Photovoltaic power generation capability prediction method and system based on space-time diagram neural network, and medium

Abstract

The invention discloses a photovoltaic power generation capability prediction method and system based on a space-time diagram neural network, and a medium. The method comprises the following steps: preprocessing historical data of a power system; a plurality of power stations are divided into a plurality of clusters, a power system is abstracted into a topological graph, and each cluster is regarded as a node of the topological graph; learning time features of the topological graph by adopting a time self-attention network model; learning spatial features of the topological graph by adopting a spatial graph convolutional network model; stacking the time self-attention network and the space graph convolutional network by adopting a factor type structure; and performing prediction by using a full connection layer. According to the method, the photovoltaic power generation system is modeled into the graph model, so that the spatial relationship and the time dependence among different power stations can be effectively captured, more efficient data processing is realized, redundant information is reduced, data streams can be optimized, and the requirements on storage resources are reduced by reducing direct operation on original data, so that the system is more efficient and more reliable. And the processing speed of the hardware is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of renewable energy power generation, and particularly to a method, a system and a medium for predicting photovoltaic power generation capacity. Background Art

[0002] In recent years, in response to the call of building a clean, low-carbon, safe and efficient energy system in China, renewable energy power generation stations in China (such as wind power generation and photovoltaic power generation) have developed rapidly. However, these renewable energy power generations have strong intermittency, randomness and volatility, and different spatio-temporal resource endowments, load levels and power grid facilities have a huge impact on the operation of the existing power system. Therefore, the prediction of photovoltaic power generation capacity has become a key technology to solve this problem. The prediction of photovoltaic power generation is helpful to improve the operation efficiency, supply-demand balance and stability of the power system, can provide better energy scheduling suggestions for the power dispatching department, and will also bring huge economic benefits and environmental protection benefits.

[0003] According to research, the research on the prediction of photovoltaic power generation capacity at home and abroad is mainly divided into the following categories:

[0004] (1) Physical method. As Figure 1 shown, this method uses methods such as solar geometry and radiation calculation formulas, and photovoltaic system performance modeling for prediction. First, the calculation formula is obtained by combining geographical location information, photovoltaic module parameters and weather models, and then the parameters of the photovoltaic system to be predicted are input into the formula to obtain the prediction result. The modeling process of this method is cumbersome, the calculation formula has poor robustness, and abnormal conditions such as photovoltaic module aging and extreme weather changes are not considered.

[0005] (2) Statistical method. This method constructs a regression model (or other traditional machine learning models, such as support vector regression) based on input and output for prediction. It relies on a large amount of historical data to reduce the error of the model, and does not deeply mine the potential features in the data.

[0006] (3) Neural network. In a real dataset, the non-linearity of solar irradiance, the volatility of the photovoltaic system, and the complexity of the weather system make traditional machine learning methods appear powerless in the face of complex prediction tasks. In recent years, the deep neural network DNN and the long short-term memory network LSTM have achieved remarkable results in predicting solar irradiance and photovoltaic power. It learns time features from time series data for prediction, but has poor interpretability.

[0007] Among the above three prediction methods, the physical method has a cumbersome modeling process and poor model robustness. The statistical method relies on a large amount of historical data, and the prediction effect is not ideal. The model interpretability of the neural network method is relatively poor; at the same time, ordinary neural networks are more proficient in time series data prediction and cannot achieve the expected effect when calculating model parameters based on high-dimensional input vectors, and it is necessary to consider integrating feature information other than time series features. Since these methods all require the input of a large amount of historical power data, their training speed is relatively slow, the memory overhead is large, and they do not have the real-time performance of prediction, so it is necessary to consider accelerating the model training and prediction speed. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to propose a photovoltaic power generation capacity prediction method, system and medium based on a spatio-temporal graph neural network to overcome the deficiencies of the prior art.

