CN116451824B - Photovoltaic power prediction method based on neural network and honeypot optimization - Google Patents

Abstract

The invention relates to a photovoltaic power prediction method based on neural network and honeypot optimization, and provides a photovoltaic power generation prediction model based on a badger algorithm and an enhanced counter-propagation neural network aiming at the defect that random initial weight and bias seriously affect the performance of a BP neural network. The solar irradiance per hour and the ambient temperature of the peltier cell are used as input data for the predictive model, and the historical power generation per hour (power) is used as the expected output. HBAs are introduced in the BP neural network to optimize initial weights and biases. The simulation results of various indexes show that the prediction accuracy of the HBA-BP neural network is improved to a certain extent relative to a BP prediction model.

Description

A photovoltaic power prediction method based on neural network and honeypot optimization

Technical Field

The invention belongs to a power prediction algorithm for improving prediction accuracy and efficiency, and relates to a photovoltaic power prediction method based on neural network and honeypot optimization.

Background Art

Metaheuristic algorithms are widely used as one of the main techniques for obtaining optimal solutions in various complex problems in multiple engineering and scientific research fields. They are very useful in solving complex problems where deterministic algorithms are stuck in local optima because they have the versatility to find several local solutions, such as the optimal design of structural engineering problems in practical applications, renewable energy systems, deep neural network (DNNs) model optimization, and other applications. Metaheuristic algorithms are not perfect. One weakness is that the quality of the optimal solution depends on the number of parameter values and the stopping condition of the algorithm, which is usually determined by the number of iterations. Many metaheuristic algorithms have been developed in recent years. For example, particle swarm optimization algorithm (PSO), whale optimization algorithm (WOA), chameleon swarm algorithm (CSA), African vulture optimization algorithm (AVOA), honeypot algorithm (HBA), etc. Among them, the honeypot algorithm has a higher accuracy in solving the optimal solution. Back propagation (BP) neural network as a prediction algorithm has been applied to power generation prediction of photovoltaic (PV) systems, but the prediction accuracy in practical applications has always been a problem. Therefore, we combine the honeypot algorithm and the back propagation neural network to improve the control performance of the PV system. The results show that the prediction accuracy and efficiency are better than the traditional back propagation neural network model.

The honeypot optimization algorithm is different from other metaheuristic algorithms. Inspired by the intelligent foraging behavior of honey badgers, an effective search strategy for solving optimization problems is mathematically developed. Through controlled randomization technology, HBA can maintain sufficient population diversity even at the end of the search process (here, the population is the weight and bias, and the solution is the optimal weight and bias), while the (BP) neural network has a relatively low prediction accuracy in the power generation prediction of photovoltaic systems, which is caused by the random setting of the initial weights and biases. Therefore, there is still room for improvement in the BP neural network for power generation prediction of photovoltaic systems.

Summary of the invention

Technical issues to be solved

In order to avoid the shortcomings of the prior art, the present invention proposes a photovoltaic power prediction method based on neural network and honeypot optimization.

Based on the above analysis, the present invention can transform the BP neural network algorithm. By combining the advantages of the HBA algorithm, that is, considering the influence of the initial weight and bias, the performance of the algorithm can be improved so that the BP neural network algorithm can obtain better dynamic performance in the application of the PV system. HBA is used to adjust the weight and bias to improve the accuracy of the BP neural network algorithm. The specific contributions of this work are as follows:

(1) Based on the fact that BP neural network algorithm has been applied to the prediction of photovoltaic power generation, an optimization algorithm with enhanced accuracy (HBA-BP) is proposed to enhance the accuracy of power generation prediction in PV systems.

(2) Compare and analyze the prediction results of HBA-BP for photovoltaic power generation with those of BP.

(3) The proposed HBA-BP is evaluated using the performance indicators of the optimization design problem, including mean absolute error (MAE) and root mean square error (RMSE), and fully considering the impact of temperature variation of the PV system.

