A photovoltaic power prediction method

Abstract

A photovoltaic power prediction method relates to the field of photovoltaic power prediction. The present invention is to solve the problems of existing photovoltaic power prediction methods, such as complex calculations, poor accuracy, susceptibility to noise and overfitting, and difficulty in model integration. The photovoltaic power prediction method described in the present invention clusters weather forecast data and photovoltaic power data within a historical time period to obtain the weather type to which the data belongs; performs modal decomposition on the weather forecast data and photovoltaic power data within the historical time period; inputs the modally decomposed data into a photovoltaic power prediction model under the corresponding weather type according to the weather type to which it belongs, obtains a photovoltaic power prediction curve, and uses the photovoltaic power prediction curve to predict future photovoltaic power.

Description

Technical Field

[0001] The invention belongs to the field of photovoltaic power prediction. Background Art

[0002] In recent years, solar power generation technology has developed rapidly, and the use of renewable energy has increased, and the proportion of non-fossil energy will rise to half of primary energy. Since photovoltaic power generation is affected by factors such as weather, time, and season, accurate prediction of power output can help the power grid optimize energy scheduling, balance supply and demand, reduce the demand for backup power, and reduce operation and maintenance costs. This is crucial to achieving efficient use of renewable energy and ensuring the stability of power supply.

[0003] At present, the methods for photovoltaic power prediction include modeling analysis, data-driven method and modeling data fusion method. However, the modeling analysis method is more complicated in calculation and lacks accuracy; the data-driven method relies on a large amount of historical data and is easily affected by data noise and overfitting problems; although the modeling data fusion method combines multiple methods, the integration and optimization of the model is difficult and not easy to implement. Summary of the invention

[0004] The present invention aims to solve the problems of existing photovoltaic power prediction methods, such as complex calculation, poor accuracy, susceptibility to noise and overfitting, and difficulty in model integration. A photovoltaic power prediction method is provided.

[0005] A photovoltaic power prediction method, comprising:

[0006] Cluster the weather forecast data and photovoltaic power data in the historical time period to obtain the weather type to which the data belongs;

[0007] Perform modal decomposition on weather forecast data and photovoltaic power data within the historical time period;

[0008] The modal decomposed data is input into the photovoltaic power prediction model under the corresponding weather type according to the corresponding weather type to obtain the photovoltaic power prediction curve, and the photovoltaic power prediction curve is used to realize the prediction of future photovoltaic power.

[0009] Furthermore, the above clustering of weather forecast data and photovoltaic power data within the historical time period includes:

[0010] Initialization: Randomly generate the initial annealing temperature, and the initial population gen0 includes the clustering features of weather forecast data and photovoltaic power data in the historical time period;

[0011] S1: At the annealing temperature T, for the population gen at the g-th iteration g Perform selection, crossover and mutation operations to generate a new population gen g ′;

[0012] S2: Use FCM clustering algorithm to calculate the new population gen g ′, each cluster center is a weather type;

[0013] S3: Determine whether the number of iterations g reaches the maximum number of clustering iterations. If yes, execute S4; otherwise, execute S5.

[0014] S4: Determine whether the maximum number of annealing iterations has been reached, if yes, output the weather type, otherwise execute S6;

[0015] S5: Using simulated annealing algorithm to generate g ′ generates new individuals and adds them to the new population gen g ′, we obtain the population gen under the g+1th iteration g+1 , make g=g+1 and then return to S1;

[0016] S6: Perform annealing operation on the simulated annealing temperature T using the simulated annealing algorithm. After updating the temperature T, set g=0 and then return to S1.

[0017] Furthermore, in the above S2, the new population gen is calculated g After finding the cluster centers of ′, the cluster centers are modified, including:

[0018] Calculate the population gen g The distance from each sample point to each cluster center in ′;

[0019] Identify sample points whose distances meet the preset isolated point threshold as isolated points;

[0020] A weight is assigned to each sample point, where the weight of the isolated point is smaller than that of the non-isolated point, so that the cluster center is corrected.

