Space-time combination prediction model construction method for photovoltaic power generation power prediction, photovoltaic power generation power prediction method and related device

Through the space-time combined prediction model, combined with time and space correlation information, the problem of low accuracy caused by the singularity of the existing photovoltaic power prediction model is solved, and more efficient photovoltaic power prediction is achieved.

CN119090057APending Publication Date: 2024-12-06INFORMATION & COMM CO OF STATE GRID XINJIANG ELECTRIC POWER CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411099494.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing photovoltaic power prediction model is single, and cannot combine the correlation between time and space, resulting in low prediction accuracy.

Method used

Using the spatial and temporal combined prediction model, the final photovoltaic power prediction value is generated by the improved Pyraformer model and the hybrid jump graph convolutional network model with RNN position coding layer, combining time and spatial correlation information, and a splicing and fusion module is used to generate the final photovoltaic power generation power prediction value.

Benefits of technology

The accuracy of photovoltaic power generation prediction is improved, and more accurate prediction results are generated by combining time and space-related information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090057A_ABST
    Figure CN119090057A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of photovoltaic power generation, in particular to a space-time combination prediction model construction method for photovoltaic power generation power prediction, a photovoltaic power generation power prediction method and a related device.The space-time combination prediction model construction method comprises the steps that a training sample library is obtained and divided into a training set, a verification set and a test set; training a space-time combination prediction model by using the training set in combination with a training termination condition, wherein the space-time combination prediction model comprises an improved Pyraformer model, a mixed jump graph convolutional network model with an RNN position coding layer and a splicing fusion module; according to the invention, the space-time combination prediction model is provided, and the model combines the advantages of a mixed jump graph convolutional network with an RNN position coding layer and an improved Pyraformer, so that the prediction accuracy of the space-time combination prediction model is improved, and the prediction accuracy of the space-time combination prediction model is improved. And a final photovoltaic power generation power prediction value is obtained by combining prediction information in two aspects of space correlation and time correlation, so that the photovoltaic power generation power prediction accuracy of the space-time combination prediction model is effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of photovoltaic power generation, and is a method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, a photovoltaic power generation prediction method and related devices. Background Art

[0002] In recent years, solar energy has played an increasingly important role in the global power system. Photovoltaic power generation, as one of the main forms of solar energy, not only provides clean energy for the world and reduces dependence on fossil fuels, but also has the characteristics of low operation and maintenance costs, thus bringing significant economic benefits. Although photovoltaic technology has made significant progress in converting solar energy into electrical energy, it still faces some challenges, such as the fact that power fluctuations are greatly affected by the external environment. When photovoltaic systems are connected to the power grid, they may have a certain impact on the energy network, and the long-term power uncertainty makes the planning and maintenance of the economic system difficult. Therefore, by making short-term and medium-term forecasts of photovoltaic power generation, the negative impact of power grid fluctuations can be effectively alleviated. However, the existing photovoltaic power generation power prediction model is single and cannot combine the correlation of time and space to predict photovoltaic power generation, which easily leads to the problem of low prediction accuracy. Summary of the invention

[0003] The present invention provides a method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, a photovoltaic power generation prediction method and related devices, which overcome the shortcomings of the above-mentioned prior art and can effectively solve the problem that the photovoltaic power generation prediction model in the prior art is single and cannot combine the correlation of time and space to predict photovoltaic power generation, which easily leads to the problem of low prediction accuracy.

[0004] One of the technical solutions of the present invention is achieved by the following measures: a method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, comprising:

[0005] A training sample library is obtained and divided into a training set, a validation set and a test set, wherein the training sample library includes a plurality of sample data groups, each of which includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and labeling information of the current photovoltaic power generation value of the single power station;

[0006] The spatiotemporal combined prediction model is trained using the training set in combination with the training termination condition, wherein the spatiotemporal combined prediction model includes an improved Pyraformer model, a hybrid skip graph convolutional network model with an RNN position encoding layer, and a splicing and fusion module. The improved Pyraformer model outputs a photovoltaic power prediction value based on time correlation, and the hybrid skip graph convolutional network model with an RNN position encoding layer outputs a photovoltaic power prediction value based on spatial correlation. The splicing and fusion module splices and combines the two prediction values ​​and assigns weights to obtain the final photovoltaic power prediction value of a single power station.

[0007] The validation set and test set are used to verify and test the trained spatiotemporal combined prediction model respectively, the model parameters of the spatiotemporal combined prediction model are determined, and the spatiotemporal combined prediction model that meets the test and evaluation requirements is output.

[0008] The following are further optimizations and / or improvements to the above technical solutions:

[0009] The above-mentioned improved Pyraformer model includes an encoder and a decoder. The encoder includes a global branch and a local branch. The global branch includes a global convolution kernel Gconv to extract global information. The local branch includes two modules, CVM and Pyraformer. CVM captures cross-dimensional dependencies, and Pyraformer captures cross-temporal dependencies. The decoder includes an attention mechanism and a fully connected layer. The attention mechanism integrates global information and local information, and the model output result is obtained by inference through a fully connected layer.

[0010] The hybrid skip graph convolutional network model with the RNN position encoding layer includes an RNN position encoding layer and a hybrid skip graph convolutional network. The position embedding and final representation of the historical photovoltaic power station power sequence are obtained through the RNN position encoding layer. The hybrid skip graph convolutional network includes two hybrid skip propagation layers, which process the inflow and outflow information of the nodes respectively. The final output information is the sum of the outputs of the two hybrid skip propagation layers.

[0011] The above-mentioned splicing and fusion module includes a splicing layer and a fully connected layer. The splicing layer splices and combines the photovoltaic power generation prediction value based on time correlation and the photovoltaic power generation prediction value based on spatial correlation. The fully connected layer assigns weights to the spliced ​​and combined photovoltaic power generation prediction values ​​and outputs the final photovoltaic power generation prediction value.