Technical Solution

A photovoltaic power prediction method based on neural network and honeypot optimization, characterized by the following steps:

Step 1: normalize three types of original data respectively, the three types of original data include light intensity y ₁ , ambient temperature y ₂ and power y ₃ , and obtain parameter data _yn after normalization of the three types of original data;

Wherein, y _n (n＝1,2,3) represents the normalized data (-1<y _n <1) corresponding to the original data y _m (m＝1,2,3); y _m,min and y _m,max are the lower and upper boundaries of the three categories of corresponding original data respectively;

After normalization, each type of data is randomly divided into two groups: training data and test data;

Step 2: Initialize the parameters of the BP neural network. The parameters are defined as a vector consisting of weights and biases:

x _i =lb _i +rand×(ub _i -lb _i ); rand∈[0,1]

Among them, ub _i is the upper boundary of the parameter, lb _i is the lower boundary of the parameter; rand is a random value in [0,1], and i represents the i-th parameter;

The three types of training data are input into the BP neural network, and the BP neural network outputs the power value corresponding to the parameter;

Step 3: Evaluate the power values under all parameters:

The minimum mean square error is the optimal parameter value, that is, the vector consisting of the corresponding weight and bias is saved in T _pos ;

Step 4: Calculate the intensity I, i.e. the parameter concentration:

Among them: the value is 2, I _i represents the parameter concentration, and d _i represents the distance between the parameter and other parameters;

Update the decrement factor α:

Where: t represents the current number of iterations, C is a constant, and t _max is the maximum number of iterations;

Step 5: Update the parameter x _new and return to step 3;

The parameter x _new is calculated as:

When R<0.5:

x _new ＝x _prey +βIFx _prey +r ₁ αd _i F|cos(2πr ₂ )×[1-cos(2πr ₃ )]|

When R ≥ 0.5:

x _new ＝x _prey +βIFx _prey +r ₄ αd _i F

Where x _new represents the new parameters, the ability to solve the best parameters is defined as β ≥ 1, R, r ₁ , r ₂ , r ₃ and r ₄ are independent random numbers between 0 and 1, and x _prey represents the current best parameters;

The F is used to change the direction of parameter value update:

Step 6: Train the weights and biases in _Tpos :

Where T _pos represents the best set of weights and biases, W _t+1 represents the weights of (t+1) training times, and the defined learning rate lr is also called the step size. Loss is a measure of the difference between the expected and predicted values in training, and W _t represents the weights of t training times;

Step 7: Substitute the weights and biases obtained from the above training into the following formula and output the predicted value Y _p :

Among them, w _ij and b _j are the weights and bias parameters from the input layer to the hidden layer nodes. w _jk and b _k represent the weights and biases from the hidden layer nodes to the output layer nodes, and Yp represents the predicted value, that is, the predicted power. is the activation function of the nodes in the hidden layer, is the activation function of the output layer nodes, m is the number of output layer nodes, n is the number of input layer nodes, i is the network node of the input layer, j is the network node of the hidden layer, and k is the network node of the output layer.

The two groups of data include 80% training data and 20% testing data.

The constant C has a value of 2.

The default value of β is 6.

Beneficial Effects

The present invention proposes a photovoltaic power prediction method based on neural network and honeypot optimization. In view of the imperfection that random initial weights and biases seriously affect the performance of BP neural network, a photovoltaic power generation prediction model based on honey badger algorithm and enhanced back propagation neural network is proposed. The hourly solar irradiance and ambient temperature of Perth are used as input data of the prediction model, and the historical hourly power generation (power) is used as the expected output. HBA is introduced into the BP neural network to optimize the initial weights and biases. The simulation results of various indicators show that the prediction accuracy of HBA-BP neural network is improved to a certain extent compared with the BP prediction model.

In the prior art, although the power generation of the PV system can be predicted by the BP neural network, the prediction accuracy will be affected by the BP structure; specifically, the initial weights and biases are randomly set, resulting in low prediction accuracy. The optimized photovoltaic power generation expectation model (HBA-BP) based on the integration of BP neural network and HBA proposed in the present invention will use HBA to adjust the weights and biases. Then, based on the data simulation experiment in Perth, the comparison of the prediction results of HBA-BP and BP models shows that HBA can improve the performance of BP neural network to obtain more accurate results.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1: The method flow is shown in Figure 1: (where t represents the current iteration number and t _max represents the maximum iteration number.)

Figure 2: Power prediction comparison test diagram

Figure 3: Daily power forecast comparison

Figure 4: Prediction errors of BP and HBA-BP

DETAILED DESCRIPTION

The present invention will now be further described with reference to the embodiments and the accompanying drawings:

The method flow of the embodiment is shown in FIG1 : (where t represents the current iteration number, and t _max represents the maximum iteration number.)

(1) Prepare and normalize the data and divide it into training and testing sets.

Where _yn (n=1,2,3) represents the normalized parameter data corresponding to the original parameter _ym (m=1,2,3) (-1< _yn <1). _y1 , _y2 , _y3 are the original parameters ( _y1 is the light intensity; _y2 is the ambient temperature; _y3 is the power). _ym,min and _ym,max are the lower and upper boundaries of the original parameters, respectively. The normalized parameter data are then randomly divided into two groups, namely 80% of the training data and 20% of the test data.