[0021] Furthermore, the mean, standard deviation, maximum value, number of peaks and troughs, coefficient of variation, kurtosis and skewness of the total horizontal radiation variable of weather forecast data were selected as clustering features.

[0022] Furthermore, the expressions of the above coefficient of variation B, kurtosis F and skewness S are:

[0023]

[0024] Where Y is the total horizontal radiation variable, σ is the standard deviation of Y, is the mean value of Y, m is the mth total horizontal radiation variable sample, and M is the total number of samples of Y.

[0025] Furthermore, the above-mentioned modal decomposition of weather forecast data and photovoltaic power data in the historical time period includes:

[0026] The weather forecast data and photovoltaic power data in the historical period are used as the data to be decomposed D(t), and a white noise signal is added to the data to be decomposed D(t) to obtain the data containing noise D(T);

[0027] Decomposing the data D(T) using the ICEEMDAN algorithm to obtain a set of modal components;

[0028] The PCA algorithm is used to calculate the correlation coefficient of each modal component and arrange them in descending order;

[0029] Delete the modal component with the smallest correlation coefficient, merge the remaining modal components, and output the data after modal decomposition.

[0030] Furthermore, the photovoltaic power prediction model includes: convolutional neural network CNN and BiLSTM network,

[0031] The convolutional neural network CNN is used to extract spatial features from input data.

[0032] The BiLSTM network is used to capture the temporal dependency of the spatial features, identify long-term dynamic changes, and obtain a photovoltaic power prediction curve.

[0033] Furthermore, the above-mentioned convolutional neural network CNN and BiLSTM network are networks whose hyperparameters are adjusted using the Bayesian algorithm.

[0034] Furthermore, the calculation process of the above BiLSTM network includes:

[0035] h t =LSTM(s t ,h t-1 ),

[0036] h t ′=LSTM(s t ,h t ' -1 ),

[0037] o t =W ho h t +W h'o h t ′+z o ′,

[0038] Among them, h t and h t′ are the hidden states of the forward and backward LSTM layers of the BiLSTM network, LSTM(·) represents the LSTM operation, and W ho and W h′o denote the weights of the forward and backward LSTM layers, respectively, and z o ′ is the bias of the BiLSTM output layer, o t is the output result of the BiLSTM output layer, s t The input data for the BiLSTM input layer.

[0039] Furthermore, the above convolutional neural network CNN can be expressed as:

[0040] W1=RELU(X*K1+b1),

[0041] H1=max(W1)+b2,

[0042] W2=RELU(H1*K2+b3),

[0043] H2=max(W2)+b4,

[0044] y C =Sigmoid(H2×K3+b5),

[0045] Where W1 and W2 are the outputs of the two convolutional layers in the convolutional neural network CNN, H1 and H2 are the outputs of the two pooling layers in the convolutional neural network CNN, * represents the convolution operation, X is the input of the convolutional neural network CNN, K1, K2, K3 are all weight matrices, b1, b2, b3, b4 and b5 are all bias coefficients, y C is the output of the convolutional neural network CNN.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The cluster center correction strategy based on GASA-FCM hybrid clustering and isolated point variable weight enhances the global and local search capabilities and reduces the impact of noise and isolated points on clustering results; the improved adaptive noise complete set empirical mode decomposition can effectively eliminate the impact of randomness and volatility in data information on prediction accuracy; the introduced Bayesian optimization algorithm can better mine the spatiotemporal characteristics of data, improve the overall prediction accuracy of the model, and ensure the safe, stable and economical operation of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the flow chart of GASA-FCM;

[0049] Figure 2 This is the flowchart of ICEEMDAN;

[0050] Figure 3 is a flow chart of a photovoltaic power prediction method;

[0051] Figure 4 The model prediction accuracy under different weather conditions in spring, (a) the model prediction result of sunny weather in spring, (b) the model prediction result of sunny to cloudy weather in spring, (c) the model prediction result of cloudy weather in spring, (d) the model prediction result of rainy weather in spring;

[0052] Figure 5 The model prediction accuracy under different weather conditions in autumn, (a) the model prediction results for sunny weather in autumn, (b) the model prediction results for sunny to cloudy weather in autumn, (c) the model prediction results for cloudy weather in autumn, and (d) the model prediction results for rainy weather in autumn. DETAILED DESCRIPTION