[0012] The second technical solution of the present invention is achieved by the following measures: a photovoltaic power generation prediction method based on a spatiotemporal combined prediction model, comprising:

[0013] Obtaining historical photovoltaic power generation data and meteorological data of the target power station, and historical photovoltaic power generation data of multiple power stations adjacent to the target power station;

[0014] The historical photovoltaic power generation data and meteorological data of the target power station, as well as the historical photovoltaic power generation data of multiple power stations adjacent to the target power station, are input into a spatiotemporal combined prediction model to obtain a predicted value of the photovoltaic power generation of the target power station, wherein the spatiotemporal combined prediction model is obtained by the spatiotemporal combined prediction model construction method for photovoltaic power generation prediction.

[0015] The following are further optimizations and / or improvements to the above technical solutions:

[0016] The above also includes verifying the rationality of selecting the historical photovoltaic power generation data of multiple power stations after obtaining the historical photovoltaic power generation data and meteorological data of the target power station and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station, including:

[0017] Set rationality verification conditions to determine whether the historical photovoltaic power generation data of multiple power stations meet all rationality verification conditions, where the rationality verification conditions include:

[0018] The Moran index is used to perform spatial autocorrelation analysis on the historical photovoltaic power data of multiple power stations to determine whether there is spatial autocorrelation in the historical photovoltaic power data of multiple power stations.

[0019] Granger causality is used to verify whether the historical photovoltaic power generation data of multiple power stations can improve the prediction accuracy of the target power station;

[0020] In response to no, the historical photovoltaic power generation data of the plurality of power stations are reselected.

[0021] The third technical solution of the present invention is achieved by the following measures: a device for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, comprising:

[0022] A sample acquisition unit acquires a training sample library and divides it into a training set, a validation set and a test set, wherein the training sample library includes a plurality of sample data groups, each of which includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and labeling information of the current photovoltaic power generation value of the single power station;

[0023] A training unit, which trains the spatiotemporal combined prediction model using the training set in combination with the training termination condition, wherein the spatiotemporal combined prediction model includes an improved Pyraformer model, a hybrid skip graph convolutional network model with an RNN position encoding layer, and a splicing and fusion module. The improved Pyraformer model outputs a photovoltaic power prediction value based on time correlation, and the hybrid skip graph convolutional network model with an RNN position encoding layer outputs a photovoltaic power prediction value based on spatial correlation. The splicing and fusion module splices and combines the two prediction values ​​and assigns weights to obtain the final photovoltaic power prediction value of a single power station.

[0024] The verification and testing unit uses the verification set and the test set to verify and test the trained spatiotemporal combined prediction model respectively, determines the model parameters of the spatiotemporal combined prediction model, and outputs the spatiotemporal combined prediction model that meets the test and evaluation requirements.

[0025] The fourth technical solution of the present invention is achieved by the following measures: a photovoltaic power generation prediction device based on a spatiotemporal combined prediction model, comprising:

[0026] A data acquisition unit, which acquires historical photovoltaic power generation data and meteorological data of a target power station, and historical photovoltaic power generation data of multiple power stations adjacent to the target power station;

[0027] The prediction unit inputs the historical photovoltaic power generation data and meteorological data of the target power station, and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station into the spatiotemporal combined prediction model to obtain the photovoltaic power generation power prediction value of the target power station, wherein the spatiotemporal combined prediction model is obtained by constructing the spatiotemporal combined prediction model for photovoltaic power generation prediction.

[0028] The following are further optimizations and / or improvements to the above technical solutions:

[0029] The above also includes a data verification unit, which verifies the rationality of selecting the historical photovoltaic power generation data of the multiple power stations after obtaining the historical photovoltaic power generation data and meteorological data of the target power station and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station, including:

[0030] Set rationality verification conditions to determine whether the historical photovoltaic power generation data of multiple power stations meet all rationality verification conditions, where the rationality verification conditions include:

[0031] The Moran index is used to perform spatial autocorrelation analysis on the historical photovoltaic power data of multiple power stations to determine whether there is spatial autocorrelation in the historical photovoltaic power data of multiple power stations.

[0032] Granger causality is used to verify whether the historical photovoltaic power generation data of multiple power stations can improve the prediction accuracy of the target power station;

[0033] In response to no, the historical photovoltaic power generation data of the plurality of power stations are reselected.

[0034] The present invention proposes a spatiotemporal combined prediction model, which combines the advantages of a hybrid skip graph convolutional network with an RNN position encoding layer and an improved Pyraformer, and splices and combines the spatial correlation prediction information realized by the hybrid skip graph convolutional network and the temporal correlation prediction information realized by the improved Pyraformer model through a splicing layer and a fully connected layer, and distributes weights to obtain the final prediction result of the target power station. The final photovoltaic power generation prediction value is obtained by combining the prediction information of spatial correlation and temporal correlation, thereby effectively improving the photovoltaic power generation prediction accuracy of the spatiotemporal combined prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Attached Figure 1 The figure is a flow chart of a model building method of the present invention.

[0036] Attached Figure 2 It is a structural schematic diagram of the spatiotemporal combined prediction model in the present invention.

[0037] Attached Figure 3 It is a schematic diagram of the structure of the improved Pyraformer model in the present invention.

[0038] Attached Figure 4 Schematic diagram of the structure of the RNN position encoding layer in the present invention.

[0039] Attached Figure 5 Schematic diagram of the structure of the hybrid skip graph convolutional network in the present invention.

[0040] Attached Figure 6 The figure is a flow chart of another model building method of the present invention.

[0041] Attached Figure 7 The figure is a schematic flow chart of a photovoltaic power generation prediction method according to the present invention.

[0042] Attached Figure 8 The figure is a schematic flow chart of another photovoltaic power generation prediction method of the present invention.

[0043] Attached Fig. 9 This is a schematic diagram of a Moran's index result in the present invention.

