Distributed photovoltaic power generation prediction method and system based on dual-channel dynamic space-time graph

Abstract

The present invention discloses a distributed photovoltaic power generation prediction method and system based on a dual-channel dynamic space-time graph. The method achieves accurate prediction of spectrum trends through a first feature extraction channel, achieves adjusted prediction of influencing factors through a second feature extraction channel, and realizes meteorological time-shift correction prediction under neighbor node learning based on a distributed photovoltaic graph network. It comprehensively considers most of the important influencing factors that need to be considered in distributed photovoltaic time series prediction, greatly improving the accuracy of photovoltaic prediction results. In addition, it adopts spectrum attention, self-attention, and graph attention mechanisms to concentrate most of the training parameters of the neural network on the influencing factors that need to be paid attention to, focusing on and extracting only important information as much as possible, further improving the accuracy of photovoltaic prediction results, and can also output prediction results of multiple distributed photovoltaic fields at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed photovoltaic power generation prediction, and in particular to a distributed photovoltaic power generation prediction method and system based on a dual-channel dynamic space-time graph, an electronic device, and a computer-readable storage medium. Background Art

[0002] Distributed photovoltaic systems face challenges in the safe operation of power grids and their ability to absorb electricity as they develop worldwide. The fluctuating and intermittent nature of distributed photovoltaic power output poses challenges to the stability and safe operation of power grids. The increasing number of distributed photovoltaic power stations, especially those connected at low voltage levels, increases the safety risks faced by frontline power distribution operators in their daily operations and repairs. To address this, researchers are dedicated to developing various methods to make photovoltaic output more stable so that stable power can be safely utilized. Once the output of all distributed photovoltaic nodes can be accurately predicted, various scheduling methods can be used to achieve stable output scheduling. Therefore, how to simultaneously predict the power generation of all distributed photovoltaic nodes has become a key issue that needs to be addressed.

[0003] Existing photovoltaic power generation prediction methods generally use LSTM networks or convolutional neural networks to predict the power generation of a single photovoltaic node, but are unable to predict the power generation of multiple photovoltaic nodes simultaneously. Furthermore, existing photovoltaic power generation prediction methods typically use temporal convolutional networks to extract data features. However, the convolution operation cannot focus on key information. When extracting time series features, the temporal convolutional network will evenly affect noise signal features and useful signal features, resulting in poor model prediction accuracy. Furthermore, cloud movement, a key influencing factor in distributed photovoltaics, can cause time lags in the output between photovoltaic nodes. This means that the future output of another photovoltaic node may affect the output of the current photovoltaic node. Existing photovoltaic power generation prediction methods do not consider the impact of this time lag on power generation prediction, which also leads to poor prediction accuracy. Summary of the Invention

[0004] The present invention provides a distributed photovoltaic power generation prediction method and system based on a dual-channel dynamic space-time graph, an electronic device, and a computer-readable storage medium to solve the technical problems that existing photovoltaic power generation prediction methods cannot simultaneously predict the power generation power of multiple photovoltaic nodes and the prediction results have poor accuracy.

[0005] According to one aspect of the present invention, a distributed photovoltaic power generation prediction method based on a dual-channel dynamic spatiotemporal graph is provided, comprising the following contents:

[0006] Collect historical data from multiple distributed photovoltaic fields, including light intensity, temperature, dust particle concentration, time, and power generation;

[0007] Based on historical data, a variable matrix is constructed for each distributed photovoltaic field. The correlation between multiple distributed photovoltaic fields is evaluated based on the variable matrix. A distributed photovoltaic graph network is constructed based on the correlation evaluation results. The data dimensions of the variable matrix include light intensity, temperature, dust particle concentration, and power generation.

[0008] A dual-channel prediction model is constructed based on a distributed photovoltaic graph network. The dual-channel prediction model is trained using historical data from multiple distributed photovoltaic fields. After training, the model parameters are saved. The dual-channel prediction model includes a first feature extraction channel using a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel using a self-attention mechanism and a graph attention mechanism.

[0009] The trained dual-channel prediction model is used to simultaneously predict the power generation of multiple distributed photovoltaic fields.

[0010] Furthermore, the process of constructing a variable matrix for each distributed photovoltaic field based on historical data, evaluating the correlation between multiple distributed photovoltaic fields based on the variable matrix, and constructing a distributed photovoltaic map network based on the correlation evaluation results includes the following:

[0011] Each distributed photovoltaic field is regarded as a photovoltaic node. Based on the light intensity, temperature, dust particle concentration and power generation of each photovoltaic node, an L×4 variable matrix is constructed, where L represents the length of the historical data sequence. Each variable matrix is normalized by row and reshaped into a one-dimensional variable vector. The Pearson correlation coefficient between each photovoltaic node and other photovoltaic nodes is calculated based on the one-dimensional variable vector. If the Pearson correlation coefficient between two photovoltaic nodes is greater than a preset threshold, it is determined that a node edge exists between the two photovoltaic nodes. Otherwise, it is determined that no node edge exists between the two photovoltaic nodes. In this way, a distributed photovoltaic graph network that can be dynamically adjusted in real time is constructed.

[0012] Furthermore, the dual-channel prediction model includes:

[0013] The input layer architecture includes a first input layer and a second input layer in parallel. The first input layer is used to perform triple data embedding and fusion on the input data to perform multi-dimensional amplification processing on the data features. The second input layer is used to convert the input data into a text code sequence and place the power generation at the center of the text code sequence.

[0014] The middle layer architecture includes a multi-layer learning model. Each layer of the learning model includes multiple feature extraction modules and multiple graph attention networks with one-to-one correspondence. Each feature extraction module includes a parallel multidimensional feature extractor and a Bert model. The multidimensional feature extractor uses a spectral attention mechanism to extract features of the first channel from the fused features output by the first input layer. The Bert model uses a self-attention mechanism to extract features of the second channel from the text encoding sequence output by the second input layer. The graph attention network is constructed based on the network structure of a distributed photovoltaic graph network. The graph attention network uses a graph attention mechanism to self-learn the hysteresis characteristics of the target photovoltaic field based on the neighboring photovoltaic fields, so as to perform meteorological time-shift correction on the features extracted from the first and second channels respectively and output them.