[0053] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0054] Reference Figures 1 to 3 Specifically describe this embodiment, a photovoltaic power prediction method based on similar day clustering analysis based on genetic simulated annealing optimized fuzzy C-means clustering algorithm (GASA-FCM) and principal component analysis optimized adaptive noise complete set empirical mode Bayesian optimized convolutional neural network-bidirectional long short-term memory network model (ICEEMDAN-CNN-Bys-BiLSTM) described in this embodiment includes:

[0055] Step 1: Clustering of NWP (weather forecast value) data and photovoltaic power data, including:

[0056] (1) Select clustering features

[0057] Photovoltaic power data includes: photovoltaic power output by the photovoltaic power station and solar panel temperature.

[0058] NWP data include: wind speed, solar radiation intensity, ambient temperature and ambient humidity.

[0059] The mean, standard deviation, maximum value, peak-to-trough number, coefficient of variation, kurtosis and skewness of the total horizontal radiation variable of NWP data are selected as clustering features. The expressions of the coefficient of variation B, kurtosis F and skewness S are:

[0060]

[0061]

[0062] In the formula, σ is the standard deviation of the total horizontal radiation variable, Y is the total horizontal radiation variable, is the average value of the total horizontal radiation variable, Y m is the mth total horizontal radiation variable sample, and M is the total number of samples of the total horizontal radiation variable.

[0063] The selected clustering features comprehensively consider multiple aspects of NWP data, such as central tendency, dispersion, extreme values, and distribution patterns, which helps to more comprehensively characterize weather characteristics for subsequent efficient clustering.

[0064] (2) Calculate cluster centers based on the hybrid clustering algorithm of genetic simulated annealing optimized FCM (GASA-FCM)

[0065] The FCM clustering algorithm (Fuzzy C-means clustering) integrates fuzzy theory and realizes flexible fuzzy partitioning. The sample set of the FCM clustering algorithm is denoted as Q = (q1, q2, ..., q J ), the task is to divide it into I fuzzy groups and calculate the cluster center c of each group i , in order to achieve the purpose of minimizing the objective function, the objective function is:

[0066]

[0067] Among them, u ij is the membership degree of the jth sample to the ith class, and:

[0068]

[0069] i=1,2,...,I,k=1,2,...,I,j=1,2,...,J,U is the membership matrix, d ij is the distance from the jth sample to the i-th cluster center, d kj is the distance from the jth sample to the kth cluster center, l is the membership factor, and J is the total number of samples.

[0070] The Lagrange multiplier λ is introduced to construct a new objective function to constrain the objective function.

[0071]

[0072] The genetic simulated annealing algorithm combines the local search capability of the simulated annealing algorithm with the global search capability of the genetic algorithm, and both the global and local search capabilities are improved. This algorithm not only strengthens the overall search process, but also reduces the requirements for parameters, showing significant advantages. Its steps are:

[0073] 1) Initialize the population: randomly generate a group of individuals, each of which represents a possible solution to the problem;

[0074] 2) Selection: According to fitness, select a part of individuals as parents to retain excellent solutions;

[0075] 3) Select two individuals from the parent generation and generate new individuals through crossover operation;

[0076] 4) Performing mutation operations on newly generated individuals helps to find new solutions in the search space;

[0077] 5) Update the current population based on the results of selection, crossover, and mutation;

[0078] 6) Repeat the steps of selection, crossover, mutation, and update until the stopping condition is met.

[0079] The simulated annealing operation performs local search through random search to avoid falling into the local optimal solution. The algorithm goal is to avoid the local optimal solution trap in the search space and thus find the global optimal solution.

