A short-term prediction method for distributed photovoltaic power generation based on big data

Abstract

The present invention belongs to the technical field of photovoltaic power generation, and specifically relates to a short-term prediction method for distributed photovoltaic power generation based on big data; the method comprises: constructing an undirected weighted heterogeneous graph to obtain original node features of all nodes; extracting connectivity features and importance features from the heterogeneous graph; using a power time series graph convolutional neural network to process the importance features and time series features to obtain a first updated node feature of the node; using a graph attention mechanism to process the importance features and the first updated node features to obtain a second updated node feature; using a multi-head path graph attention mechanism to process the connectivity features and the first updated node features to obtain a third updated node feature; fusing three types of updated node features to obtain a comprehensive feature and inputting it into a multi-layer perceptron to obtain the predicted power generation of a photovoltaic power station; the present invention can improve the prediction efficiency of the power generation and help maintain the stability of the power grid.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, and in particular relates to a short-term prediction method for distributed photovoltaic power generation based on big data. Background Art

[0002] Under the current background of global energy structure adjustment, the development and utilization of renewable energy has become one of the important development directions. Photovoltaic power generation, as an important component of renewable energy, has the advantages of extensive resources, environmental protection and cleanliness, and plays an important role in the energy field. Among them, distributed photovoltaic power stations are widely deployed on building roofs, parking lots, agricultural facilities and other places. They can not only meet the electricity needs of users nearby and reduce the transmission pressure of the power grid, but also effectively improve energy utilization efficiency and reduce carbon emissions. However, due to the influence of environmental factors such as weather and climate, the power generation of photovoltaic power generation systems shows great volatility and uncertainty, which brings challenges to the prediction of photovoltaic power generation.

[0003] Traditional distributed photovoltaic power generation prediction methods are mostly based on a single statistical model or physical model, such as time series analysis, support vector machines, artificial neural networks, etc. Although these methods can capture the changing trend of photovoltaic power generation to a certain extent, they are often unable to deeply explore the complex relationship between power generation, environment, and time during photovoltaic power generation. In addition, traditional methods are prone to problems such as overfitting and insufficient generalization when dealing with large-scale, high-dimensional, and heterogeneous data. They are unable to fully explore the deep information contained in massive data, resulting in limited prediction accuracy. In recent years, graph neural networks have been widely used in various fields due to their advantages in dealing with unstructured data problems, but there are currently few solutions that apply graph structures to the field of photovoltaic power generation prediction. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes a distributed photovoltaic power generation short-term prediction method based on big data, the method comprising: obtaining photovoltaic power station related data and inputting it into a trained distributed photovoltaic power generation short-term prediction model to obtain the photovoltaic power station predicted power generation;

[0005] The training process of the distributed photovoltaic power generation short-term prediction model includes:

[0006] S1: Obtain relevant data of photovoltaic power plants, construct an undirected weighted heterogeneous graph based on the relevant data of photovoltaic power plants, and obtain the original node features of all nodes in the undirected weighted heterogeneous graph;

[0007] S2: Extracting connectivity features and importance features from undirected weighted heterogeneous graphs;

[0008] S3: According to the undirected weighted heterogeneous graph, the power time series graph convolutional neural network is used to process the time series features in the importance features and the original node features to obtain the first updated node features of the node;

[0009] S4: According to the undirected weighted heterogeneous graph, the graph attention mechanism is used to process the importance features and the first updated node features to obtain the second updated node features of the node;

[0010] S5: According to the undirected weighted heterogeneous graph, the multi-head path graph attention mechanism is used to process the connectivity features and the first updated node features to obtain the third updated node features of the node;

[0011] S6: Fusing the second updated feature and the third updated feature of the node itself to obtain the final feature of the node; splicing the final features of the three nodes to obtain the comprehensive features of the power station;

[0012] S7: Input the comprehensive features into a multi-layer perceptron for processing to obtain the predicted power generation of the photovoltaic power station;

[0013] S8: Calculate the total model loss and adjust the model parameters according to the total model loss to obtain a trained distributed photovoltaic power generation short-term prediction model.