[0015] The output layer architecture includes multiple decoding layers. The number of decoding layers is the same as the number of distributed photovoltaic fields. That is, each node of the graph attention network is connected to a corresponding decoding layer. Each decoding layer includes a dual-channel fully connected layer and a single-channel fully connected layer, which is used to perform weighted fusion of the corrected features of the two channels and simultaneously output the power generation prediction results of multiple distributed photovoltaic fields.

[0016] Furthermore, the data embedding method adopted by the first input layer includes time feature coding, value coding and position coding.

[0017] Furthermore, the graph attention network calculates the self-learned weighted features based on the following formula,

[0018]

[0019] Where h′ i (P) represents weighted features, P represents the number of heads in the multi-head attention mechanism, || represents the vector concatenation operation, and the subscripts i and j represent the target photovoltaic field and the neighboring photovoltaic field, respectively. represents the attention coefficient of the neighboring photovoltaic field j calculated by the p-th head, represents the attention coefficient of target photovoltaic field i calculated by the p-th head, W p represents the learnable parameters of the p-th head, h j represents the feature vector of the neighboring photovoltaic field j extracted by the multidimensional feature extractor or Bert model, h i Represents the feature vector of target photovoltaic field i extracted by multidimensional feature extractor or Bert model, N i represents the set of neighbor nodes of the target photovoltaic field i, and σ() represents the activation function.

[0020] Furthermore, each head calculates the attention coefficient of the neighboring photovoltaic field based on the following formula:

[0021]

[0022] Among them, α ij represents the attention coefficient of the neighboring photovoltaic field j, e ij and e ik They represent the similarity coefficients between the neighboring photovoltaic field j and the neighboring photovoltaic field k and the target photovoltaic field i, respectively, and LeakyReLU() represents the activation function.

[0023] Furthermore, each head calculates the similarity coefficient between the neighboring photovoltaic field and the target photovoltaic field based on the following formula:

[0024] e ij =a([Wh i ||Wh j ]),j∈N i

[0025] Among them, a() represents the mapping function that maps high-dimensional features to real numbers e, W represents the learnable parameters, and h i and h j They represent the feature vectors of target photovoltaic field i and neighboring photovoltaic field j extracted by the multidimensional feature extractor or Bert model respectively.

[0026] In addition, the present invention also provides a distributed photovoltaic power generation prediction system based on a dual-channel dynamic spatiotemporal graph, comprising:

[0027] A data acquisition module is used to collect historical data of multiple distributed photovoltaic fields, where the historical data types include light intensity, temperature, dust particle concentration, time and power generation;

[0028] A graph network construction module is used to construct a variable matrix for each distributed photovoltaic field based on historical data, evaluate the correlation between multiple distributed photovoltaic fields based on the variable matrix, and construct a distributed photovoltaic graph network based on the correlation evaluation results. The data dimensions of the variable matrix include light intensity, temperature, dust particle concentration, and power generation;

[0029] A model training module is used to build a dual-channel prediction model based on a distributed photovoltaic graph network, train the dual-channel prediction model using historical data from multiple distributed photovoltaic fields, and save the model parameters after training. The dual-channel prediction model includes a first feature extraction channel using a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel using a self-attention mechanism and a graph attention mechanism.

[0030] The power generation prediction module is used to simultaneously predict the power generation of multiple distributed photovoltaic fields using the trained dual-channel prediction model.

[0031] In addition, the present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the steps of the above method by calling the computer program stored in the memory.

[0032] In addition, the present invention also provides a computer-readable storage medium for storing a computer program for predicting distributed photovoltaic power generation based on a dual-channel dynamic spatiotemporal diagram, wherein the computer program executes the steps of the method described above when running on a computer.

[0033] Compared with the existing technology, this application has the following beneficial effects:

[0034] The distributed photovoltaic power generation prediction method based on a dual-channel dynamic spatiotemporal graph of the present invention, after collecting historical data from multiple distributed photovoltaic fields, can construct a variable matrix based on the historical data, evaluate the correlations between the multiple distributed photovoltaic fields, and build a distributed photovoltaic graph network. Unlike traditional graph networks, this distributed photovoltaic graph network can automatically adjust the neighboring edges of each node to achieve real-time dynamic network updates. The distributed photovoltaic graph network is used to analyze the mutual influence between different photovoltaic nodes. Furthermore, a dual-channel prediction model is constructed based on the distributed photovoltaic graph network. The dual-channel model includes a first feature extraction channel that uses a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel that uses a self-attention mechanism and a graph attention mechanism. The first channel uses a frequency domain feature attention enhancement mechanism, focusing on high-frequency spectrum portions and treating low-frequency spectrum portions as noise and not learning them. This pays more attention to the overall data trend. It also integrates the meteorological data spectrum with the focus of the photovoltaic output spectrum of interest. Combined with dimensional amplification and fusion under multiple meteorological factors, it provides more accurate predictions of data details and trends. The second channel, thanks to its self-attention mechanism, can extract features within sequences, between sequences, and between cycles, compensating for the first channel's limitation of being unable to fine-tune future PV output using timestamps and future meteorological factors. Furthermore, both feature extraction channels employ a graph attention mechanism based on the constructed distributed PV graph network. This mechanism self-learns based on the lag characteristics of neighboring PV nodes regarding the target PV node, fusing the features of the current PV node with those of its neighboring nodes to determine the impact of the neighboring nodes' future output on the current PV node's output. This improves the prediction accuracy of individual PV nodes and enables simultaneous prediction of every PV node in the distributed PV graph network. Therefore, the distributed photovoltaic power generation prediction method based on the dual-channel dynamic space-time graph of the present invention realizes accurate prediction of spectrum trends through the first feature extraction channel, realizes adjustment prediction of influencing factors through the second feature extraction channel, and realizes meteorological time-shift correction prediction under neighbor node learning based on the distributed photovoltaic graph network. It comprehensively considers most of the important influencing factors that need to be considered in distributed photovoltaic time series prediction, greatly improves the accuracy of photovoltaic prediction results, and adopts spectrum attention, self-attention, and graph attention mechanisms to concentrate most of the training parameters of the neural network on the influencing factors that need to be paid attention to, and focuses on and extracts only important information as much as possible, further improving the accuracy of photovoltaic prediction results, and can also output prediction results of multiple distributed photovoltaic fields at the same time.