[0080] Based on the above two methods, Figure 1 As shown in the figure, the design process of the FCM (GASA-FCM) hybrid clustering algorithm based on genetic simulated annealing optimization is as follows:

[0081] Initialization: Randomly generate the initial annealing temperature, and the initial population gen0 includes the clustering features of weather forecast data and photovoltaic power data in the historical time period;

[0082] S1: At the annealing temperature T, for the population gen at the g-th iteration g Perform selection, crossover and mutation operations to generate a new population gen g ′;

[0083] S2: Use FCM clustering algorithm to calculate the new population gen g ′, each cluster center is a weather type;

[0084] S3: Determine whether the current iteration number g reaches the maximum clustering iteration number, if yes, execute S4, otherwise execute S5;

[0085] S4: Determine whether the simulated annealing temperature T meets the annealing iteration termination condition (reaching the maximum number of annealing iterations), if yes, output the weather type, otherwise execute S6;

[0086] S5: Using simulated annealing algorithm to generate g ′ generates new individuals and adds them to the new population gen g ′, we obtain the population gen under the g+1th iteration g+1 , make g=g+1 and then return to S1;

[0087] S6: Perform annealing operation on the simulated annealing temperature T using the simulated annealing algorithm. After updating the temperature T, set g=0 and then return to S1.

[0088] (3) Modify the weighted cluster center of the data outliers

[0089] In order to compensate for the influence of noise points and isolated points on the estimation accuracy and avoid the significant shift of cluster centers due to the extreme values ​​of these points, a cluster center correction strategy based on isolated point variable weights is proposed to assign a weight to each cluster feature sample point. These weights reflect the influence of each cluster feature sample point on the cluster center position. Isolated points may be assigned lower weights to reduce their influence on the center position.

[0090] The specific steps are as follows:

[0091] 1) Calculate the population gen g The distance from each sample point to each cluster center in ′;

[0092] 2) Identify sample points whose distances meet the preset isolated point threshold as isolated points;

[0093] 3) Assign a weight to each sample point, where the weight of the isolated point is smaller than the weight of the non-isolated point, so that the cluster center is corrected.

[0094] In order to verify the superiority of the GASA-FCM hybrid clustering and cluster center correction strategy based on isolated point variable weight, it is compared with the GASA-FCM algorithm and the FCM algorithm, and the Davies-Bouldin index (DB index) is used as an evaluation to evaluate the clustering results. This index represents the ratio of the average distance within each cluster to the distance between the nearest clusters. The lower the value, the better the clustering effect, as shown in the following formula:

[0095]

[0096] Where: e is the average distance within cluster e, σ g is the average distance within cluster g, r(c e ,c g ) is the distance between cluster center e and cluster center g.

[0097] This step uses the selected clustering features as input variables for analysis and calculation, and uses the GASA-FCM algorithm for cluster analysis to obtain four weather types: sunny, sunny to cloudy, cloudy, and rainy. Different weather types have different effects on photovoltaic power generation. For example, on sunny days, the solar radiation is strong and the photovoltaic power generation is high; on rainy days, the solar radiation is weak and the photovoltaic power generation is low. By clustering into different weather types, a more accurate prediction model can be established for the photovoltaic power generation characteristics under different weather types. In this way, when predicting, the prediction model that best suits the current weather type can be selected based on real-time weather data, thereby improving the prediction accuracy.

[0098] Step 2: Perform modal decomposition on clustered data

[0099] Principal component analysis (PCA) is a technique for processing high-dimensional data features. It can remove complex noise while retaining key information to improve data processing performance. The steps of PCA are as follows:

[0100] 1) The original feature space data X = {x1, x2, ..., x n} Standardization, get the standardized X * .

[0101] 2) Solve for X * The covariance matrix of .

[0102] 3) For cov(X * ) performs eigenvalue decomposition to obtain its eigenvalues ​​and corresponding eigenvectors.

[0103] 4) Sort by eigenvalue size and construct an eigenvector matrix based on these eigenvalues.

[0104] 5) Convert the original data into a new space based on these eigenvectors.