[0014] Preferably, the process of constructing an undirected weighted heterogeneous graph includes: taking each power station as a power station node, each power station node is connected to each corresponding time node to form an edge, each time node is connected to the corresponding environment node to form an edge, and the weight of the edge is 1, so as to obtain multiple subgraphs; calculating the average distance between all power stations, if the distance between the power station nodes in two subgraphs is less than the average distance, there is an edge between the two power station nodes, and the weight of the edge is the distance between the two nodes, otherwise there is no edge, so as to obtain an undirected weighted heterogeneous graph.

[0015] Preferably, the process of obtaining the original node features of all nodes includes:

[0016] The original features of the power station node include the geographical location features, equipment parameter features, maintenance status features and power station layout features of the power station; the original features of the time node include timestamp features, astronomical features, seasonal identification, holiday identification, power demand features and historical performance features; the original features of the environment node include meteorological observation data, extreme weather identification, meteorological forecast information, lighting condition features, and air quality features;

[0017] The original features corresponding to each node are encoded and concatenated to obtain the original node features of all nodes.

[0018] Preferably, step S2 comprises:

[0019] S21: Construct a connection matrix C as a connectivity feature. The elements C in the matrix ij It represents the length of the shortest path between node i and node j in an undirected weighted heterogeneous graph. All elements on the diagonal of matrix C are 0.

[0020] S22: Construct an importance score matrix as an importance feature. Only the elements on the diagonal of the matrix have values. ii Represents the importance score of node i in the undirected weighted heterogeneous graph.

[0021] Furthermore, the calculation formula of the importance score is:

[0022]

[0023] Among them, I ii represents the importance score of node i, N represents the number of nodes in the graph, and a ij Represents the element in the i-th row and j-th column of the adjacency matrix A of the graph. If there is an edge between node i and node j, then a ij =a ji =1, d(i,j) represents the shortest path length from node i to node j, σ st represents the number of shortest paths between node s and node t, σ st (i) represents the number of nodes i passed through in the shortest path.

[0024] Preferably, step S3 specifically includes:

[0025] The undirected weighted heterogeneous graph includes power station nodes, time nodes and environment nodes. The power time series graph convolutional neural network is used to process the importance features and the time series features in the original node features to obtain new time series features. Among them, the time series features are the original node features of the time nodes and environment nodes.

[0026] Each time node or environment node is spliced ​​with the original node feature and the new time series feature corresponding to the node to obtain the first updated node feature of the corresponding node; the original node feature of the power station node is used as the first updated node feature of the power station node.

[0027] Furthermore, the formula for obtaining the new time series features is:

[0028]

[0029] in, represents the output feature vector of the current layer at time t, I ii represents the importance score of node i, K represents the width of the convolution kernel, represents the weight vector of the l-th layer convolution kernel at the delayed k position, d (l)represents the expansion rate of the convolution kernel of the lth layer, Indicates that node i is in td (l) The time series characteristics at time k, b (l) represents the bias term of the lth layer, Represents the output feature vector of the previous layer at time t, and ReLU() represents the ReLU activation function.

[0030] Preferably, in step S4, the formula for processing the importance feature and the first update node feature by the graph attention mechanism is:

[0031]

[0032] Among them, h′ i represents the second updated node feature of node i, α represents the weighting coefficient, and h i represents the first updated node feature of node i, h j represents the first updated node feature of node j, h k represents the first updated node feature of node k, d represents the feature dimension in the graph attention mechanism calculation; W q , W k , W v , W u represents the first, second, third, and fourth learnable weight coefficients, b q , b k , b v , b u represents the first, second, third, and fourth bias terms; N(i) represents the neighboring nodes of node i; LeakyRelu represents the LeakyRelu activation function, softmax represents the normalization function, ReLU represents the ReLU activation function, and I jj represents the importance score of node j, I kk represents the importance score of node k, w ij represents the weight of the edge between node i and node j.