[0009] To solve the above technical problem, the present invention proposes a photovoltaic power generation capacity prediction method based on a spatio-temporal graph neural network, including the following steps:

[0010] S1: Clean and standardize the historical data of the power system to obtain standardized data, where the standardized data includes the location information of each grid station, weather and climate data, and equipment parameters;

[0011] S2: Based on the standardized data, classify power generation stations according to the similarity of location information, weather and climate data, and equipment parameters through a clustering algorithm, and construct a topological graph representing the structure of the power system, where each clustering node corresponds to a cluster of power generation stations with similar characteristics;

[0012] S3: Input the topological graph into a time self-attention network model, and generate a time feature matrix representing the correlation between historical power generation capacity and future prediction by extracting the time series features of each node in the topological graph;

[0013] S4: Based on the generator mechanism and prior knowledge, fuse weather and climate data through a spatial graph convolutional network model, analyze the spatial correlation between nodes in the topological graph, and generate a spatial feature matrix reflecting the mutual influence of power generation stations;

[0014] S5: Perform factorized cross-stacking on the time feature matrix and the spatial feature matrix, and generate integrated spatio-temporal features through iterative update of spatio-temporal dependence;

[0015] S6: Input the integrated spatio-temporal features into a fully connected layer, and output the predicted value of photovoltaic power generation capacity based on the physical mechanism mapping relationship between equipment parameters and photovoltaic power generation capacity.

[0016] Further, the specific steps of S1 include:

[0017] S1-1: Perform consistency alignment processing on the multi-source heterogeneous fields of the historical data of the power system to generate a standardized field set containing a unified naming specification and dimension system, eliminate field redundancy conflicts to construct a reliable data stream;

[0018] S1-2: Based on the standardized field set, fill in the missing data with the mean values of the previous and subsequent moments within a sliding time window to generate a complete time series data set with continuity;

[0019] S1-3: According to the difference in data acquisition frequencies between the meteorological station and the power sensor, perform time-frequency synchronization fusion on the complete time series data set to generate a high-resolution data stream with a unified time baseline;

[0020] S1-4: Based on the high-resolution data stream, establish a dynamic weight relationship between each field and the photovoltaic power generation capacity through a feature importance analysis algorithm, and output a feature impact vector with physical interpretation to guide model training.

[0021] Further, the step S2 specifically includes:

[0022] S2-1: Perform Z-score normalization processing on the multi-dimensional features in the standardized data to generate a normalized feature matrix, and eliminate the dimensional differences of position information, weather and climate data, and equipment parameters;

[0023] S2-2: Based on the normalized feature matrix, use the K-means++ algorithm to initialize the clustering centers, and generate a set of differential clustering centers by maximizing the distance between the initial center points to ensure the feature separation degree between categories;

[0024] S2-3: Construct a hybrid loss function that combines spatial distance and equipment parameters, perform iterative optimization on the set of differential clustering centers, and output an optimized clustering center and corresponding category labels that reflect the common characteristics of the power station;

[0025] S2-4: Based on the optimized clustering center, dynamically calculate the difference entropy value between the within-class dispersion and the random reference distribution through the Gap Statistic method, and determine the optimal number of clusters K that maximizes the difference entropy to achieve adaptive category division;

[0026] S2-5: Map the category labels corresponding to the optimal number of clusters K to the nodes of the topological graph to generate a simplified topological graph representing the regional power generation association, where each node represents a feature cluster with a similar power generation pattern, and reduce the computational complexity of the subsequent spatio-temporal model by reducing the original site order of magnitude (N→K).

[0027] Further, when extracting the time series features of each node in the topological graph in the step S3, the Transformer model is used to capture the time relationship between different time steps.

[0028] Further, in step S4, when analyzing the spatial correlation between nodes in the topological graph and generating a spatial feature matrix reflecting the mutual influence of power generation stations, the topological graph is converted into an adjacency matrix form; at the same time, the degree of each node is calculated to obtain a degree matrix, and the mutual relationship between nodes is established.

[0029] Further, in step S5, when factorially cross-stacking the time feature matrix and the spatial feature matrix, the graph data is processed recursively to capture spatio-temporal dependence relationships.

[0030] A photovoltaic power generation capacity prediction system based on a spatio-temporal graph neural network, comprising: a computer-readable storage medium and a processor;

[0031] The computer-readable storage medium is used to store executable instructions;

[0032] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the photovoltaic power generation capacity prediction method based on the spatio-temporal graph neural network.

[0033] A non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the photovoltaic power generation capacity prediction method based on the spatio-temporal graph neural network is implemented.