[0105] The ICEEMDAN (Completely Adaptive Noise Ensemble Empirical Mode Decomposition) algorithm with adaptive addition of white noise can effectively reduce the modal artifact problem, but it still cannot completely eliminate the influence of redundant components and false components, which affects the selection of principal components. Therefore, this step uses a combination of principal component analysis and fractal dimension to improve ICEEMDAN. PCA can extract effective fault information through data dimensionality reduction, and use a cross analysis method to calculate the fractal dimension before and after dimensionality reduction to reconstruct the optimal component group. According to fractal theory, fractal dimension is positively correlated with signal stability. The improved ICEEMDAN algorithm flow chart is shown in the figure. Figure 2 The specific process is as follows:

[0106] SA: Collect NWP data and PV power data in the historical period as the data to be decomposed D(t);

[0107] SB: adding a white noise signal to the data to be decomposed D(t) to obtain data D(T) containing noise;

[0108] SC: Decomposing the data D(T) using the ICEEMDAN algorithm to obtain a set of modal components;

[0109] SD: PCA algorithm is used to calculate the correlation coefficient of each modal component and arrange them in descending order;

[0110] SE: Delete the modal component with the smallest correlation coefficient;

[0111] SF: Merge the remaining modal components and output the data after modal decomposition.

[0112] Step 3: According to the different weather types obtained by clustering in step 1, a CNN-Bys-BiLSTM photovoltaic power prediction model under different weather types is built, and an optimization algorithm is introduced to adjust the prediction model. The input data includes: photovoltaic power data and NWP data; the output is the photovoltaic power prediction curve.

[0113] (1) Use convolutional neural network (CNN) to extract spatial features from input data and identify key local patterns, which can be expressed as:

[0114] W1=RELU(X*K1+b1),

[0115] H1=max(W1)+b2,

[0116] W2=RELU(H1*K2+b3),

[0117] H2=max(W2)+b4,

[0118] y C =Sigmoid(H2×K3+b5),

[0119] Where W1 and W2 are the outputs of convolutional layer 1 and convolutional layer 2 in the convolutional neural network CNN, respectively; H1 and H2 are the outputs of pooling layer 1 and pooling layer 2 in the convolutional neural network CNN, respectively; * represents convolution operation; K1, K2, K3 are all weight matrices; b1, b2, b3, b4 and b5 are all bias coefficients; y C is the output feature vector of the CNN layer.

[0120] (2) The spatial features extracted by CNN are further processed using a BiLSTM (bidirectional long short-term memory) network to capture temporal dependencies and identify long-term dynamic changes.

[0121] In the BiLSTM network structure, the same input data is input to the forward LSTM layer and the backward LSTM layer to calculate the hidden state h of the forward LSTM layer. t and the hidden state h of the backward LSTM layer t ′. The forward computation is performed from time 1 to time t in the forward LSTM layer, and the backward computation is performed from time t to time 1 in the backward LSTM layer. At each time point, the output of the forward hidden state and the backward hidden state are obtained and saved. Then, the two hidden states are connected to calculate the output of the BiLSTM.

[0122] The calculation process of BiLSTM is shown as follows:

[0123] h t =LSTM(s t ,h t-1 ),

[0124] h t ′=LSTM(s t ,h t ' -1 ),

[0125] o t =W ho h t +W h'o h t ′+z o ′,

[0126] Among them, LSTM(·) represents the LSTM operation, W ho and W h′o Respectively represent the forward LSTM weight and the backward LSTM weight, z o ′ is the bias of the BiLSTM output layer, o t Output result of BiLSTM output layer, s t Input data to the BiLSTM input layer.

[0127] (3) The Bayesian (Bys) algorithm is used to improve the model performance by adjusting the hyperparameters of CNN and BiLSTM to form an efficient integrated prediction model.

[0128] The CNN model is used to deeply mine the spatial characteristics of photovoltaic power data and the seasonal characteristics of NWP (weather forecast value) data. Then the expected improvement, improvement probability and upper confidence bound parameters of the Bayesian optimization BiLSTM model are used. These parameters allow Bayesian optimization to explore various areas of the search space, sample points that may improve the optimization, avoid sampling points that are unlikely to improve the optimization, and perform deep mining to obtain the time correlation of the input data sequence, thereby accurately obtaining the spatiotemporal distribution characteristics of photovoltaic power data and NWP data, and effectively improving the prediction accuracy of the BiLSTM model. CNN is used to deeply mine the spatial and seasonal characteristics of the input data, and Bayesian optimization BiLSTM is used to obtain the time correlation of the input data sequence. The photovoltaic power prediction model of CNN-Bys-Bilstm is constructed to achieve accurate prediction of photovoltaic power.