[0033] Preferably, in step S5, the process of processing the connectivity feature and the first updated node feature using the multi-head path graph attention mechanism includes:

[0034] For power station nodes with loops, a loop is randomly selected as the input path, and the connectivity features and the first updated node features of the power station nodes on the path are processed using the multi-head path graph attention mechanism to obtain the third updated node features of the power station nodes;

[0035] The calculation formula of the multi-head path graph attention mechanism is:

[0036]

[0037] Among them, h″ i represents the third updated node feature of node i, concat represents the concatenation operation, P represents the input path, softmax represents the normalization function, and W a represents the learnable weight matrix, represents the feature matrix after linear transformation of the first updated node feature of node i in path P under the first attention head, represents the feature matrix after linear transformation of the first updated node feature of node j in path P under the first attention head, represents the feature matrix after linear transformation of the first updated node feature of node j in path P, C ij represents the element in the i-th row and j-th column of the connection matrix, d h represents the feature dimension in the calculation of the multi-head path graph attention mechanism, H represents the total number of attention heads, and W O represents the output weight matrix, represents the feature matrix after linear transformation of the first updated node feature of node i in path P under the Hth attention head, Represents the feature matrix after linear transformation of the first updated node feature of node j in path P under the Hth attention head.

[0038] The beneficial effects of the present invention are:

[0039] 1. Comprehensive data perspective: By constructing a large dynamic undirected weighted heterogeneous graph, taking photovoltaic power stations, time, and environment as nodes of the graph, a more comprehensive data perspective is provided. Separately constructing a subgraph from the data of each photovoltaic power station is conducive to retaining and mining the individual characteristics of the power station. Connecting each subgraph by distance enhances the spatial correlation understanding of the model. In addition, multi-dimensional information such as time series characteristics and importance characteristics is extracted, which helps to more comprehensively analyze the impact of different factors on electric power.

[0040] 2. Extract more accurate node features: The power time series graph convolutional neural network can better extract the time series features of the nodes. The graph attention mechanism based on node importance and the multi-head path graph attention mechanism simultaneously consider the influence of adjacent nodes and path nodes on node updates, which can better explore the influence of neighbor nodes and improve the expressiveness and generalization capabilities of the prediction model.

[0041] 3. Complete more efficient power system dispatching and planning: The present invention can effectively improve the prediction efficiency of power generation and help the power dispatching center predict power output in advance, maintain power grid stability, and ensure people's livelihood needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1This is a flow chart of the training of the distributed photovoltaic power generation short-term prediction model in the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] The present invention proposes a distributed photovoltaic power generation short-term prediction method based on big data, such as Figure 1 As shown, the method includes the following contents: obtaining relevant data of the photovoltaic power station and inputting it into a trained distributed photovoltaic power generation short-term prediction model to obtain the predicted power generation of the photovoltaic power station.

[0045] The training process of the distributed photovoltaic power generation short-term prediction model includes:

[0046] S1: Obtain relevant data of photovoltaic power stations, construct an undirected weighted heterogeneous graph based on the relevant data of photovoltaic power stations, and obtain the original node features of the nodes in the undirected weighted heterogeneous graph.

[0047] Preferably, the photovoltaic power station related data include the geographical location characteristics of the power station (longitude, latitude, altitude, surface reflectivity), equipment parameter characteristics (photovoltaic panel type, component power, inverter specifications), maintenance status characteristics (photovoltaic panel cleanliness, maintenance status category, fault type, maintenance cycle, last maintenance date, distance between the last maintenance date and now, next expected maintenance date, maintenance completion rate, component life assessment), power station layout characteristics (photovoltaic array orientation, photovoltaic array angle, average photovoltaic array spacing), timestamp characteristics, astronomical characteristics (sun altitude angle, solar azimuth), seasonal identification (season, solar term), holiday identification (whether it is a weekend, whether it is a national holiday), power demand characteristics Characteristics (average daily industrial electricity demand, average daily residential electricity demand, demand peak and demand trough difference), and historical performance characteristics (total power generation in the past 24 hours, average power generation in the past 24 hours, total power generation in the past ten days, average power generation in the past ten days, average power generation when faults occurred in the past, average power generation in the same season in the past, average power generation on holidays in the past, etc.), meteorological observation data (temperature, humidity, air pressure), extreme weather signs (whether it exceeds the historical highest temperature, whether it is lower than the historical lowest temperature, whether there is heavy rainfall, whether it is disastrous weather), lighting condition characteristics (sunshine time, solar radiation), and air quality characteristics (air quality index, pollutant concentration).