[0035] In addition, the distributed photovoltaic power generation prediction system based on the dual-channel dynamic space-time graph of the present invention also has the above advantages.

[0036] In addition to the above-described purposes, features, and advantages, this application also has other purposes, features, and advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0038] Figure 1 It is a flow chart of a distributed photovoltaic power generation prediction method based on a dual-channel dynamic space-time graph in a preferred embodiment of the present application.

[0039] Figure 2 It is a schematic diagram of the network structure of the intermediate layer architecture of the dual-channel prediction model in the preferred embodiment of the present application.

[0040] Figure 3 This is a schematic diagram of three data embedding methods of the Bert model in the preferred embodiment of the present application for embedding time series data.

[0041] Figure 4 It is a schematic diagram of the principle of feature processing performed by the first feature extraction channel in the preferred embodiment of the present application.

[0042] Figure 5 It is a schematic diagram of the principle of feature processing performed by the second feature extraction channel in the preferred embodiment of the present application.

[0043] Figure 6 It is a schematic diagram of the network structure of the output layer architecture of the dual-channel prediction model in the preferred embodiment of the present application.

[0044] Figure 7 This is a schematic diagram of the module structure of a distributed photovoltaic power generation prediction system based on a dual-channel dynamic space-time graph in another embodiment of the present application. DETAILED DESCRIPTION

[0045] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0046] Reference Figure 1 The preferred embodiment of the present application provides a distributed photovoltaic power generation prediction method based on a dual-channel dynamic spatiotemporal graph, including the following contents:

[0047] Step S1: collecting historical data of multiple distributed photovoltaic fields, wherein the historical data types include light intensity, temperature, dust particle concentration, time and power generation;

[0048] Step S2: constructing a variable matrix for each distributed photovoltaic field based on historical data, evaluating the correlation between multiple distributed photovoltaic fields based on the variable matrix, and constructing a distributed photovoltaic graph network based on the correlation evaluation results, wherein the data dimensions of the variable matrix include light intensity, temperature, dust particle concentration, and power generation;

[0049] Step S3: constructing a dual-channel prediction model based on the distributed photovoltaic graph network, using historical data from multiple distributed photovoltaic fields to train the dual-channel prediction model, and saving the model parameters after training. The dual-channel prediction model includes a first feature extraction channel using a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel using a self-attention mechanism and a graph attention mechanism.

[0050] Step S4: using the trained dual-channel prediction model to simultaneously predict the power generation of multiple distributed photovoltaic fields.

[0051] It can be understood that the distributed photovoltaic power generation prediction method based on a dual-channel dynamic spatiotemporal graph in this embodiment, after collecting historical data from multiple distributed photovoltaic fields, can construct a variable matrix based on the historical data, evaluate the correlations between the multiple distributed photovoltaic fields, and build a distributed photovoltaic graph network. Unlike traditional graph networks, this distributed photovoltaic graph network can automatically adjust the neighboring edges of each node to achieve real-time dynamic network updates. The distributed photovoltaic graph network is used to analyze the mutual influence between different photovoltaic nodes. Furthermore, a dual-channel prediction model is constructed based on the distributed photovoltaic graph network. The dual-channel model includes a first feature extraction channel that uses a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel that uses a self-attention mechanism and a graph attention mechanism. The first channel uses a frequency domain feature attention enhancement mechanism, focusing on high-frequency spectrum portions and treating low-frequency spectrum portions as noise and not learning them. This increases the focus on the overall data trend. It also integrates the meteorological data spectrum with the focus of the photovoltaic output spectrum. Combined with dimensional amplification and fusion under multiple meteorological factors, it provides more accurate predictions of data details and trends. The second channel, thanks to its self-attention mechanism, can extract features within sequences, between sequences, and between cycles, compensating for the first channel's limitation of being unable to fine-tune future PV output using timestamps and future meteorological factors. Furthermore, both feature extraction channels employ a graph attention mechanism based on the constructed distributed PV graph network. This mechanism self-learns based on the lag characteristics of neighboring PV nodes regarding the target PV node, fusing the features of the current PV node with those of its neighboring nodes to determine the impact of the neighboring nodes' future output on the current PV node's output. This improves the prediction accuracy of individual PV nodes and enables simultaneous prediction of every PV node in the distributed PV graph network. Therefore, the distributed photovoltaic power generation prediction method based on the dual-channel dynamic space-time graph of the present invention realizes accurate prediction of spectrum trends through the first feature extraction channel, realizes adjustment prediction of influencing factors through the second feature extraction channel, and realizes meteorological time-shift correction prediction under neighbor node learning based on the distributed photovoltaic graph network. It comprehensively considers most of the important influencing factors that need to be considered in distributed photovoltaic time series prediction, greatly improves the accuracy of photovoltaic prediction results, and adopts spectrum attention, self-attention, and graph attention mechanisms to concentrate most of the training parameters of the neural network on the influencing factors that need to be paid attention to, and focuses on and extracts only important information as much as possible, further improving the accuracy of photovoltaic prediction results, and can also output prediction results of multiple distributed photovoltaic fields at the same time.