[0129] The accuracy of the photovoltaic power prediction model is evaluated using the mean absolute error (MAE) and root mean square error (RMSE), and the relevant formula is:

[0130]

[0131]

[0132] Where N represents the data point for photovoltaic power prediction; P true is the actual value of photovoltaic power; P fore is the predicted value of photovoltaic power; P cap is the total installed capacity of the photovoltaic power station. MAE It is the ratio of mean absolute error to installed capacity, and is usually expressed in percentage in practical applications. RMSE is the ratio of the root mean square error to the installed capacity.

[0133] (4) Predicting PV power under actual operating conditions

[0134] The prediction results of the ICEEMDAN-CNN-Bys-BiLSTM model, CNN-Bys-BiLSTM model, CNN-BiLSTM model, and CNN-GRU model are compared and analyzed.

[0135] Figure 4 and Figure 5The photovoltaic power prediction results of various prediction models under four weather conditions in spring and autumn. As can be seen from the figure, each model can accurately predict the photovoltaic power change trend under different seasons and different weather conditions, proving that the machine learning method has good prediction performance in photovoltaic power prediction. In addition, the ICEEMDAN-CNN-Bys-BiLSTM prediction model proposed in the present invention can more accurately describe the distribution characteristics of photovoltaic power compared with the other prediction models, proving that the prediction performance of the ICEEMDAN-CNN-Bys-BiLSTM model proposed in the present invention is better than that of the other models.

[0136] Table 1 shows the average prediction error of each prediction model under different seasonal conditions (average prediction error under four weather conditions), where the last row is the average prediction error of all seasons. As shown in Table 1, among all prediction models, the average prediction error of the ICEEMDAN-CNN-Bys-BiLSTM model is the smallest, indicating that the prediction accuracy of the ICEEMDAN-CNN-Bys-BiLSTM model is higher than that of other prediction models, proving the rationality and advancement of the ICEEMDAN-CNN-Bys-BiLSTM prediction model.

[0137] Table 1 Average prediction error of each prediction model

[0138]

[0139] In summary, the photovoltaic power prediction method proposed in the present invention effectively improves the prediction accuracy compared with other methods. Among them, the cluster center correction strategy based on GASA-FCM hybrid clustering and isolated point variable weight enhances the global and local search capabilities and reduces the influence of noise and isolated points on the clustering results; based on the improved adaptive noise complete set empirical mode decomposition, the influence of randomness and volatility in data information on prediction accuracy can be effectively eliminated; the introduced Bayesian optimization algorithm can better mine the spatiotemporal characteristics of data, improve the prediction accuracy of the model as a whole, and ensure the safe, stable and economical operation of the power grid.

[0140] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with a single embodiment may be used in other described embodiments.

Claims

1. A photovoltaic power prediction method, characterized in that: include: Cluster the weather forecast data and photovoltaic power data in the historical time period to obtain the weather type to which the data belongs; Perform modal decomposition on weather forecast data and photovoltaic power data within the historical time period; The modal decomposed data is input into the photovoltaic power prediction model under the corresponding weather type according to the corresponding weather type to obtain the photovoltaic power prediction curve, and the photovoltaic power prediction curve is used to realize the prediction of future photovoltaic power.

2. A photovoltaic power prediction method according to claim 1, characterized in that: The clustering of weather forecast data and photovoltaic power data within the historical time period includes: Initialization: Randomly generate the initial annealing temperature, and the initial population gen0 includes the clustering features of weather forecast data and photovoltaic power data in the historical time period; S1: At the annealing temperature T, for the population gen at the g-th iteration g Perform selection, crossover and mutation operations to generate a new population gen g ′; S2: Use FCM clustering algorithm to calculate the new population gen g ′, each cluster center is a weather type; S3: Determine whether the number of iterations g reaches the maximum number of clustering iterations. If yes, execute S4; otherwise, execute S5. S4: Determine whether the maximum number of annealing iterations has been reached, if yes, output the weather type, otherwise execute S6; S5: Using simulated annealing algorithm to generate g ′ generates new individuals and adds them to the new population gen g ′, we obtain the population gen under the g+1th iteration g+1 , make g=g+1 and then return to S1; S6: Perform annealing operation on the simulated annealing temperature T using the simulated annealing algorithm. After updating the temperature T, set g=0 and then return to S1.