[0034] The present invention has the following beneficial effects:

[0035] (1) High prediction accuracy: By modeling the photovoltaic power generation system as a graph model, the present invention fully considers spatial features and time features, and realizes high-precision prediction of photovoltaic power generation. This method overcomes the limitations of traditional physical and statistical methods and shows significant superiority in terms of prediction accuracy and stability. Through experimental verification based on the historical data of multiple photovoltaic power stations in a certain area and evaluation using indicators such as mean square error (MSE) and root mean square error (RMSE), the results show that the prediction error of this method is reduced by 20% compared with traditional methods (such as long short-term memory network).

[0036] (2) Strong adaptability: The present invention models multiple photovoltaic power generation stations as a topological graph and uses subsequent algorithms to learn features and make predictions. This modeling method enables the method to adapt to different regions and types of photovoltaic power generation systems, demonstrating broad applicability. In the experiment, we selected the data of photovoltaic power stations in three different regions for testing, and the results showed that the method could maintain stable prediction performance in different environments.

[0037] (3) Efficient resource utilization: By optimizing the processing of input data, the present invention reduces the demand for storage resources and computing resources, thereby improving the training speed and hardware processing efficiency. In the same hardware environment, the training time when the method achieves the optimal fitting effect is shortened by 23% compared with the traditional training method. This feature enables the model to quickly respond to changes and adapt to real-time prediction requirements. In actual prediction applications, the method can simultaneously respond to the prediction indicators of multiple sites. As the number of target sites increases, the response speed is increased by 12%-25% compared with the traditional method. Description of the Drawings

[0038] Figure 1 It is a schematic diagram of the module structure of the physical method adopted for existing prediction.

[0039] Figure 2 It is a network architecture diagram corresponding to a photovoltaic power generation capacity prediction method based on a spatio-temporal graph neural network of the present invention.

[0040] Figure 3 It is a topological graph generated in a photovoltaic power generation capacity prediction method based on a spatio-temporal graph neural network of the present invention.

[0041] Figure 4 It is a stacked factor structure diagram of a stacked time self-attention network and a spatial graph convolutional network in a photovoltaic power generation capacity prediction method based on a spatio-temporal graph neural network of the present invention. Detailed Embodiments

[0042] In practical applications, as a distributed complex system, the power generation behaviors and patterns of different types of power generation equipment in different regions of a photovoltaic power generation system are significantly different. Therefore, when performing photovoltaic power generation prediction, the spatial characteristics of the power generation systems in each region must be considered. By extracting the spatial characteristics of the data of multiple photovoltaic power generation equipment / systems in the same region, the present invention not only effectively reduces the complexity of data processing, but also enhances the adaptability of the model to different power generation behaviors. At the same time, when extracting spatial characteristics, the present invention pays attention to the photovoltaic power generation mechanism and prior knowledge, and has strong interpretability.

[0043] An embodiment of the present invention proposes a photovoltaic power generation capacity prediction method based on a spatio-temporal graph neural network, including the following steps:

[0044] S1: Clean and standardize the historical data of the power system to obtain standardized data, where the standardized data includes the location information of each site in the power grid, weather and climate data, and equipment parameters;

[0045] S2: Based on the standardized data, classify power stations according to the similarity of location information, weather and climate data, and equipment parameters through a clustering algorithm, and construct a topological graph representing the structure of the power system, where each clustering node corresponds to a cluster of power stations with similar characteristics;

[0046] S3: Input the topological graph into a temporal self-attention network model, and generate a temporal feature matrix representing the correlation between historical power generation capacity and future prediction by extracting the time series features of each node in the topological graph;

[0047] S4: Based on the generator principle and prior knowledge, fuse weather and climate data through a spatial graph convolutional network model, analyze the spatial correlation between nodes in the topological graph, and generate a spatial feature matrix reflecting the mutual influence of power stations;

[0048] S5: Perform factorized cross-stacking on the temporal feature matrix and the spatial feature matrix, and generate integrated spatio-temporal features through iterative updating of spatio-temporal dependence relationships;

[0049] S6: Input the integrated spatio-temporal features into a fully connected layer, and output the predicted value of photovoltaic power generation capacity based on the physical mechanism mapping relationship between equipment parameters and photovoltaic power generation capacity. The network architecture diagram corresponding to the prediction method of the present invention is as Figure 2 shown, where ReLU activation functions are used to connect every two layers.