[0044] Attached Fig.10 It is a schematic diagram of the structure of the model building device of the present invention.

[0045] Attached Fig.11It is a schematic diagram of the structure of the photovoltaic power prediction device of the present invention. DETAILED DESCRIPTION

[0046] The present invention is not limited by the following embodiments, and specific implementation methods can be determined based on the technical solution of the present invention and actual conditions.

[0047] The present invention will be further described below in conjunction with embodiments and drawings:

[0048] Embodiment 1: As attached Figure 1 As shown, the embodiment of the present invention discloses a method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, comprising:

[0049] Step S110, obtaining a training sample library and dividing it into a training set, a validation set and a test set, wherein the training sample library includes a plurality of sample data groups, each of which includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and labeling information of the current photovoltaic power generation value of the single power station;

[0050] Step S120, training the spatiotemporal combined prediction model using the training set in combination with the training termination condition, wherein the spatiotemporal combined prediction model includes an improved Pyraformer model, a hybrid skip graph convolutional network model with an RNN position encoding layer, and a splicing and fusion module, the improved Pyraformer model outputs a photovoltaic power prediction value based on time correlation, the hybrid skip graph convolutional network model with an RNN position encoding layer outputs a photovoltaic power prediction value based on spatial correlation, and the splicing and fusion module splices and combines the two prediction values ​​and assigns weights to obtain a final photovoltaic power prediction value of a single power station;

[0051] Step S130, using the validation set and the test set to respectively validate and test the trained spatiotemporal combined prediction model, determine the model parameters of the spatiotemporal combined prediction model, and output the spatiotemporal combined prediction model that meets the test evaluation requirements.

[0052] The embodiment of the present invention discloses a method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, and proposes a spatiotemporal combined prediction model. The model combines the advantages of a hybrid jump graph convolutional network with an RNN position encoding layer and an improved Pyraformer, and splices and combines the spatial correlation prediction information realized by the hybrid jump graph convolutional network and the temporal correlation prediction information realized by the improved Pyraformer model through a splicing layer and a fully connected layer, and distributes weights to obtain the final prediction result of the target power station. The final photovoltaic power generation prediction value is obtained by combining the prediction information of the spatial correlation and the temporal correlation, thereby effectively improving the photovoltaic power generation prediction accuracy of the spatiotemporal combined prediction model.

[0053] Embodiment 2: The embodiment of the present invention discloses a method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, comprising:

[0054] Step S210, obtain a training sample library and divide it into a training set, a validation set and a test set, wherein the training sample library includes several sample data groups, each sample data group includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and labeling information of the current photovoltaic power generation value of the single power station.

[0055] The steps above to obtain the training sample library include:

[0056] (1) obtaining a number of sample data sets, each of which includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and label information of the current photovoltaic power generation value of the single power station; wherein the meteorological data includes irradiance, wind speed, temperature, etc.;

[0057] (2) All sample data sets are preprocessed. The preprocessing process includes using linear interpolation to fill missing values, setting any negative power generation to 0, using the 3σ principle to identify outliers, marking data that exceed the 3σ boundary of the normal distribution and replacing them with the average of the adjacent data;

[0058] (3) All sample data sets are divided into training set, validation set and test set according to 70%, 20% and 10% respectively.

[0059] Step S220, the spatiotemporal combined prediction model is trained using the training set in combination with the training termination condition, wherein the spatiotemporal combined prediction model includes an improved Pyraformer model, a hybrid skip graph convolutional network model with an RNN position encoding layer, and a splicing and fusion module, the improved Pyraformer model outputs a photovoltaic power prediction value based on time correlation, the hybrid skip graph convolutional network model with an RNN position encoding layer outputs a photovoltaic power prediction value based on spatial correlation, and the splicing and fusion module splices and combines the two prediction values ​​and assigns weights to obtain the final photovoltaic power prediction value of a single power station.

[0060] In the above steps, the training termination condition may be, but is not limited to, the number of training iterations, such as 200.

[0061] As attached Figure 2As shown in the above steps, during training, the historical photovoltaic power generation data and meteorological data of a single power station are input into the improved Pyraformer model to obtain a photovoltaic power generation power prediction value based on time correlation. The improved Pyraformer model includes an encoder and a decoder. The encoder includes a global branch and a local branch. The global branch includes a global convolution kernel Gconv to extract global information. The local branch includes two modules, CVM and Pyraformer, where CVM captures cross-dimensional dependencies and Pyraformer captures cross-time dependencies. The decoder includes an attention mechanism and a fully connected layer. The attention mechanism integrates global information and local information, and the model output result is obtained by reasoning through a fully connected layer.

[0062] Specifically, as attached Figure 3 As shown in Figure 2, the global branch of the encoder is used to capture the global information of the photovoltaic sequence, that is, the long-term dependencies in the photovoltaic power generation sequence. The global branch includes a global convolution kernel Gconv, which converts the entire photovoltaic input sequence X∈R I×d (I is the input length, d is the number of dimensions, i.e. the number of covariates such as photovoltaic power generation and irradiance, wind speed, temperature, etc.) is input into the global convolution kernel, whose complexity is sublinear (logarithmic) with the sequence length. The global convolution kernel uses a long kernel that spans the entire length of the input sequence, so that it can capture long-term dependencies. Use K∈R I×d represents a learnable global kernel, Y∈R I×d Represents the output. The global convolution operation is expressed as:

[0063] Y=X*K (1)

[0064] Where * is the convolution operator. Although the complexity of the global convolution operation is O(N 2 ), but it can be realized by fast Fourier transform represented by F, reducing the complexity to O(NlogN), expressed as:

[0065] X*K=F -1 (F(X)·F(K)) (2)