[0052] It can be understood that in step S1, after research, the biggest factors affecting the efficiency of solar photovoltaic power generation are light radiation intensity, temperature and the number of particles in the air. The influence of these three factors accounts for the majority of the factors affecting photovoltaic power generation efficiency. Therefore, the present invention mainly collects historical data of these three factors. For example, there are N groups of distributed photovoltaic fields in a region, and the light intensity, temperature, dust particle density and power generation of each photovoltaic field are respectively They represent the light intensity, temperature, dust particle density and power generation of the i-th distributed photovoltaic field at time t, and the historical light intensity data collected from N distributed photovoltaic fields can be expressed as The historical data collection results of the other three are similar and will not be repeated here.

[0053] It can be understood that in step S2, the process of constructing a variable matrix of each distributed photovoltaic field based on historical data, evaluating the correlation between multiple distributed photovoltaic fields based on the variable matrix, and constructing a distributed photovoltaic map network based on the correlation evaluation results includes the following:

[0054] Each distributed photovoltaic field is regarded as a photovoltaic node. Based on the light intensity, temperature, dust particle concentration and power generation of each photovoltaic node, an L×4 variable matrix is constructed, where L represents the length of the historical data sequence. Each variable matrix is normalized by row and reshaped into a one-dimensional variable vector. The Pearson correlation coefficient between each photovoltaic node and other photovoltaic nodes is calculated based on the one-dimensional variable vector. If the Pearson correlation coefficient between two photovoltaic nodes is greater than a preset threshold, it is determined that a node edge exists between the two photovoltaic nodes. Otherwise, it is determined that no node edge exists between the two photovoltaic nodes. In this way, a distributed photovoltaic graph network that can be dynamically adjusted in real time is constructed.

[0055] The calculation formula of Pearson correlation coefficient is: Among them, cov() represents covariance, X i and X j Represent the one-dimensional variable vectors of the i-th and j-th photovoltaic nodes, represents the correlation coefficient between the i-th and j-th PV nodes, σ i and σ j Represents the standard deviation of the one-dimensional variable vector of the i-th and j-th photovoltaic nodes respectively. When , it is determined that there is a connection edge between the i-th photovoltaic node and the j-th photovoltaic node, and the two have a mutual influence. When , there is no connection edge between the i-th photovoltaic node and the j-th photovoltaic node, and there is no mutual influence between the two, thus constructing a distributed photovoltaic graph network, and the network will be dynamically adjusted in real time according to time changes.

[0056] It can be understood that in step S3, the dual-channel prediction model specifically includes:

[0057] The input layer architecture includes a first input layer and a second input layer in parallel. The first input layer is used to perform triple data embedding and fusion on the input data to perform multi-dimensional amplification processing on the data features. The second input layer is used to convert the input data into a text code sequence and place the power generation at the center of the text code sequence.

[0058] The middle layer architecture includes a multi-layer learning model. Each layer of the learning model includes multiple feature extraction modules and multiple graph attention networks with one-to-one correspondence. Each feature extraction module includes a parallel multidimensional feature extractor and a Bert model. The multidimensional feature extractor uses a spectral attention mechanism to extract features of the first channel from the fused features output by the first input layer. The Bert model uses a self-attention mechanism to extract features of the second channel from the text encoding sequence output by the second input layer. The graph attention network is constructed based on the network structure of a distributed photovoltaic graph network. The graph attention network uses a graph attention mechanism to self-learn the hysteresis characteristics of the target photovoltaic field based on the neighboring photovoltaic fields, so as to perform meteorological time-shift correction on the features extracted from the first and second channels respectively and output them.

[0059] The output layer architecture includes multiple decoding layers. The number of decoding layers is the same as the number of distributed photovoltaic fields. That is, each node of the graph attention network is connected to a corresponding decoding layer. Each decoding layer includes a dual-channel fully connected layer and a single-channel fully connected layer, which is used to perform weighted fusion of the corrected features of the two channels and simultaneously output the power generation prediction results of multiple distributed photovoltaic fields.

[0060] The data embedding methods used in the first input layer include time feature coding, value coding and position coding. Specifically, the time feature coding is the dimension amplification after the time tag decomposition, which is specifically achieved by the full connection layer connection amplification, and the amplification dimension is d model , time feature encoding can amplify the characteristics of time labels, so that the model can better learn the corresponding characteristics of periodic labels in periodic data. Value encoding uses one-dimensional convolution operation to extract the waveform characteristics of input features and uses d model A one-dimensional convolution kernel expands the dim dimension of the input data to d model Dimension, through the multi-dimensional extraction of waveform features, thereby amplifying the overall characteristics of the data waveform. In addition, the position encoding is to embed the data into the trigonometric function by using the cross encoding of sin and cos, using d modelThe trigonometric function of this period is embedded into the waveform data, which introduces multiple dimensions of position information into the input sequence data, so that the model can capture the order and relative position relationship in the data, and enlarge the latitude of the input data from dim×L to d model × L. Among them, d model Generally, it is 32. Among them, these three encoding mechanisms are commonly used encoding mechanisms, and the specific encoding principles are not described here.

[0061] It can be understood that the present invention adopts three commonly used encoding mechanisms for data embedding, integrating the data position relationship, the data's own characteristics and the time label characteristics, so that the hidden relationship in the original data is easier to be learned by the subsequent network, and realizes data mining information amplification. The data after feature amplification has richer information content, ensuring that the multidimensional feature extractor can extract richer and more comprehensive data features, which is conducive to improving the accuracy of the prediction results.

[0062] In addition, the second input layer can convert the input data into a text encoding sequence and place the power generation data in the center of the text encoding sequence, so that the information of the upper and lower paragraphs can be captured through the context learning of the Bert model, so that the power generation is associated with the meteorological factor data and time data before and after. For example, for the historical data example in Table 1, text reverse encoding is performed, and the two rows of data are encoded as [TS][M]01[D]01[HH]08[MM]00[P]100[L]12[T]23[DU]78[SEP][M]01[D]01[HH]09[MM]00[P]120[L]14[T]24[DU]79, where [TS] represents the start of the time series, [M] represents the month, [D] represents the date, and [HH] represents the hour. [MM] represents minutes, [P] represents electric power, [L] represents light intensity, [T] represents temperature, [DU] represents dust particle size, and [SEP] represents the separation of different time points. In this way, each set of historical data can be connected into a character sentence. For example, the sentence connected in the first row of Table 1 is [M]01[D]01[HH]08[MM]00[P]100[L]12[T]23[DU]78, thereby encoding a set of input historical data into text data of length 8.