[0050] Step S1 specifically includes the following steps:

[0051] S1-1: Data cleaning, perform consistency alignment processing on multi-source heterogeneous fields of historical power system data, generate a standardized field set containing a unified naming specification and dimension system, eliminate field redundancy conflicts to construct a reliable data stream. Photovoltaic power generation data comes from different manufacturers and devices, with different formats and acquisition frequencies, making data cleaning and standardization complex. During the cleaning process, first solve the problems of data inconsistency and missing values. For the historical power grid data, including the location information, weather and climate parameters, and equipment status of each power station, unified standardization processing is required. For fields included in some data sets but not in others, the default method is adopted, and the values are assigned as the mode, median, or average according to the specific meaning of the fields; if a field appears in a small number of data sets and cannot provide effective information, consider discarding the field. For observations with a large proportion of missing values, if their impact on the overall analysis is small, they can also be discarded. In addition, for outliers of a certain field (such as extreme values of weather parameters), they can be changed to the mode, median, or average according to the business meaning of the field.

[0052] S1-2: Data filling. Based on the standardized field set, the missing data is filled with the average values of the adjacent moments within the sliding time window to generate a complete time series data set with continuity. For the data of the same power station, since it is continuous time series data, it should include parameters such as the weather condition, solar irradiance, and power generation at each fixed time interval t. If data is missing at a certain moment due to collection, observation, or other errors, the average value of the adjacent moments before and after this moment is used for filling to ensure the continuity of the time series.

[0053] S1-3: Data alignment. According to the difference in data collection frequencies between the weather station and the power sensor, the complete time series data set is subjected to time-frequency synchronization fusion to generate a high-resolution data stream with a unified time baseline. Usually, the data collection frequency of the power station is between 5 minutes and 60 minutes, and the collection frequencies of different data sets may vary. For data sets with similar frequencies, they can be directly merged; for data sets with large frequency differences, it is recommended to perform model training separately in the subsequent steps and then perform feature fusion to ensure the effectiveness of each feature in the model.

[0054] S1-4: Model the fields in the data set. Based on the high-resolution data stream, establish the dynamic weight relationship between each field and the photovoltaic power generation capacity through the feature importance analysis algorithm, and output the feature influence vector with physical explanations to guide model training. Suppose there are N photovoltaic power stations in a certain area, and the fields that may be included in the data set include but are not limited to: the location of the power station, weather and climate parameters (such as temperature, humidity, wind speed, etc.), solar irradiance, power generation, equipment parameters (such as inverter status, component type, etc.). Systematically model these fields to ensure that the model can accurately reflect the influence of each feature on the prediction of photovoltaic power generation capacity. The modeling of the fields that may be included in the data set is shown in Table 1:

[0055] Table 1

[0056]

[0057]

[0058] Step S2 specifically includes the following steps:

[0059] S2-1: Data normalization. Perform Z-score normalization on the multi-dimensional features in the standardized data to generate a normalized feature matrix, eliminating the dimensional differences in location information, weather and climate data, and equipment parameters. During the clustering process, considering that the K-means algorithm divides data based on Euclidean distance, dimensions with larger means and variances may have a greater impact on the clustering results. Therefore, before clustering, it is necessary to normalize and unify the units of the features of each dimension. In addition, to improve the clustering effect, outliers need to be removed in step S1 to ensure that the clustering is based on valid data.

[0060] S2-2: Initialize cluster centers. Based on the normalized feature matrix, use the K-means++ algorithm to initialize the clustering centers, generating a set of differential clustering centers by maximizing the distance between the initial center points to ensure the feature separation degree between categories. Use the idea of the K-means++ algorithm to initialize the cluster centers - select K cluster centers one by one, and the sample points farther away from other cluster centers are more likely to be selected as the next cluster center. First, randomly select a sample point from the data set χ as the first initial clustering center c 1 ; then, calculate the shortest distance between each sample and the currently existing clustering centers, denoted as D(x), and calculate the probability of each sample point being selected as the next clustering center Finally, select the sample point corresponding to the maximum probability value as the next cluster center, where χ represents the set of sample points; repeat the above step until K cluster centers are selected.

[0061] In clustering, the similarity between samples is calculated comprehensively using the similarity of the historical power generation data of photovoltaic power stations and the Euclidean distance based on location information, weather and climate, and equipment parameters. The specific calculation method is as follows:

[0062] (1) Similarity of historical power generation data: Use the dynamic time warping (DTW) algorithm to calculate the similarity of historical power generation data between power stations, which can effectively capture changes in power generation patterns.