[0066] Compared to local convolutions with fixed kernel size, global convolutions require the same kernel size as the input sequence length I. For long sequences, it is difficult to simply parameterize the convolution kernel like local convolution. Therefore, efficient convolution kernel parameterization is necessary. SGConv composes the global convolution kernel through a series of sub-kernels of increasing size. The size of each larger sub-kernel is twice that of the previous sub-kernel, but each sub-kernel is upsampled from the same number of parameters. This parameterization ensures that the number of parameters is logarithmically proportional to the input length. In addition, a weighted combination of sub-kernels with weight decay is adopted, and smaller weights are assigned to larger sub-kernels, which can provide a good inductive bias for long sequence modeling, thereby helping to improve performance. The global convolution kernel is defined as:

[0067]

[0068] In the formula, w i ∈R d is the i-th subkernel k i The parameter N = [log2(I / d)] + 1 represents the number of scales. Upsampling operations (implemented in the form of linear interpolation) are used to form sub-kernels of different scales; Indicates that w i Upsample to length 2 max[i -1,0]d. Z is the normalization constant, which ensures that the convolution operation does not change the scale of the input; α is the attenuation coefficient, which controls the attenuation speed.

[0069] Based on the above, given a kernel K parameterized by multi-scale sub-kernels mask , the process of global convolution operation is defined as:

[0070] Y=F -1 (F(X)·F(K mask )) (4)

[0071] The local branch of the encoder is used to capture the local information of the photovoltaic sequence, that is, the dependencies between nearby time nodes. The local branch includes two modules, CVM and Pyraformer. CVM is used to capture cross-dimensional dependencies, and Pyraformer is used to capture cross-time dependencies. The specific contents are as follows:

[0072] The CVM module proposes cross-variable attention to learn the dependencies between variables, which consists of a multi-head attention and a feedforward network. The input photovoltaic sequence X is a two-dimensional tensor with a shape of I×d, where d represents the number of variables. Before input, the sequence X first needs to be flipped. Cross-variable attention is the key part of multi-head attention and is defined as:

[0073]

[0074] Where Q is the query matrix, K is the key matrix, and V is the value matrix. Q, K, and V are generally obtained by linearly transforming the original input X, and d is the number of variables.

[0075] The Pyraformer model is a neural network based on a multi-resolution pyramid attention mechanism. The multi-scale photovoltaic sequence is generated by the coarse-grained building block (CSCM) and input into the pyramid attention module (PAM) for intra-scale and inter-scale information transfer between nodes. Finally, the output of PAM is passed through a linear layer to obtain the output result. Unlike the original attention mechanism, in the pyramid attention mechanism, each node only needs to pay attention to a set of neighboring nodes. Assume is the lth node on scale s, where s = 1, 2, ... S represents the order of scales from bottom to top, A set of neighbor nodes Defined as:

[0076]

[0077] In the formula, Includes three groups of nodes and in Indicates that the same scale includes The set of adjacent nodes including itself, represents the C child nodes in the C-ary tree, Represents the parent node. Based on the above formula, the node The attention at can be simplified to:

[0078]

[0079] As attached Figure 3 As shown in the figure, the decoder includes an attention mechanism and a fully connected layer. In order to improve the utilization of global and local information, the attention mechanism is used to integrate global and local information, and the model output result is obtained by inference through a fully connected layer. The decoder first linearly maps the global and local information, and then uses the global information as the query vector (q), and the local information as the key vector (k) and the value vector (v), which are input into the Attention layer. In this way, the correlation between each global-local token pair can be captured, and the local information is prompted to adaptively explore their relationship with the global information, making it more representative and informative. It can be specifically expressed as:

[0080] q=MLP(Z global ) (8)

[0081] k=MLP(Z local ),v=MLP(Z local) (9)

[0082]

[0083] Therefore, a global convolution module (Gconv) is introduced and incorporated into the global branch of the model encoder, making full use of the global convolution kernel's modeling ability for long-term dependencies, thereby effectively capturing the global information of the photovoltaic power generation sequence. A cross-variable module (CVM) is proposed, and CVM and Pyraformer are combined to form a local branch of the model encoder. In this way, the dependencies between different variables can be learned through the cross-variable attention mechanism, and the cross-time dependencies of the photovoltaic sequence can be learned through Pyraformer, thereby efficiently capturing the local information of the photovoltaic power generation sequence. Therefore, the attention mechanism is used to maximize the integration of global information and local information, which significantly improves the prediction accuracy.

[0084] As attached Figure 2 As shown in the above steps, during training, the historical photovoltaic power generation data of multiple power stations adjacent to a single power station are input into the hybrid jump graph convolutional network model with the RNN position encoding layer to obtain the photovoltaic power generation power prediction value based on spatial correlation. The hybrid jump graph convolutional network model with the RNN position encoding layer includes the RNN position encoding layer and the hybrid jump graph convolutional network. The position embedding and final representation of the historical photovoltaic power station power sequence are obtained through the RNN position encoding layer. The hybrid jump graph convolutional network includes two hybrid jump propagation layers, which process the inflow and outflow information of the node respectively. The final output information is the sum of the outputs of the two hybrid jump propagation layers.

[0085] The structure of the RNN position encoding layer is shown in the following figure. Figure 4 As shown in the figure, specifically, the RNN position encoding layer aims to utilize the sequence modeling characteristics of RNN to generate position encodings with strict temporal order information. Instead of passing the input data directly through the RNN layer, the temporal order information is explicitly introduced in the position encoding only through the RNN, and this information is injected into the feature-driven sequence labeling.

[0086] The historical power sequence x of each PV power station i First, it is divided into multiple blocks, each of which has a length of P, and S represents the step length between different blocks, which will form a sequence Where N is the number of blocks, calculated as Next, let A linear layer is used to generate the representation of each block. Each representation is fed into the corresponding RNN unit. The hidden states of the last layer are concatenated to get the position embedding. The specific process is as follows:

[0087]

[0088] In the formula, RNN(·) can be any RNN structure model, such as RNN, LSTM and GRU; C represents the number of RNN layers, The vector representation of the jth block of the power sequence of the i-th PV plant; represents the hidden state obtained from the corresponding j-th RNN unit in the c-th layer; it is worth noting that at the 0th layer, i.e., the initial input, POS (i) and E (i) They represent the position embedding and final representation of the i-th PV power station power sequence respectively.