[0063] Table 1. Example of collected historical data

[0064]

[0065] It is understandable that Figure 2As shown, the middle layer architecture of the dual-channel prediction model includes a multi-layer learning model, and each layer of the learning model includes multiple feature extraction modules and multiple graph attention networks corresponding to each other. Each feature extraction module includes a parallel multidimensional feature extractor and a Bert model, that is, the multidimensional feature extractor serves as the first feature extraction channel, and the Bert model serves as the second feature extraction channel. The multidimensional feature extractor of each layer is connected to a graph attention network, and the Bert model of each layer is also connected to a graph attention network. The multidimensional feature extractor uses the spectral attention mechanism to extract the features of the first channel of the fusion features output by the first input layer. The Bert model uses the self-attention mechanism to extract the features of the second channel of the text encoding sequence output by the second input layer. The graph attention network is constructed based on the network structure of the distributed photovoltaic graph network. Each photovoltaic node in the distributed photovoltaic graph network is used as a node in the graph attention network. The graph attention network uses the graph attention mechanism to self-learn the hysteresis characteristics of the target photovoltaic field according to the neighboring photovoltaic fields, so as to perform meteorological time-shift correction on the features extracted from the first channel and the second channel respectively and output them.

[0066] Specifically, the multi-dimensional feature extractor (Uni-dim extractor block) is composed of d model Each extractor block is used to extract the implicit information in one-dimensional data. Each extractor block first uses the fast Fourier transform (FFT) method to amplify the features in a single dimension, that is, to extract the features in the d dimension. model Each dimension uses FFT amplification and innovatively uses the efficient two-dimensional image feature extraction module Inception to extract d model Dimension de FFT processing data performs two-dimensional convolution feature extraction, and finally d model The convolution extraction data of the dimension is transformed and the features are added and fused. Among them, the input dimension and output dimension of the multidimensional feature extractor are both d model ×L. The Uni-dim extractor block is an existing neural network module, and the specific principle will not be repeated here.

[0067] As a mature pre-training model, the Bert model can memorize most photovoltaic scenarios based on its strong memory capacity, thereby obtaining the universal spatiotemporal characteristics of distributed photovoltaics. It does not require photovoltaic data in a specific time and space, and can make relatively accurate predictions based on only some factors such as meteorological and time data, and has high portability. Figure 3As shown, the BERT model has three data embedding methods: token embedding, segment embedding, and position embedding. All three data embedding methods are existing technologies, and the specific principles are not detailed here. When applied to time series data, the BERT model's three data embedding methods enable it to not only understand the specific meaning of each data point, but also capture the data's temporal order and possible contextual relationships. This understanding capability is crucial for efficient time series analysis, predicting future trends, and identifying abnormal patterns. Ultimately, the three embedding dimensions are combined to produce a 768-dimensional sequence feature representation.

[0068] In addition, the Graph Attention Networks (GAT) is constructed based on the distributed photovoltaic field graph network. Each distributed photovoltaic field (i.e., photovoltaic node) in the distributed photovoltaic field graph network is used as a vertex in the graph attention network, and the connection edge between two photovoltaic nodes is used as the connection between two vertices in the graph attention network, thereby constructing a GAT-based neighboring node feature learning framework. Its purpose is to perform graph attention mechanism self-learning on the hysteresis characteristics of the target photovoltaic field based on the neighboring photovoltaic fields, so as to improve the accuracy of the prediction results of a single photovoltaic field. Among them, the graph attention network calculates the self-learned weighted features based on the following formula,

[0069]

[0070] Where h′ i (P) represents weighted features, P represents the number of heads in the multi-head attention mechanism, || represents the vector concatenation operation, and the subscripts i and j represent the target photovoltaic field and the neighboring photovoltaic field, respectively. represents the attention coefficient of the neighboring photovoltaic field j calculated by the p-th head, represents the attention coefficient of target photovoltaic field i calculated by the p-th head, W p represents the learnable parameters of the p-th head, h j represents the feature vector of the neighboring photovoltaic field j extracted by the multidimensional feature extractor or Bert model, h i Represents the feature vector of target photovoltaic field i extracted by multidimensional feature extractor or Bert model, N i represents the neighbor node set of the target photovoltaic field i, and σ() represents the activation function. It can be understood that, as a preferred embodiment,

[0071] Among them, each head calculates the attention coefficient of the neighboring photovoltaic field based on the following formula:

[0072]

[0073] Among them, α ij represents the attention coefficient of the neighboring photovoltaic field j, e ij and e ik They represent the similarity coefficients between the neighboring photovoltaic field j and the neighboring photovoltaic field k and the target photovoltaic field i, respectively, and LeakyReLU() represents the activation function.

[0074] Each head calculates the similarity coefficient between the neighboring photovoltaic field and the target photovoltaic field based on the following formula:

[0075] e ij =a([Wh i ||Wh j ]),j∈N i

[0076] Among them, a() represents the mapping function that maps high-dimensional features to real numbers e, W represents the learnable parameters, and h i and h j They represent the feature vectors of target photovoltaic field i and neighboring photovoltaic field j extracted by the multidimensional feature extractor or Bert model, Wh i ||Wh j This means that the eigenvector h i and h j After multiplying by the shared parameter W, vector splicing is performed to increase the dimension of the vertex features in the graph attention network. The spliced feature vector is then mapped to the real number e through the mapping function a(). The mapping function a() and the learnable parameter W are used to learn the correlation between vertices i and j.