[0048] An undirected weighted heterogeneous graph is constructed based on the relevant data of photovoltaic power stations: each power station is regarded as a power station node, each power station node is connected to each corresponding time node to form an edge, each time node is connected to the corresponding environment node to form an edge, and the weight of the edge is 1, so as to obtain multiple subgraphs; the average distance between all power stations is calculated, if the distance between the power station nodes in two subgraphs is less than the average distance, there is an edge between the two power station nodes, and the weight of the edge is the distance between the two nodes, otherwise there is no edge, so as to obtain an undirected weighted heterogeneous graph.

[0049] The process of obtaining the original node features of all nodes includes:

[0050] The original features of the power station node include the geographical location features, equipment parameter features, maintenance status features and power station layout features of the power station; the original features of the time node include timestamp features, astronomical features, seasonal identification, holiday identification, power demand features and historical performance features; the original features of the environment node include meteorological observation data, extreme weather identification, meteorological forecast information, lighting condition features, and air quality features;

[0051] The original features corresponding to each node are encoded and spliced. Specifically, each node performs one-hot processing on the discrete features in the corresponding original features, and splices the processing results with continuous features to obtain the original node features of all nodes; among them, the discrete features are photovoltaic panel type, inverter specifications, maintenance status category, fault type, photovoltaic array orientation, season, solar term, whether it is a weekend, whether it is a national holiday, whether it exceeds the historical highest temperature, whether it is lower than the historical lowest temperature, whether it is heavy precipitation, whether it is disastrous weather, and the rest are continuous features.

[0052] S2: Extracting connectivity and importance features from undirected weighted heterogeneous graphs.

[0053] S21: Construct a connection matrix C as a connectivity feature. The elements C in the matrix ij It represents the length of the shortest path between node i and node j in an undirected weighted heterogeneous graph. All elements on the diagonal of matrix C are 0.

[0054] S22: Construct an importance score matrix as an importance feature. Only the elements on the diagonal of the matrix have values. ii Represents the importance score of node i in the undirected weighted heterogeneous graph; the calculation formula of the importance score is:

[0055]

[0056] Among them, I ii represents the importance score of node i, N represents the number of nodes in the graph, and a ijRepresents the element in the i-th row and j-th column of the adjacency matrix A of the graph. If there is an edge between node i and node j, then a ij =a ji =1, d(i,j) represents the shortest path length from node i to node j, σ st represents the number of shortest paths between node s and node t, σ st (i) represents the number of nodes i passed through in the shortest path.

[0057] S3: According to the undirected weighted heterogeneous graph, the power time series graph convolutional neural network is used to process the importance features and the time series features in the original node features to obtain the first updated node features of the node.

[0058] The time series features are the original node features of the time nodes and environment nodes.

[0059] The formula for processing the importance features and time series features of the power time series graph convolutional neural network is:

[0060]

[0061] in, represents the output feature vector of the current layer at time t, I ii represents the importance score of node i, K represents the width of the convolution kernel, represents the weight vector of the l-th layer convolution kernel at the delayed k position, d (l) represents the expansion rate of the convolution kernel of the lth layer, Indicates that node i is in td (l) The time series characteristics at time k, b (l) represents the bias term of the lth layer, Represents the output feature vector of the previous layer at time t, and ReLU() represents the ReLU activation function.

[0062] The output of the last layer of the network is the new time series features of the nodes.

[0063] In the power generation prediction scenario, historical power generation can reveal the regularity of power generation and better explore the impact of different factors on power generation. The power time series graph convolutional neural network can better process time series features. When designing the power time series graph convolutional neural network, dilated convolution is added. Dilated convolution introduces gaps in the calculation process, allowing the convolution kernel to operate across a larger time step, thereby increasing the receptive field while maintaining a smaller number of parameters, thereby capturing longer-distance dependencies.

[0064] Each time node or environment node is spliced ​​with the original node feature and the new time series feature corresponding to the node to obtain the first updated node feature of the corresponding node; the original node feature of the generator node is used as the first updated node feature of the generator node.

[0065] S4: According to the undirected weighted heterogeneous graph, the graph attention mechanism is used to process the importance features and the first updated node features to obtain the second updated node features of the node.

