\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}{(P_{mi}-P_{pi})}^2}$

(2) Mean absolute percent error MAPE of

$MAPE=100\×\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}\frac{|P_{mi}-P_{pi}|}{P_{mi}}$

Wherein n is the number of samples in the power generation period of the photovoltaic power station, P_ratedAt rated power, P_piPredicted power for i period, P_miIs the actual power of the i period.