(2) Initialization parameters (the parameters below refer to the vector composed of weights and biases) (the i below represents the i-th parameter)

x _i =lb _i +rand×(ub _i -lb _i ); rand∈[0,1] (2)

Among them, ub _i is the upper boundary of the parameter, lb _i is the lower boundary of the parameter; rand is a random value in [0,1].

(3) Evaluate the power values under all parameters (weights and biases) and save the best parameter values (weights and biases) to T _pos , that is, the weights and biases corresponding to the minimum mean square error are the best parameters.

Where Fitness _i is the mean square error (MSE) of the ith power, P _p and _Pa represent the predicted value and true value of the PV system output power, respectively. d refers to the ith power, and M is the total number of power values.

(4) Update the decrement factor α and calculate the intensity I.

Among them, C is a constant with a value of 2, _Ii represents the parameter concentration (the parameter here refers to the vector composed of weights and biases), _di represents the distance between the parameter and other parameters, and α is a time-varying parameter.

(5) Update the parameters (weights and biases) by the following formula and return to step (3) for evaluation and comparison with the original T _pos . If the mean square error under x _new is smaller, x _new is assigned to T _pos .

When R<0.5:

x _new ＝x _prey +βIFx _prey +r ₁ αd _i F|cos(2πr ₂ )×[1-cos(2πr ₃ )]| (5a)

When R ≥ 0.5:

x _new ＝x _prey +βIFx _prey +r ₄ αd _i F (5b)

Where x _new represents the new parameters (weights and biases). The ability to find the best parameters can be defined as β ≥ 1 (default = 6). R, r ₁ , r ₂ , r ₃ and r ₄ are independent random numbers between 0 and 1, and x _prey represents the current best parameters. As a flag, F is used to change the direction of parameter value update (the BP training process includes forward propagation and back propagation), as shown below:

(6) Train the weights and biases within _Tpos .

Where T _pos represents the best set of weights and biases, W _t+1 represents the weights of (t+1) training times, and the defined learning rate lr is also called the step size. Loss is a measure of the difference between the expected and predicted values in training, and W _t represents the weights of t training times.

(7) After substituting the weights and biases obtained from the above training into the following formula, simulation and prediction are performed, and the predicted value _Yp output is compared with _y3 .

Among them, w _ij and b _j are the weights and bias parameters from the input layer to the hidden layer nodes. w _jk and b _k represent the weights and biases from the hidden layer nodes to the output layer nodes, and Yp represents the predicted value, that is, the predicted power. is the activation function of the nodes in the hidden layer, is the activation function of the output layer nodes, m is the number of output layer nodes, n is the number of input layer nodes, i is the network node of the input layer, j is the network node of the hidden layer, and k is the network node of the output layer.

In the implementation example

(1) Parameter settings:

In order to clearly demonstrate the scientificity and reliability of the proposed network optimization, 400 sets of data were used to train the BP neural network and the HBA-BP neural network respectively. The remaining 100 sets of data were used as test sets to verify the performance of the two algorithms.

Table 1. Parameter setting table

parameter Numeric parameter Numeric Input Layer 1 Dataset size 30 Output layer nodes 1 Maximum number of iterations 100 Hidden Layer 1 Dimensions of the dataset 17 Hidden layer nodes 4 Data settings 500 Output Layer 1 Learning Rate 0.1 Output layer nodes 1 Number of training sessions 100

(2) When the hourly average solar irradiance and ambient temperature are used as input and the hourly power generation is the output expectation, the prediction results and overall error are shown in Figure 2. The power prediction results of the HBA-BP model are presented together with the actual power generation. This shows that the prediction performance of the HBA-BP model is better than that of the BP model.

(3) In order to obtain more obvious performance of BP and HBA-BP, the 24-hour prediction results are shown in Figure 3;

As shown in Figure 3, the daily power forecast results of the HBA-BP and BP models are presented together with the actual power generation. The prediction of HBA-BP is closer to the actual value than that of BP, which shows that the prediction performance of the HBA-BP model is better than that of the BP model. That is, the predicted results will be more accurate.

(4) The prediction errors of the BP model and the HBA-BP model are compared in Figure 4. The prediction error of BP is significantly higher than that of HBA-BP, so HBA-BP exhibits better nonlinear fitting ability than the BP neural network prediction model.

(5) In order to more accurately and effectively verify the superiority of the HBBA-BP prediction model, the mean absolute error (MAE) and root mean square error (RMSE) were applied to evaluate the performance of the HBA-BP prediction model.

The quantitative evaluation of the BP neural network prediction model and the HBA-BP prediction model is shown in Table 2.