[0063] (2) Location information: Use longitude and latitude coordinates to convert location information into plane coordinates to calculate the straight-line distance between power stations.

[0064] (3) Weather and climate: Consider features such as temperature, humidity, and wind speed, and use the standardized data for distance calculation.

[0065] (4) Equipment parameters: Include the efficiency, inclination angle, and component aging degree of photovoltaic panels, and calculate the parameter similarity using physical formulas.

[0066] Combining the above parameters, the similarity D(x i , x j ) calculation formula is as follows:

[0067]

[0068] Among them, lat and long respectively represent the latitude and longitude of power stations i and j, and f k is the standardized weather climate and equipment parameter characteristics, and DTW(h i , h j ) is the dynamic time warping distance between the historical power generation data of power stations i and j.

[0069] S2-3: Define the loss function and perform clustering operations. Construct a hybrid loss function that combines spatial distance and equipment parameters, and iteratively optimize the set of differential clustering centers to output the optimized clustering centers and corresponding class labels that reflect the common characteristics of power stations. The distance metric is defined as the similarity D(x i , x j ) between two photovoltaic power stations; the loss function is defined as where N represents the total number of power stations, x i represents the i-th sample, c i represents the cluster to which x i belongs, represents the center point of cluster c i . First, fix the center point and adjust the class to which each sample belongs to reduce the value of the loss function; then fix the class of each sample and adjust the center point to continue reducing the value of the loss function until the loss is reduced to a minimum value and the clustering division converges.

[0070] S2-4: Use the Gap Statistic to determine the value of K. Based on the optimized clustering centers, dynamically calculate the difference entropy value between the within-class dispersion and the random reference distribution through the Gap Statistic method, and determine the optimal number of clusters K that maximizes the difference entropy to achieve adaptive class division. The selection of the value of K is generally based on the results of multiple experiments. The Gap Statistic method can automatically find the best value of K for the clustering effect and avoid overfitting or underfitting.

[0071] S2-5: Map the class labels corresponding to the optimal number of clusters K to the nodes of the topological graph to generate a simplified topological graph representing the regional power generation association, where each node represents a feature cluster with a similar power generation pattern, and reduce the computational complexity of the subsequent spatio-temporal model by reducing the original number of sites (N → K). After the above steps, finally, a large number of photovoltaic power stations are divided into reasonable K clusters. According to the obtained multiple clusters, the complex power system is abstracted into a topological graph that is convenient for analysis and processing. Regarding each cluster as a node, a topological graph similar to Figure 3 can be obtained.

[0072] After completing steps S1 - S2, the obtained dataset contains two types of data: time - series data and spatial topological data. First, each photovoltaic power station has its own historical power generation data, which have obvious time - series characteristics. As time goes by, they show periodicity and trend, so they are called time - series data. Second, after clustering and partitioning, the spatial relationship between power stations is modeled as a graph topological structure, which dynamically adjusts over time, and is called spatial topological data. There is a spatio - temporal dependence relationship between these two types of data, and in the subsequent steps S3 - S5, these two types of data will be fused for modeling. In this context, these two types of data are collectively referred to as graph data.

[0073] In step S3, a time self - attention network model is used to learn the time characteristics of each power station. Here, the Transformer model is used to capture the time relationship between different time steps. The core of the Transformer model is the dot - product attention mechanism, and its calculation formula is where Q represents the query matrix query, K represents the key matrix key, K T is the transpose of K, represents the dimensional vector of Q, Softmax is the normalization function, V represents the value matrix value; before using the Transformer model, it is necessary to perform positional encoding on the time - series data first. The encoding formula used here is as follows:

[0074]

[0075] where pos is the position in the sequence, d model is the length of the feature vector of the positional information encoding, i represents the i - th element of the positional information encoding feature vector, and the odd - numbered bits in the encoding vector are encoded with the cosine function cos, and the even - numbered bits are encoded with the sine function sin.

[0076] In step S4, a spatial graph convolutional network model is used to learn the spatial characteristics of each power station. Compared with the spectrum - based graph convolutional network, the spatial graph convolutional network overcomes the dependence on the Laplacian matrix, and its definition is where A represents the adjacency matrix, D is the degree matrix, ω is the parameter to be learned, represents the Laplacian matrix L, I N represents the identity matrix.