[0089] The structure of the hybrid skip graph convolutional network is shown in the attached figure. Figure 5 As shown in the figure, specifically, the hybrid jump graph convolutional network is a graph convolutional neural network based on a hybrid jump propagation layer proposed by MTGNN. The goal of this network is to fuse the information of a node with that of its neighbors to process the spatial dependency in the graph. Specifically, it includes two hybrid jump propagation layers, which process the inflow and outflow information of the node respectively. The final output information is the sum of the outputs of the two hybrid jump propagation layers.

[0090] The hybrid jump propagation layer consists of two steps, namely information propagation and information selection. The horizontal information flow direction corresponds to information propagation, and the vertical information flow direction corresponds to information selection. In the information propagation step, multi-hop neighbor information is propagated through the graph structure. Its definition is as follows:

[0091]

[0092] In the formula, H (0) is the original state of each node after RNN position encoding; in order to alleviate over-smoothing and prevent the hidden state of the node from converging to a point, we use (A+I) to establish the autocorrelation of the node and use the hyperparameter α to retain a certain proportion of the original state of the node during the propagation process; is the degree matrix of (A+I); is the normalized adjacency matrix.

[0093] When only the information propagation step is used, some node information will be lost. In the extreme case where there is no spatial dependency, aggregating neighborhood information will only introduce useless noise to each node, so an information selection step needs to be introduced to act as a feature selector. This step uses a weighted matrix to summarize the results of different propagation depths, which is defined as follows:

[0094]

[0095] In the formula, K represents the depth of information propagation, and weight is a learnable parameter, which contains K randomly initialized variables. The Softmax function is used to assign different weights to multi-hop neighbor information, and then the results are added to get the final output

[0096] As attached Figure 2 As shown, in the above steps, the splicing and fusion module includes a splicing layer and a fully connected layer. The splicing layer splices and combines the photovoltaic power generation prediction value based on time correlation and the photovoltaic power generation prediction value based on space correlation. The fully connected layer assigns weights to the spliced ​​and combined photovoltaic power generation prediction values ​​and outputs the final photovoltaic power generation prediction value.

[0097] The number of neurons in the above-mentioned fully connected layer can be set, thereby completing photovoltaic power generation prediction at multiple scales.

[0098] Step S230, using the validation set and the test set to respectively validate and test the trained spatiotemporal combined prediction model, determine the model parameters of the spatiotemporal combined prediction model, and output the spatiotemporal combined prediction model that meets the test evaluation requirements.

[0099] In the above steps, the validation set is used to verify the spatiotemporal combined prediction model, the model parameters of the spatiotemporal combined prediction model are adjusted, the spatiotemporal combined prediction model is tested using the test set, and the test results are evaluated and analyzed. The evaluation and analysis methods include mean square error index (MSE), mean absolute error index (MAE), root mean square error index (RMSE), determination coefficient index (R 2 ), as shown below:

[0100]

[0101] Where N is the number of samples used to calculate the prediction error, y i is the actual photovoltaic power value, is the predicted value, is the average of N actual photovoltaic power values. MSE is a frequently used indicator to evaluate the effect of time series prediction. MAE can better reflect the true situation of the error between the predicted value and the actual value because it is not so sensitive to outliers. RMSE is used to measure the deviation between the observed value and the predicted value. The lower the values ​​of the above three indicators, the better the model prediction performance. 2 Indicates the proportion of variance explained by the model and shows the correlation between the predicted and actual values. 2 The larger the value, the better the model prediction performance.

[0102] Therefore, meeting the test and evaluation requirements is the threshold condition for the above indicators, and the specific threshold is set based on actual conditions.

[0103] Embodiment 3: As attached Figure 6 As shown, the embodiment of the present invention discloses a method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, comprising:

[0104] Step S310, obtaining a training sample library and dividing it into a training set, a validation set and a test set, wherein the training sample library includes a plurality of sample data groups, each of which includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and labeling information of the current photovoltaic power generation value of the single power station;

[0105] Step S320, verifying the rationality of selecting the historical photovoltaic power generation data of multiple power stations in each sample data group, includes:

[0106] Set rationality verification conditions to determine whether the historical photovoltaic power generation data of multiple power stations meet all rationality verification conditions, where the rationality verification conditions include:

[0107] The Moran index is used to perform spatial autocorrelation analysis on the historical photovoltaic power data of multiple power stations to determine whether there is spatial autocorrelation in the historical photovoltaic power data of multiple power stations.

[0108] Granger causality is used to verify whether the historical photovoltaic power generation data of multiple power stations can improve the prediction accuracy of the target power station;

[0109] In response to no, reselecting the historical photovoltaic power generation data of the plurality of power stations;

[0110] Step S330, training the spatiotemporal combined prediction model using the training set in combination with the training termination condition, wherein the spatiotemporal combined prediction model includes an improved Pyraformer model, a hybrid skip graph convolutional network model with an RNN position encoding layer, and a splicing and fusion module, the improved Pyraformer model outputs a photovoltaic power prediction value based on time correlation, the hybrid skip graph convolutional network model with an RNN position encoding layer outputs a photovoltaic power prediction value based on spatial correlation, and the splicing and fusion module splices and combines the two prediction values ​​and assigns weights to obtain a final photovoltaic power prediction value of a single power station;

[0111] Step S340, using the validation set and the test set to respectively validate and test the trained spatiotemporal combined prediction model, determine the model parameters of the spatiotemporal combined prediction model, and output the spatiotemporal combined prediction model that meets the test evaluation requirements.