3. A photovoltaic power prediction method according to claim 2, characterized in that: In S2, the new population gen is calculated g After finding the cluster centers of ′, the cluster centers are modified, including: Calculate the population gen g The distance from each sample point to each cluster center in ′; Identify sample points whose distances meet the preset isolated point threshold as isolated points; A weight is assigned to each sample point, where the weight of the isolated point is smaller than that of the non-isolated point, so that the cluster center is corrected.

4. A photovoltaic power prediction method according to claim 2 or 3, characterized in that: The mean, standard deviation, maximum value, number of peaks and troughs, coefficient of variation, kurtosis and skewness of the total horizontal radiation variable of weather forecast data were selected as clustering features.

5. A photovoltaic power prediction method according to claim 4, characterized in that: The expressions of the coefficient of variation B, kurtosis F and skewness S are respectively: Where Y is the total horizontal radiation variable, σ is the standard deviation of Y, is the mean value of Y, m is the mth total horizontal radiation variable sample, and M is the total number of samples of Y.

6. A photovoltaic power prediction method according to claim 1, characterized in that: The modal decomposition of the weather forecast data and photovoltaic power data within the historical time period includes: The weather forecast data and photovoltaic power data in the historical period are used as the data to be decomposed D(t), and a white noise signal is added to the data to be decomposed D(t) to obtain the data containing noise D(T); Decomposing the data D(T) using the ICEEMDAN algorithm to obtain a set of modal components; The PCA algorithm is used to calculate the correlation coefficient of each modal component and arrange them in descending order; Delete the modal component with the smallest correlation coefficient, merge the remaining modal components, and output the data after modal decomposition.

7. A photovoltaic power prediction method according to claim 1, characterized in that: The photovoltaic power prediction model Including: Convolutional Neural Network CNN and BiLSTM network, The convolutional neural network CNN is used to extract spatial features from input data. The BiLSTM network is used to capture the temporal dependency of the spatial features, identify long-term dynamic changes, and obtain a photovoltaic power prediction curve.

8. A photovoltaic power prediction method according to claim 7, characterized in that: The convolutional neural network CNN and the BiLSTM network are both networks whose hyperparameters have been adjusted using the Bayesian algorithm.

9. A photovoltaic power prediction method according to claim 8, characterized in that: The calculation process of the BiLSTM network includes: h t =LSTM(s t ,h t-1 ), h t ′=LSTM(s t ,h t ′ -1 ), o t =W ho h t +W h'o h t ′+z o ′, Among them, h t and h t ′ are the hidden states of the forward and backward LSTM layers of the BiLSTM network, LSTM(·) represents the LSTM operation, and W ho and W h′o denote the weights of the forward and backward LSTM layers, respectively, and z o ′ is the bias of the BiLSTM output layer, o t is the output result of the BiLSTM output layer, s t The input data for the BiLSTM input layer.

10. A photovoltaic power prediction method according to claim 7, characterized in that: The convolutional neural network CNN can be expressed as: W1=RELU(X*K1+b1), H1=max(W1)+b2, W2=RELU(H1*K2+b3), H2=max(W2)+b4, y C =Sigmoid(H2×K3+b5), Where W1 and W2 are the outputs of the two convolutional layers in the convolutional neural network CNN, H1 and H2 are the outputs of the two pooling layers in the convolutional neural network CNN, * represents the convolution operation, X is the input of the convolutional neural network CNN, K1, K2, K3 are all weight matrices, b1, b2, b3, b4 and b5 are all bias coefficients, y C is the output of the convolutional neural network CNN.

Images