[0077] When using a spatial graph convolutional network to learn the spatial features of a photovoltaic power generation system, it is necessary to convert the topological graph obtained in S2-5 into the form of an adjacency matrix. At the same time, calculate the degree of each node to obtain the degree matrix. When the topological graph is complex, the dimension of the adjacency matrix is high, making it difficult to input it into the spatial graph convolutional network for convolution operations. Therefore, a sampling aggregation method is adopted here to improve the network's processing ability for high-dimensional adjacency matrices, and its aggregation operator is expressed as follows:

[0078]

[0079] In the formula, represents the set of neighbor nodes of node u, represents the sampling aggregation operation and the embedding of node u, and W k represents the weight matrix of the k-th order embedding.

[0080] By performing graph convolution and fusing multiple spatial correlations, the algorithm complexity can be further reduced and the prediction efficiency can be improved.

[0081] Step S5, stack the time feature matrix and the spatial feature matrix using a factorized structure, and the specific stacking method is as Figure 4 shown.

[0082] This stacking method captures the dependencies of spatio-temporal relationships in a recursive form. For each time step in the graph data, the graph topology and time series data at this step are sequentially processed by the spatial learning network and the time learning network to learn spatial and time dependencies respectively. Each stacked factor in this model can be expressed as:

[0083] f(X,A) = σ(AXW 0 )

[0084] u t = σ(W u [f(A,X t ),h t-1 +b u )

[0085] r t = σ(W r [f(A,X t ),h t-1 +b r )

[0086] c t = tanh(W c [f(A,X t ),(r t *h t-1 )]+b c )

[0087] h t= u t * h t-1 + (1 - u t ) * c t

[0088] In the formula, u t , r t , c t , h t respectively represent the outputs of each factor, W 0 , W u , W r , W c respectively represent the parameter matrices of each factor, b u , b r , b c respectively represent the bias matrices of each factor, f(A, X t ) represents the output of the time learning network at time step t.

[0089] In step S6, a fully connected layer is added after the spatio-temporal graph learning module to map the learned feature representation to the label space of the sample. Input the relevant parameters (geographical location, solar irradiance, weather data, equipment configuration, etc.) of the photovoltaic power generation system to be predicted, perform data preprocessing, clustering division and modeling into a topological graph according to steps S1 and S2, and after passing through the fully connected layer, map to obtain the results such as the predicted power generation power that characterize the photovoltaic power generation ability, and at the same time, the weights and changes of various physical factors input can be reflected in the change of the predicted power.

[0090] By modeling the photovoltaic power generation system as a graph model, the present invention can effectively capture the spatial relationship and time dependence between different power stations, thereby realizing more efficient data processing. This modeling method not only reduces redundant information, but also optimizes the data stream. By reducing the direct operation on the original data, the demand for storage resources is reduced, and thus the processing speed of the hardware is improved.

[0091] In addition, there is an inherent correlation between the prediction factors in the present invention and the historical data of the power system, which conforms to the photovoltaic power generation law. Therefore, it has wide applicability and can adapt to different regions and different types of photovoltaic power generation systems.

[0092] Another embodiment of the present invention provides a photovoltaic power generation ability prediction system based on a spatio-temporal graph neural network, including: a computer-readable storage medium and a processor;

[0093] The computer-readable storage medium is used to store executable instructions;

[0094] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the photovoltaic power generation ability prediction method based on the spatio-temporal graph neural network.

[0095] Another embodiment of the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the photovoltaic power generation capacity prediction method based on the spatio-temporal graph neural network described above is implemented.

[0096] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take 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.) containing computer-usable program code.

[0097] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a machine for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 a device for the functions specified in one block or multiple blocks.