[0112] Embodiment 4: As attached Figure 7As shown, the embodiment of the present invention discloses a photovoltaic power generation prediction method based on a spatiotemporal combined prediction model, comprising:

[0113] Step S410, acquiring historical photovoltaic power generation data and meteorological data of the target power station, and historical photovoltaic power generation data of multiple power stations adjacent to the target power station;

[0114] Step S420, inputting the historical photovoltaic power generation data and meteorological data of the target power station, and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station into the spatiotemporal combined prediction model to obtain the photovoltaic power generation power prediction value of the target power station, wherein the spatiotemporal combined prediction model is obtained by the spatiotemporal combined prediction model construction method for photovoltaic power generation prediction.

[0115] Embodiment 5: As attached Figure 8 As shown, the embodiment of the present invention discloses a photovoltaic power generation prediction method based on a spatiotemporal combined prediction model, comprising:

[0116] Step S510, acquiring historical photovoltaic power generation data and meteorological data of a target power station, and historical photovoltaic power generation data of a plurality of power stations adjacent to the target power station;

[0117] Step S520, after obtaining the historical photovoltaic power generation data and meteorological data of the target power station, and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station, verify the rationality of selecting the historical photovoltaic power generation data of the multiple power stations, including:

[0118] (1) Setting rationality verification conditions to determine whether the historical photovoltaic power generation data of multiple power stations meet all rationality verification conditions, where the rationality verification conditions include:

[0119] The Moran index is used to perform spatial autocorrelation analysis on the historical photovoltaic power data of multiple power stations to determine whether there is spatial autocorrelation in the historical photovoltaic power data of multiple power stations.

[0120] Granger causality is used to verify whether the historical photovoltaic power generation data of multiple power stations can improve the prediction accuracy of the target power station;

[0121] (2) In response to "no", reselect the historical photovoltaic power generation data of the plurality of power stations. If "yes", execute step S530;

[0122] Step S530, inputting the historical photovoltaic power generation data and meteorological data of the target power station, and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station into the spatiotemporal combined prediction model to obtain the photovoltaic power generation power prediction value of the target power station, wherein the spatiotemporal combined prediction model is obtained by the spatiotemporal combined prediction model construction method for photovoltaic power generation prediction.

[0123] In the above step S520, the Moran index is a widely used index for characterizing spatial autocorrelation, and its definition is as follows:

[0124]

[0125] Where n is the total number of PV plants in the study area. i and j represent the i-th PV plant and the j-th PV plant. Zi and Zj represent the PV power generation values ​​of the i-th PV plant and the j-th PV plant at a given time point. Z is the average PV power generation value of all PV plants studied at a given time point. W ij is the spatial weight between the ith PV site and the jth PV site. If the ith PV site is adjacent to the jth PV site, then W ij is equal to 1. Otherwise, W ij is equal to 0.

[0126] The Moran index is used to perform spatial autocorrelation analysis on the historical photovoltaic power generation data of multiple power stations. Fig. 9 The Moran index result graph shown in the figure shows that the photovoltaic power generation value is zero at night, which may cause the denominator to be zero, making it impossible to calculate the Moran index. Therefore, the Moran index from 06:00 to 19:00 was calculated. Figure 2 It can be seen that the photovoltaic power generation of different photovoltaic stations has spatial correlation at every moment of the day, and the Moran index is about 0.3, showing a strong spatial autocorrelation.

[0127] Furthermore, the Moran index can be normalized. The normalized Moran index is defined as follows:

[0128]

[0129] In the above step S520, the Granger causality test is a widely used test method for testing whether adding the historical values ​​of another related time series helps to improve the prediction accuracy of the studied time series. The detailed process of the Granger causality test is as follows: If adding the historical values ​​of time series X helps to improve the prediction accuracy of time series Y, then there is a causal relationship. In this case, time series X is called the Granger cause of time series Y, which is recorded as X→Y.

[0130] In order to test the causal relationship between time series X and time series Y, two linear autoregression models are used as follows:

[0131]

[0132] If the variance of the error E1 of the model considering variable Y (coefficient A12 is not zero) is statistically significantly smaller than the model without considering variable Y (coefficient A12 is equal to zero), it means that adding the historical values ​​of time series Y can significantly reduce the forecast error of time series X. In this case, variable Y is the Granger cause of variable X. Similarly, if the variance of the error E2 of the model including variable X (coefficient A21 is not zero) is statistically significantly smaller than the model without variable X (coefficient A21 is equal to zero), it means that adding the historical values ​​of time series X can significantly reduce the forecast error of time series Y. In this case, variable X is the Granger cause of variable Y.

[0133] In order to test whether variable Y is the variable X that causes Granger, the F test is used. First, the residual sum of squares RSS of the model that only uses the historical values ​​of X to predict future X and the model that uses the historical values ​​of time series X and time series Y to predict future X are calculated respectively. The specific calculation process is as follows:

[0134]

[0135] Based on RSS1 and RSS2, the test statistic for the F test (S1) can be calculated as follows.

[0136]

[0137] Where l is the number of lags considered and T is the number of forecast values.

[0138] The S1 statistic follows an F distribution with l and (T-2l-1) degrees of freedom. To test whether Y causes X (Y→X), an F test is performed under the null hypothesis that Y does not cause X and the alternative hypothesis that Y causes X. In practice, the p-value is calculated based on an F distribution with l and (T-2l-1) degrees of freedom. If the p-value is less than a given significance level α, such as 0.01, the null hypothesis is rejected, and there is reason to believe that Y causes X (Y→X), that is, introducing historical values ​​of Y can significantly reduce the prediction error of X.

[0139] For example, for each of the 7 power plants, a Granger causality test is performed between it and each of the other 6 power plants. The test p-values ​​are listed in Tables (a) to (g), where the numbers represent the power plant ID in the first row of each table.