[0077] It can be understood that the graph attention network of the present invention first allocates attention to each neighboring node based on the correlation between different nodes, so that each node performs dual-channel information fusion of the characteristics of the neighboring nodes, and draws on the multi-head attention mechanism of the transformer architecture to use multiple convolution kernels to enhance the robustness of the model, thereby abstracting the location influence information such as cloud layer and wind speed of the current node into the past information of the neighboring nodes, that is, performing lagging feature learning, which can drive the entire model to learn the time-shift characteristics of the surrounding distributed photovoltaic fields, thereby performing meteorological time-shift correction on the features extracted by the dual channels.

[0078] It is understandable that Figure 4 As shown, the historical data sequence of the i-th distributed photovoltaic field The dimensions include e i 、T i d i 、p i Four dimensions, namely dim is 4, L is the length of the historical data sequence, T Lis the time label sequence, the input data sequence and T L First, the triple embedding of the first input layer is weighted to obtain the fusion feature Then the feature h is extracted by the multi-dimensional feature extractor Uni-dim extractor block i , and then updated to feature h′ after the first layer of GAT learning i1 , where h i and h′ i1 The latitude of the i-th distributed photovoltaic field is d model ×L features. Figure 5 As shown, the collected historical data sequence is converted into the text encoding of the second input layer. T and e to be collected i , t i d i 、p i The historical data of these five dimensions are integrated and encoded into text data of length 8. The length of 8L represents L groups of input data of the i-th distributed photovoltaic field, and L is the length of the historical data sequence. First, the triple encoding embedding of the Bert model is weighted to obtain the fusion word embedding feature Among them, 768 is the default embedding dimension of the Bert model, and the fusion word embedding feature Then it is sent to the multi-layer Transformer network of the Bert model to process multiple groups of dimensions to capture upstream and downstream related information. The Bert model outputs the final representation h of each word. i , and then after the first layer of GAT learning, the updated feature is h i1 , where h i With h i1 are all word-unit context features of the i-th distributed photovoltaic field with a dimension of 8L×768. Figure 2 , h output by the first layer learning model i1 and h′ i1 , and then output h after the second layer learning model i2 and h′ i2 , iterate continuously, and finally get the output h of the m-th layer learning model im and h′ im , m does not exceed 6, and is generally 2 to 4. It can be understood that in the middle layer architecture, the two feature extraction channels are parallel to each other and have no intersection, that is, the features after amplification by the first input layer are After the m-layer learning model, the feature update process is as follows: i1 , h′ i2 ,…,h′ im , and the features after the second input layer encoding conversion After the m-layer learning model, the feature update process is as follows: i1 、h i2 ,…,h im , the two are only fused in the final output layer.

[0079] In addition, the Bert model is a pre-trained data embedding and feature extraction model based on the transformer architecture. It is often used for semantic understanding. The Bert model can obtain bidirectional contextual information in the sequence, that is, the impact of the future on the present and the impact of the present on the future, while the transformer obtains unidirectional information, that is, the impact of the present on the future. In existing time series predictions, it is usually only considered that the future is affected by the present state. The effect of the future learned by the Bert model on the present is not beneficial, but may lead to prediction bias. Therefore, the transformer architecture is often used for time series prediction, and the Bert model is not used for time series prediction. In the second feature extraction channel of the present invention, the Bert model is combined with the GAT network. The Bert model masks each power generation in the text encoding sequence through the masked language model MLM, so that the power generation is associated with the previous and next weather and timestamps. The contextual relationship of the power in the current sentence is captured through Token Embedding (internal relationship in the sentence) and Segment Embedding (contextual relationship in the sentence), and the surrounding node information of the current node is obtained through the GAT network for feature fusion. The impact characteristics of the future output of the adjacent point on the current output of the current node can be obtained, and the self-learning of the hysteresis feature is realized, which makes up for the deviation caused by the Bert model capturing the impact of the future on the present, so that the hysteresis characteristics and meteorological effects of distributed photovoltaics can be combined for refined prediction.

[0080] It can be understood that the output layer of the model is immediately followed by the middle layer, which includes multiple decoding layers. The number of decoding layers is the same as the number of distributed photovoltaic fields. That is, each node of the graph attention network is connected to a decoding layer. Each decoding layer includes a dual-channel fully connected layer and a single-channel fully connected layer, which is used to perform weighted fusion on the features corrected by the two channels and simultaneously output the power generation prediction results of multiple distributed photovoltaic fields. For example, Figure 6 As shown, the last layer of the intermediate layer learns the h′ output by the model im 、h imThe output is sent to a dual-channel fully connected layer, and the two outputs of the dual-channel fully connected layer are then simultaneously input to a single-channel fully connected layer to achieve decoding of the prediction task. It can be understood that the present invention self-learns the important coefficients of each factor in a fully connected manner at the output layer, and realizes the fusion and precise correction of dual-channel features through a multi-layer perceptron network, avoiding the prediction bias caused by the inaccurate weights of the first feature extraction channel and the second feature extraction channel that are manually and subjectively assigned. Moreover, each vertex in the graph attention network corresponds to a prediction value, so that multiple distributed photovoltaic fields can be predicted simultaneously.

[0081] It can be understood that the dual-channel prediction model of the present invention can achieve precise prediction through the first feature extraction channel and general prediction through the second feature extraction channel. Both of them are also combined with the meteorological time-shift correction of the graph attention network to construct a large photovoltaic prediction model based on the dual-channel dynamic space-time graph, which can not only be quickly applied to different distributed photovoltaic scenarios, but also has high prediction accuracy.

[0082] It can be understood that in step S4, when making predictions, the collected historical data of multiple distributed photovoltaic fields are input into the trained dual-channel prediction model, and the output prediction value of each distributed photovoltaic field at multiple moments in the future can be output.