Table 2. Performance comparison between HBA-BP prediction model and BP prediction model

Performance indicators BP Model HBA-BP model Mean Square Error (MSE) 305.15*10 ^-6 5.585*10 ^-6 Mean Absolute Error (MAE) 1.846 0.264 Root mean square error (RMSE) 0.00175 0.000236

The results show that after the introduction of HBA into the BP prediction model, the three indicators all dropped significantly. The RMSE value dropped dramatically from 0.00175 to 0.000236, a decrease of 7.5 times, which reflects the superior prediction performance and accuracy of the HBA-BP prediction model, and will play a more excellent role in practical applications.

In view of the fact that the performance of BP neural network is affected by random weights and biases, we proposed a photovoltaic power generation prediction model based on honey badger algorithm and enhanced back propagation neural network to optimize weights and biases to obtain the best effect, thereby improving the prediction performance of neural network, that is, improving its prediction accuracy. The simulation results of various indicators after optimization show that the prediction accuracy of HBA-BP neural network is greatly improved compared with BP prediction model, showing excellent prediction effect, which greatly improves the real-time and accuracy requirements for PV system control.

Claims (4)

1. A photovoltaic power prediction method based on neural network and honeypot optimization, characterized by the following steps: Step 1: normalize three types of original data respectively, the three types of original data include light intensity y ₁ , ambient temperature y ₂ and power y ₃ , and obtain parameter data _yn after normalization of the three types of original data; Wherein, _yn (n=1, 2, 3) represents the normalized data (-1＜ _yn ＜1) corresponding to the original data _ym (m=1, 2, 3); _ym,min and _ym,max are the lower and upper boundaries of the three categories of corresponding original data respectively; After normalization, each type of data is randomly divided into two groups: training data and test data; Step 2: Initialize the parameters of the BP neural network. The parameters are defined as a vector consisting of weights and biases: x _i =lb _i +rand×(ub _i -lb _i ); rand∈[0, 1] Where ub _i is the upper boundary of the parameter, lb _i is the lower boundary of the parameter; rand is a random value in [0, 1], and i represents the i-th parameter; The three types of training data are input into the BP neural network, and the BP neural network outputs the power value corresponding to the parameter; Step 3: Evaluate the power values under all parameters: The minimum mean square error is the optimal parameter value, that is, the vector consisting of the corresponding weight and bias is saved in T _pos ; Step 4: Calculate the intensity I, i.e. the parameter concentration: Among them: the value is 2, I _i represents the parameter concentration, and d _i represents the distance between the parameter and other parameters; Update the decrement factor α: Where: t represents the current number of iterations, C is a constant, and t _max is the maximum number of iterations; Step 5: Update the parameter x _new and return to step 3; The parameter x _new is calculated as: When R＜0.5: x _new ＝x _prey +βIFx _prey +r ₁ αd _i F|cos(2πr ₂ )×[1-cos(2πr ₃ )]| When R ≥ 0.5: x _new ＝x _prey +βIFx _prey +r ₄ αd _i F Where x _new represents the new parameters, the ability to solve the best parameters is defined as β ≥ 1, R, r ₁ , r ₂ , r ₃ and r ₄ are independent random numbers between 0 and 1, and x _prey represents the current best parameters; The F is used to change the direction of parameter value update: Step 6: Train the weights and biases in _Tpos : Where T _pos represents the best set of weights and biases, W _t+1 represents the weight of (t+1) training times, and the defined learning rate lr is also called the step size; Loss is a measure of the difference between the expected and predicted values in training, and W _t represents the weight of t training times; Step 7: Substitute the weights and biases obtained from the above training into the following formula and output the predicted value Y _p : Among them, w _ij and b _j are the weights and bias parameters from the input layer to the hidden layer nodes; w _jk and b _k represent the weights and biases from the hidden layer nodes to the output layer nodes, and Yp represents the predicted value, that is, the predicted power; is the activation function of the nodes in the hidden layer, is the activation function of the output layer nodes, m is the number of output layer nodes, n is the number of input layer nodes, i is the network node of the input layer, j is the network node of the hidden layer, and k is the network node of the output layer. 2. The photovoltaic power prediction method based on neural network and honeypot optimization according to claim 1 is characterized in that: the training data accounts for 80% and the test data accounts for 20% of the two groups of data. 3. According to the photovoltaic power prediction method based on neural network and honeypot optimization in claim 1, it is characterized in that the constant C is 2. 4. The photovoltaic power prediction method based on neural network and honeypot optimization according to claim 1 is characterized in that: The default value of β=6.