[0140] Table a). Granger causality test p-value for site 1

[0141] 2 3-A 3-B 3-C 4 5 1 0.0 0.0 0.0 0.0 0.0 0.0

[0142] Table b). Granger causality test p-value for site 2

[0143] 1 3-A 3-B 3-C 4 5 2 0.0 0.0 0.0 0.0 0.0 0.0

[0144] Table c).3-P value of Granger causality test for site A

[0145] 1 2 3-B 3-C 4 5 3-A 0.0 0.0 0.0 0.0 0.0 0.0

[0146] Table d). Granger causality test p-value for site 3-B

[0147] 1 2 3-A 3-C 4 5 3-B 0.0 0.0 0.0 0.0 0.0 0.0

[0148] Table e). Granger causality test p-value for site 3-C

[0149] 1 2 3-A 3-B 4 5 3-C 0.0 0.0 0.0 0.0 0.0 0.0

[0150] Table f). Granger causality test p-value for site 4

[0151] 1 2 3-A 3-B 3-C 5 4 0.0 0.0 0.0 0.0 0.0 0.0

[0152] Table g). Granger causality test p value for site 5

[0153] 1 2 3-A 3-B 3-C 4 5 0.0 0.0 0.0 0.0 0.0 0.0

[0154] As can be seen from the table, all test p values ​​are less than 0.01, which means that there is a Granger causality relationship between the several sites studied, which means that introducing photovoltaic data from other nearby photovoltaic power stations can help improve the prediction results of the target power station.

[0155] Embodiment 6: As attached Fig.10 As shown, the embodiment of the present invention discloses a device for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, comprising:

[0156] A sample acquisition unit acquires a training sample library and divides it into a training set, a validation set and a test set, wherein the training sample library includes a plurality of sample data groups, each of which includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and labeling information of the current photovoltaic power generation value of the single power station;

[0157] A training unit, which trains the spatiotemporal combined prediction model using the training set in combination with the training termination condition, wherein the spatiotemporal combined prediction model includes an improved Pyraformer model, a hybrid skip graph convolutional network model with an RNN position encoding layer, and a splicing and fusion module. The improved Pyraformer model outputs a photovoltaic power prediction value based on time correlation, and the hybrid skip graph convolutional network model with an RNN position encoding layer outputs a photovoltaic power prediction value based on spatial correlation. The splicing and fusion module splices and combines the two prediction values ​​and assigns weights to obtain the final photovoltaic power prediction value of a single power station.

[0158] The verification and testing unit uses the verification set and the test set to verify and test the trained spatiotemporal combined prediction model respectively, determines the model parameters of the spatiotemporal combined prediction model, and outputs the spatiotemporal combined prediction model that meets the test and evaluation requirements.

[0159] Embodiment 7: As attached Fig.11 As shown, the embodiment of the present invention discloses a photovoltaic power generation prediction device based on a spatiotemporal combined prediction model, comprising:

[0160] A data acquisition unit, which acquires historical photovoltaic power generation data and meteorological data of a target power station, and historical photovoltaic power generation data of multiple power stations adjacent to the target power station;

[0161] The data verification unit, after obtaining the historical photovoltaic power generation data and meteorological data of the target power station and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station, verifies the rationality of the selection of the historical photovoltaic power generation data of the multiple power stations, including:

[0162] Set rationality verification conditions to determine whether the historical photovoltaic power generation data of multiple power stations meet all rationality verification conditions, where the rationality verification conditions include:

[0163] The Moran index is used to perform spatial autocorrelation analysis on the historical photovoltaic power data of multiple power stations to determine whether there is spatial autocorrelation in the historical photovoltaic power data of multiple power stations.

[0164] Granger causality is used to verify whether the historical photovoltaic power generation data of multiple power stations can improve the prediction accuracy of the target power station;

[0165] In response to no, reselecting the historical photovoltaic power generation data of the plurality of power stations;

[0166] The prediction unit inputs the historical photovoltaic power generation data and meteorological data of the target power station, and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station into the spatiotemporal combined prediction model to obtain the photovoltaic power generation power prediction value of the target power station, wherein the spatiotemporal combined prediction model is obtained by constructing the spatiotemporal combined prediction model for photovoltaic power generation prediction.

[0167] Embodiment 8: The embodiment of the present invention discloses an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the computer program is loaded and executed by the processor to implement a photovoltaic power generation power prediction method based on a spatiotemporal combined prediction model.

[0168] The processor may be a central processing unit (CPU), a general purpose processor, a digital signal processor (DSP), an ASIC, an FPGA or other programmable logic device, a transistor logic device, a hardware component or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. It may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like. The memory may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk.

[0169] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0170] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0171] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0172] The above technical features constitute the best embodiment of the present invention, which has strong adaptability and best implementation effect. Non-essential technical features can be added or reduced according to actual needs to meet the requirements of different situations.

Claims

1. A method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction, characterized in that: include: A training sample library is obtained and divided into a training set, a validation set and a test set, wherein the training sample library includes a plurality of sample data groups, each of which includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and labeling information of the current photovoltaic power generation value of the single power station; The spatiotemporal combined prediction model is trained using the training set in combination with the training termination condition, wherein the spatiotemporal combined prediction model includes an improved Pyraformer model, a hybrid skip graph convolutional network model with an RNN position encoding layer, and a splicing and fusion module. The improved Pyraformer model outputs a photovoltaic power prediction value based on time correlation, and the hybrid skip graph convolutional network model with an RNN position encoding layer outputs a photovoltaic power prediction value based on spatial correlation. The splicing and fusion module splices and combines the two prediction values ​​and assigns weights to obtain the final photovoltaic power prediction value of a single power station. The validation set and test set are used to verify and test the trained spatiotemporal combined prediction model respectively, the model parameters of the spatiotemporal combined prediction model are determined, and the spatiotemporal combined prediction model that meets the test and evaluation requirements is output.