[0098] These computer program instructions can 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 generate a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 a device for the functions specified in one block or multiple blocks.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 a device for the functions specified in one block or multiple blocks.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent substitutions can still be made to the specific embodiments of the present invention, and any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A photovoltaic power generation capacity prediction method based on spatiotemporal graph neural network, characterized in that: The following steps are involved: S1: Clean and standardize the historical data of the power system to obtain standardized data, which includes location information, weather and climate data, and equipment parameters of each station in the power grid; S2: Based on the standardized data, the power stations are classified according to the location information, weather and climate data, and equipment parameter similarity through a clustering algorithm, and a topological diagram representing the structure of the power system is constructed, wherein each clustering node corresponds to a cluster of power stations with similar characteristics; S3: inputting the topological graph into a temporal self-attention network model, and generating a temporal feature matrix representing the correlation between historical power generation capacity and future prediction by extracting the time series features of each node in the topological graph; S4: Based on the power generation mechanism and prior knowledge, weather and climate data are integrated through the spatial graph convolutional network model to analyze the spatial correlation between nodes in the topological graph and generate a spatial feature matrix reflecting the mutual influence of power stations; S5: cross-stacking the time feature matrix and the space feature matrix in a factorial manner, iteratively updating through the time-space dependency relationship, and generating an integrated time-space feature; S6: Input the integrated spatiotemporal features into a fully connected layer, and output a predicted value of photovoltaic power generation capacity based on a mapping relationship between device parameters and the physical mechanism of photovoltaic power generation capacity.

2. The photovoltaic power generation capacity prediction method based on spatiotemporal graph neural network according to claim 1 is characterized in that: The step S1 specifically includes: S1-1: Perform consistency alignment on the multi-source heterogeneous fields of the power system historical data, generate a standardized field set with a unified naming specification and dimension system, eliminate field redundancy conflicts and build a reliable data flow; S1-2: Based on the standardized field set, the missing data is filled with the mean of the previous and next moments in the sliding time window to generate a complete time series data set with continuity; S1-3: According to the difference in data collection frequency between the meteorological station and the power sensor, the complete time series data set is synchronized and fused in time and frequency to generate a high-resolution data stream with a unified time baseline; S1-4: Based on the high-resolution data stream, a dynamic weight relationship between each field and photovoltaic power generation capacity is established through a feature importance analysis algorithm, and a feature influence vector with a mechanism explanation is output to guide model training.

3. The photovoltaic power generation capacity prediction method based on spatiotemporal graph neural network according to claim 1 is characterized in that: The step S2 specifically includes: S2-1: performing Z-score normalization processing on the multidimensional features in the standardized data to generate a normalized feature matrix, eliminating the dimensional differences of the location information, weather and climate data, and equipment parameters; S2-2: Based on the normalized feature matrix, the K-means++ algorithm is used to initialize the cluster centers, and a differential cluster center set is generated by maximizing the initial center point spacing to ensure feature separation between categories; S2-3: construct a hybrid loss function integrating spatial distance and equipment parameters, iteratively optimize the difference cluster center set, and output the optimized cluster center and corresponding category label reflecting the common characteristics of the power station; S2-4: Based on the optimized cluster center, dynamically calculate the difference entropy value between the intra-class dispersion and the random reference distribution through the Gap Statistic method, determine the optimal cluster number K that maximizes the difference entropy, and realize adaptive category division; S2-5: Map the category labels corresponding to the optimal clustering number K to topological map nodes, generate a simplified topological map representing regional power generation associations, where each node represents a characteristic cluster with similar power generation patterns, and reduce the computational complexity of subsequent spatiotemporal models by reducing the order of magnitude of original sites (N→K).

4. The photovoltaic power generation capacity prediction method based on spatiotemporal graph neural network according to claim 1 is characterized in that: When extracting the time series features of each node in the topological graph in step S3, a Transformer model is used to capture the time relationship between different time steps.

5. The photovoltaic power generation capacity prediction method based on spatiotemporal graph neural network according to claim 1 is characterized in that: In step S4, the spatial correlation between nodes in the topological map is analyzed, and when a spatial feature matrix reflecting the mutual influence of power stations is generated, the topological map is converted into an adjacency matrix form; at the same time, the degree of each node is calculated to obtain a degree matrix, and the relationship between nodes is established.

6. The photovoltaic power generation capacity prediction method based on spatiotemporal graph neural network according to claim 1 is characterized in that: In step S5, when the time feature matrix and the space feature matrix are cross-stacked in a factored manner, the graph data is processed recursively to capture the spatiotemporal dependency.

7. A photovoltaic power generation capacity prediction system based on spatiotemporal graph neural network, comprising: A computer readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the photovoltaic power generation capacity prediction method based on the spatiotemporal graph neural network described in any one of claims 1 to 6.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the photovoltaic power generation capacity prediction method based on a spatiotemporal graph neural network as described in any one of claims 1 to 6.

Images