[0083] In addition, if Figure 7 As shown, another embodiment of the present invention further provides a distributed photovoltaic power generation prediction system based on a dual-channel dynamic spatiotemporal graph, preferably using the above-mentioned method, including:

[0084] A data acquisition module is used to collect historical data of multiple distributed photovoltaic fields, where the historical data types include light intensity, temperature, dust particle concentration, time and power generation;

[0085] A graph network construction module is used to construct a variable matrix for each distributed photovoltaic field based on historical data, evaluate the correlation between multiple distributed photovoltaic fields based on the variable matrix, and construct a distributed photovoltaic graph network based on the correlation evaluation results. The data dimensions of the variable matrix include light intensity, temperature, dust particle concentration, and power generation;

[0086] A model training module is used to build a dual-channel prediction model based on a distributed photovoltaic graph network, train the dual-channel prediction model using historical data from multiple distributed photovoltaic fields, and save the model parameters after training. The dual-channel prediction model includes a first feature extraction channel using a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel using a self-attention mechanism and a graph attention mechanism.

[0087] The power generation prediction module is used to simultaneously predict the power generation of multiple distributed photovoltaic fields using the trained dual-channel prediction model.

[0088] It can be understood that the distributed photovoltaic power generation prediction system based on a dual-channel dynamic spatiotemporal graph in this embodiment, after collecting historical data from multiple distributed photovoltaic fields, can construct a variable matrix based on the historical data, evaluate the correlations between the multiple distributed photovoltaic fields, and build a distributed photovoltaic graph network. Unlike traditional graph networks, this distributed photovoltaic graph network can automatically adjust the neighboring edges of each node to achieve real-time dynamic network updates. The distributed photovoltaic graph network is used to analyze the mutual influence between different photovoltaic nodes. Furthermore, a dual-channel prediction model is constructed based on the distributed photovoltaic graph network. The dual-channel model includes a first feature extraction channel that uses a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel that uses a self-attention mechanism and a graph attention mechanism. The first channel uses a frequency domain feature attention enhancement mechanism, focusing on high-frequency spectrum portions and treating low-frequency spectrum portions as noise and not learning them. This increases the focus on the overall data trend. It also integrates the meteorological data spectrum with the focus of the photovoltaic output spectrum. Combined with dimensional amplification and fusion under multiple meteorological factors, it provides more accurate predictions of data details and trends. The second channel, thanks to its self-attention mechanism, can extract features within sequences, between sequences, and between cycles, compensating for the first channel's limitation of being unable to fine-tune future PV output using timestamps and future meteorological factors. Furthermore, both feature extraction channels employ a graph attention mechanism based on the constructed distributed PV graph network. This mechanism self-learns based on the lag characteristics of neighboring PV nodes regarding the target PV node, fusing the features of the current PV node with those of its neighboring nodes to determine the impact of the neighboring nodes' future output on the current PV node's output. This improves the prediction accuracy of individual PV nodes and enables simultaneous prediction of every PV node in the distributed PV graph network. Therefore, the distributed photovoltaic power generation prediction system based on the dual-channel dynamic space-time graph of the present invention realizes accurate prediction of spectrum trends through the first feature extraction channel, realizes adjustment prediction of influencing factors through the second feature extraction channel, and realizes meteorological time-shift correction prediction under neighbor node learning based on the distributed photovoltaic graph network. It comprehensively considers most of the important influencing factors that need to be considered in distributed photovoltaic time series prediction, greatly improves the accuracy of photovoltaic prediction results, and adopts spectrum attention, self-attention, and graph attention mechanisms to concentrate most of the training parameters of the neural network on the influencing factors that need to be paid attention to, focusing on important information as much as possible and extracting only important information, further improving the accuracy of photovoltaic prediction results, and can also output prediction results of multiple distributed photovoltaic fields at the same time.

[0089] In addition, another embodiment of the present invention further provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the steps of the above method by calling the computer program stored in the memory.

[0090] In addition, another embodiment of the present invention further provides a computer-readable storage medium for storing a computer program for predicting distributed photovoltaic power generation based on a dual-channel dynamic spatiotemporal diagram, wherein the computer program executes the steps of the method described above when running on a computer.

[0091] Common computer-readable storage media include: floppy disks, flexible disks, hard disks, magnetic tape, any other magnetic media, CD-ROMs, any other optical media, punch cards, paper tape, any other physical medium with a pattern of holes, random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), flash-erasable programmable read-only memory (FLASH-EPROM), any other memory chip or cartridge, or any other medium that can be read by a computer. Instructions can further be transmitted or received via a transmission medium. The term transmission medium may include any tangible or intangible medium that can be used to store, encode, or carry instructions for execution by a machine, and includes digital or analog communication signals or other intangible media that facilitate communication of such instructions. Transmission media include coaxial cables, copper wire, and fiber optics, including the wires of a bus used to transmit a computer data signal.

[0092] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

[0093] 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 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 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 embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0094] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the 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 the 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 produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. 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.

[0095] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0097] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0098] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A distributed photovoltaic power generation prediction method based on a dual-channel dynamic space-time graph, characterized in that: Includes the following: Collect historical data from multiple distributed photovoltaic fields, including light intensity, temperature, dust particle concentration, time, and power generation; Based on historical data, a variable matrix is constructed for each distributed photovoltaic field. The correlation between multiple distributed photovoltaic fields is evaluated based on the variable matrix. A distributed photovoltaic graph network is constructed based on the correlation evaluation results. The data dimensions of the variable matrix include light intensity, temperature, dust particle concentration, and power generation. A dual-channel prediction model is constructed based on a distributed photovoltaic graph network. The dual-channel prediction model is trained using historical data from multiple distributed photovoltaic fields. After training, the model parameters are saved. The dual-channel prediction model includes a first feature extraction channel using a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel using a self-attention mechanism and a graph attention mechanism. Use the trained dual-channel prediction model to simultaneously predict the power generation of multiple distributed photovoltaic fields; The dual-channel prediction model includes: The input layer architecture includes a first input layer and a second input layer in parallel. The first input layer is used to perform triple data embedding and fusion on the input data to perform multi-dimensional amplification processing on the data features. The second input layer is used to convert the input data into a text code sequence and place the power generation at the center of the text code sequence. The middle layer architecture includes a multi-layer learning model. Each layer of the learning model includes multiple feature extraction modules and multiple graph attention networks with one-to-one correspondence. Each feature extraction module includes a parallel multidimensional feature extractor and a Bert model. The multidimensional feature extractor uses a spectral attention mechanism to extract features of the first channel from the fused features output by the first input layer. The Bert model uses a self-attention mechanism to extract features of the second channel from the text encoding sequence output by the second input layer. The graph attention network is constructed based on the network structure of a distributed photovoltaic graph network. The graph attention network uses a graph attention mechanism to self-learn the hysteresis characteristics of the target photovoltaic field based on the neighboring photovoltaic fields, so as to perform meteorological time-shift correction on the features extracted from the first and second channels respectively and output them. The output layer architecture includes multiple decoding layers. The number of decoding layers is the same as the number of distributed photovoltaic fields. That is, each node of the graph attention network is connected to a corresponding decoding layer. Each decoding layer includes a dual-channel fully connected layer and a single-channel fully connected layer, which is used to perform weighted fusion of the corrected features of the two channels and simultaneously output the power generation prediction results of multiple distributed photovoltaic fields.