2. The method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction according to claim 1, characterized in that: The improved Pyraformer model includes an encoder and a decoder. The encoder includes a global branch and a local branch. The global branch includes a global convolution kernel Gconv to extract global information. The local branch includes two modules, CVM and Pyraformer. CVM captures cross-dimensional dependencies and Pyraformer captures cross-temporal dependencies. The decoder includes an attention mechanism and a fully connected layer. The attention mechanism integrates global information and local information, and the model output result is obtained by inference through a fully connected layer.

3. The method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction according to claim 1 or 2, characterized in that: The hybrid skip graph convolutional network model with the RNN position encoding layer includes an RNN position encoding layer and a hybrid skip graph convolutional network. The position embedding and final representation of the historical photovoltaic power station power sequence are obtained through the RNN position encoding layer. The hybrid skip graph convolutional network includes two hybrid skip propagation layers, which process the inflow and outflow information of the nodes respectively. The final output information is the sum of the outputs of the two hybrid skip propagation layers.

4. The method for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction according to any one of claims 1 to 3, characterized in that: The splicing and fusion module includes a splicing layer and a fully connected layer. The splicing layer splices and combines the photovoltaic power prediction value based on time correlation and the photovoltaic power prediction value based on space correlation. The fully connected layer weights the spliced ​​and combined photovoltaic power prediction values ​​and outputs the final photovoltaic power prediction value.

5. A photovoltaic power generation prediction method based on a spatiotemporal combined prediction model, characterized in that: include: Obtaining historical photovoltaic power generation data and meteorological data of the target power station, and historical photovoltaic power generation data of multiple power stations adjacent to the target power station; The historical photovoltaic power generation data and meteorological data of the target power station, as well as the historical photovoltaic power generation data of multiple power stations adjacent to the target power station are input into a spatiotemporal combined prediction model to obtain a predicted value of the photovoltaic power generation of the target power station, wherein the spatiotemporal combined prediction model is obtained by the method as described in any one of claims 1 to 4.

6. The photovoltaic power generation prediction method based on the spatiotemporal combined prediction model according to claim 5 is characterized in that: It also includes verifying the rationality of selecting the historical photovoltaic power generation data of multiple power stations after obtaining the historical photovoltaic power generation data and meteorological data of the target power station and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station, including: Set rationality verification conditions to determine whether the historical photovoltaic power generation data of multiple power stations meet all rationality verification conditions, where the rationality verification conditions include: The Moran index is used to perform spatial autocorrelation analysis on the historical photovoltaic power data of multiple power stations to determine whether there is spatial autocorrelation in the historical photovoltaic power data of multiple power stations. Granger causality is used to verify whether the historical photovoltaic power generation data of multiple power stations can improve the prediction accuracy of the target power station; In response to no, the historical photovoltaic power generation data of the plurality of power stations are reselected.

7. A device for constructing a spatiotemporal combined prediction model for photovoltaic power generation prediction using the method according to any one of claims 1 to 4, characterized in that: include: A sample acquisition unit acquires a training sample library and divides it into a training set, a validation set and a test set, wherein the training sample library includes a plurality of sample data groups, each of which includes historical photovoltaic power generation data and meteorological data of a single power station, historical photovoltaic power generation data of multiple power stations adjacent to the single power station, and labeling information of the current photovoltaic power generation value of the single power station; A training unit, which trains the spatiotemporal combined prediction model using the training set in combination with the training termination condition, wherein the spatiotemporal combined prediction model includes an improved Pyraformer model, a hybrid skip graph convolutional network model with an RNN position encoding layer, and a splicing and fusion module. The improved Pyraformer model outputs a photovoltaic power prediction value based on time correlation, and the hybrid skip graph convolutional network model with an RNN position encoding layer outputs a photovoltaic power prediction value based on spatial correlation. The splicing and fusion module splices and combines the two prediction values ​​and assigns weights to obtain the final photovoltaic power prediction value of a single power station. The verification and testing unit uses the verification set and the test set to verify and test the trained spatiotemporal combined prediction model respectively, determines the model parameters of the spatiotemporal combined prediction model, and outputs the spatiotemporal combined prediction model that meets the test and evaluation requirements.

8. A photovoltaic power generation prediction device based on a spatiotemporal combined prediction model using the method as claimed in claim 5 or 6, characterized in that: include: A data acquisition unit, which acquires historical photovoltaic power generation data and meteorological data of a target power station, and historical photovoltaic power generation data of multiple power stations adjacent to the target power station; The prediction unit inputs the historical photovoltaic power generation data and meteorological data of the target power station, and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station into a spatiotemporal combined prediction model to obtain a predicted value of the photovoltaic power generation of the target power station, wherein the spatiotemporal combined prediction model is obtained by the method as described in any one of claims 1 to 4.

9. The photovoltaic power generation prediction device based on the spatiotemporal combined prediction model according to claim 8, characterized in that: It also includes a data verification unit, which verifies the rationality of selecting the historical photovoltaic power generation data of the multiple power stations after obtaining the historical photovoltaic power generation data and meteorological data of the target power station and the historical photovoltaic power generation data of multiple power stations adjacent to the target power station, including: Set rationality verification conditions to determine whether the historical photovoltaic power generation data of multiple power stations meet all rationality verification conditions, where the rationality verification conditions include: The Moran index is used to perform spatial autocorrelation analysis on the historical photovoltaic power data of multiple power stations to determine whether there is spatial autocorrelation in the historical photovoltaic power data of multiple power stations. Granger causality is used to verify whether the historical photovoltaic power generation data of multiple power stations can improve the prediction accuracy of the target power station; In response to no, the historical photovoltaic power generation data of the plurality of power stations are reselected.

10. An electronic device, characterized in that: It comprises a processor and a memory, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the photovoltaic power generation power prediction method based on the spatiotemporal combined prediction model as claimed in claim 5 or 6.