[0066] The formula for the graph attention mechanism to process the important features and the first updated node features is:

[0067]

[0068] Among them, h′ i represents the second updated node feature of node i, α represents the weighting coefficient, and h i represents the first updated node feature of node i, h j represents the first updated node feature of node j, h k represents the first updated node feature of node k, d represents the dimension of the feature in the graph attention mechanism calculation; W q , W k , W v , W u represents the first, second, third, and fourth learnable weight coefficients, b q , b k , b v , b u represents the first, second, third, and fourth bias terms; N(i) represents the neighboring nodes of node i; LeakyRelu represents the LeakyRelu activation function, softmax represents the normalization function, ReLU represents the ReLU activation function, and I jj represents the importance score of node j, I kk represents the importance score of node k, w ij represents the weight of the edge between node i and node j.

[0069] S5: According to the undirected weighted heterogeneous graph, a multi-head path graph attention mechanism is used to process the connectivity features and the first updated node features to obtain the third updated node features of the node.

[0070] For power station nodes with loops, a loop refers to a path with the node as the starting point and the node as the end point. A loop is randomly selected as the input path, and the multi-head path graph attention mechanism is used to process the connectivity features and the first updated node features of the power station nodes on the path to obtain the third updated node features of the power station nodes. The calculation formula of the multi-head path graph attention mechanism is:

[0071]

[0072] Among them, h″ i represents the third updated node feature of node i, concat represents the concatenation operation, P represents the input path, softmax represents the normalization function, and W a represents the learnable weight matrix, represents the feature matrix after linear transformation of the first updated node feature of node i in path P under the first attention head, represents the feature matrix after linear transformation of the first updated node feature of node j in path P under the first attention head, represents the feature matrix after linear transformation of the first updated node feature of node j in path P, C ij represents the element in the i-th row and j-th column of the connection matrix, d h represents the feature dimension in the calculation of the multi-head path graph attention mechanism, H represents the total number of attention heads, and W O represents the output weight matrix, represents the feature matrix after linear transformation of the first updated node feature of node i in path P under the Hth attention head, Represents the feature matrix after linear transformation of the first updated node feature of node j in path P under the Hth attention head.

[0073] S6: The second updated feature and the third updated feature of the node itself are integrated to obtain the final feature of the node; the final features of the three nodes are concatenated to obtain the comprehensive features of the power station.

[0074] The second updated node feature and the third updated node feature are weightedly fused to obtain the final feature of the node; wherein, if the node does not have the third updated node feature, the second updated node feature of the node is used as the final feature;

[0075] The final features of the power station node are spliced ​​with the final features of its corresponding time node and environment node to obtain the comprehensive features of the power station at the corresponding time node.

[0076] S7: Input the comprehensive features into a multi-layer perceptron for processing to obtain the predicted power generation of the photovoltaic power station.

[0077] The multi-layer perceptron includes multiple fully connected layers and prediction layers. The comprehensive features are input into the multi-layer perceptron for processing, and the predicted power generation of the photovoltaic power station at each moment can be output.

[0078] S8: Calculate the total model loss and adjust the model parameters according to the total model loss to obtain a trained distributed photovoltaic power generation short-term prediction model.

[0079] The total loss of the model can be the mean absolute error loss function. Back propagation is performed according to the loss function to adjust the model parameters. When the loss function is minimized or reaches a preset number of iterations, a trained distributed photovoltaic power generation short-term prediction model is obtained.

[0080] The photovoltaic power station related data required for prediction, namely the original features of the power station node, the original features of the time node and the original features of the environment node, are obtained; the data are input into the trained distributed photovoltaic power generation short-term prediction model for processing, and the predicted power generation of the photovoltaic power station in the prediction period can be obtained.

[0081] In summary, the present invention processes the original photovoltaic power station related data, expresses the useful information of the data through an undirected weighted heterogeneous graph, extracts the time series features, importance features, etc. in the data, retains the original node features by constructing subgraphs, and connects each subgraph by distance to dig out the data patterns between similar power stations. The power time series graph neural network effectively utilizes the time series features, and the current power generation can be predicted based on the existing features through the feature distribution in the historical data. The graph attention mechanism based on the importance of nodes allows nodes to pay attention to more important adjacent nodes when updating their own features. The multi-head path graph attention mechanism allows nodes to have a broader field of view and extract more effective features for power prediction.