2. The distributed photovoltaic power generation prediction method based on a dual-channel dynamic spatiotemporal graph according to claim 1, characterized in that: The process of constructing a variable matrix for each distributed photovoltaic field based on historical data, evaluating the correlation between multiple distributed photovoltaic fields based on the variable matrix, and constructing a distributed photovoltaic map network based on the correlation evaluation results includes the following: Each distributed photovoltaic field is regarded as a photovoltaic node. Based on the light intensity, temperature, dust particle concentration and power generation of each photovoltaic node, an L×4 variable matrix is constructed, where L represents the length of the historical data sequence. Each variable matrix is normalized by row and reshaped into a one-dimensional variable vector. The Pearson correlation coefficient between each photovoltaic node and other photovoltaic nodes is calculated based on the one-dimensional variable vector. If the Pearson correlation coefficient between two photovoltaic nodes is greater than the preset threshold, it is determined that a node edge exists between the two photovoltaic nodes. Otherwise, it is determined that no node edge exists between the two photovoltaic nodes. In this way, a distributed photovoltaic graph network that can be dynamically adjusted in real time is constructed.

3. The distributed photovoltaic power generation prediction method based on a dual-channel dynamic spatiotemporal graph according to claim 1, characterized in that: The data embedding method adopted by the first input layer includes time feature coding, value coding and position coding.

4. The distributed photovoltaic power generation prediction method based on a dual-channel dynamic spatiotemporal graph according to claim 1, characterized in that: The graph attention network calculates the self-learned weighted features based on the following formula, Where h′ i (P) represents weighted features, P represents the number of heads in the multi-head attention mechanism, || represents the vector concatenation operation, and the subscripts i and j represent the target photovoltaic field and the neighboring photovoltaic field, respectively. represents the attention coefficient of the neighboring photovoltaic field j calculated by the p-th head, represents the attention coefficient of target photovoltaic field i calculated by the p-th head, W p represents the learnable parameters of the p-th head, h j represents the feature vector of the neighboring photovoltaic field j extracted by the multidimensional feature extractor or Bert model, h i Represents the feature vector of target photovoltaic field i extracted by multidimensional feature extractor or Bert model, N i represents the set of neighbor nodes of the target photovoltaic field i, and σ() represents the activation function.

5. The distributed photovoltaic power generation prediction method based on dual-channel dynamic spatiotemporal graph according to claim 4, characterized in that: Each head calculates the attention coefficient of the neighboring photovoltaic field based on the following formula: Among them, α ij represents the attention coefficient of the neighboring photovoltaic field j, e ij and e ik They represent the similarity coefficients between the neighboring photovoltaic field j and the neighboring photovoltaic field k and the target photovoltaic field i, respectively, and LeakyReLU() represents the activation function.

6. The distributed photovoltaic power generation prediction method based on a dual-channel dynamic spatiotemporal graph according to claim 5, characterized in that: Each head calculates the similarity coefficient between the neighboring photovoltaic field and the target photovoltaic field based on the following formula: e ij =a([Wh i ||Wh j ]),j∈N i Among them, a() represents the mapping function that maps high-dimensional features to real numbers e, W represents the learnable parameters, and h i and h j They represent the feature vectors of target photovoltaic field i and neighboring photovoltaic field j extracted by the multidimensional feature extractor or Bert model respectively.

7. A distributed photovoltaic power generation prediction system based on a dual-channel dynamic space-time graph, using the distributed photovoltaic power generation prediction method based on a dual-channel dynamic space-time graph according to any one of claims 1 to 6, characterized in that: include: A data acquisition module is used to collect historical data of multiple distributed photovoltaic fields, where the historical data types include light intensity, temperature, dust particle concentration, time and power generation; A graph network construction module is used to construct a variable matrix for each distributed photovoltaic field based on historical data, evaluate the correlation between multiple distributed photovoltaic fields based on the variable matrix, and construct a distributed photovoltaic graph network based on the correlation evaluation results. The data dimensions of the variable matrix include light intensity, temperature, dust particle concentration, and power generation; A model training module is used to build a dual-channel prediction model based on a distributed photovoltaic graph network, train the dual-channel prediction model using historical data from multiple distributed photovoltaic fields, and save the model parameters after training. The dual-channel prediction model includes a first feature extraction channel using a spectral attention mechanism and a graph attention mechanism, and a second feature extraction channel using a self-attention mechanism and a graph attention mechanism. The power generation prediction module is used to simultaneously predict the power generation of multiple distributed photovoltaic fields using the trained dual-channel prediction model.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the steps of the method according to any one of claims 1 to 6 by calling the computer program stored in the memory.

9. A computer-readable storage medium for storing a computer program for predicting distributed photovoltaic power generation based on a dual-channel dynamic spatiotemporal graph, characterized in that: When the computer program is run on a computer, the steps of the method according to any one of claims 1 to 6 are executed.

Images