[0082] The present invention predicts the power generation of photovoltaic power stations under different time and environment through effective analysis of relevant data of photovoltaic power stations, which can effectively assist the power grid control center to make corresponding decisions and ensure people's power supply needs. Through iterative learning and optimization of the graph neural network model, the dynamic impact of time and environment on power generation can be simulated. The use of the graph attention mechanism realizes the dynamic assignment of different attention weights to each node in the network, which helps the model focus on the node features that have a greater impact on power, thereby improving the accuracy and pertinence of the model prediction.

[0083] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A short-term prediction method for distributed photovoltaic power generation based on big data, characterized in that: include: Obtain relevant data of the photovoltaic power station and input it into the trained distributed photovoltaic power generation short-term prediction model to obtain the predicted power generation of the photovoltaic power station; The training process of the distributed photovoltaic power generation short-term prediction model includes: Step S1: Obtain relevant data of the photovoltaic power station, construct an undirected weighted heterogeneous graph according to the relevant data of the photovoltaic power station, and obtain original node features of all nodes in the undirected weighted heterogeneous graph; Step S2: extracting connectivity features and importance features from the undirected weighted heterogeneous graph; Step S2 includes: Step S21: Construct a connection matrix C as a connectivity feature. The elements C in the matrix ij It represents the length of the shortest path between node i and node j in an undirected weighted heterogeneous graph. All elements on the diagonal of matrix C are 0. Step S22: Construct an importance score matrix as an importance feature. Only the elements on the diagonal of the matrix have values. ii Represents the importance score of node i in the undirected weighted heterogeneous graph; Step S3: According to the undirected weighted heterogeneous graph, a power time series graph convolutional neural network is used to process the importance features and the time series features in the original node features to obtain the first updated node features of the node; Step S3 specifically includes: The undirected weighted heterogeneous graph includes power station nodes, time nodes and environment nodes. The power time series graph convolutional neural network is used to process the importance features and the time series features in the original node features to obtain new time series features. Among them, the time series features are the original node features of the time nodes and environment nodes. Each time node or environment node is spliced ​​with the original node feature and the new time series feature corresponding to the node to obtain the first updated node feature of the corresponding node; the original node feature of the power station node is used as the first updated node feature of the power station node; Step S4: According to the undirected weighted heterogeneous graph, the graph attention mechanism is used to process the importance feature and the first updated node feature to obtain the second updated node feature of the node; the formula for the graph attention mechanism to process the importance feature and the first updated node feature is: Among them, h′ i represents the second updated node feature of node i, α represents the weighting coefficient, and h i represents the first updated node feature of node i, h j represents the first updated node feature of node j, h k represents the first updated node feature of node k, d represents the feature dimension in the graph attention mechanism calculation; W q , W k , W v , W u represents the first, second, third, and fourth learnable weight coefficients, b q 、b k 、b v 、b u represents the first, second, third, and fourth bias terms; N(i) represents the neighboring nodes of node i; LeakyRelu represents the LeakyRelu activation function, softmax represents the normalization function, ReLU represents the ReLU activation function, and I jj represents the importance score of node j, I kk represents the importance score of node k, w ij represents the weight of the edge between node i and node j; Step S5: According to the undirected weighted heterogeneous graph, a multi-head path graph attention mechanism is used to process the connectivity feature and the first updated node feature to obtain the third updated node feature of the node; Step S6: Fusing the second updated feature and the third updated feature of the node itself to obtain the final feature of the node; splicing the final features of the three nodes to obtain the comprehensive feature of the power station; Step S7: input the comprehensive features into a multi-layer perceptron for processing to obtain the predicted power generation of the photovoltaic power station; Step S8: Calculate the total model loss and adjust the model parameters according to the total model loss to obtain a trained distributed photovoltaic power generation short-term prediction model.

2. According to the method of short-term prediction of distributed photovoltaic power generation based on big data in claim 1, it is characterized in that: The process of constructing an undirected weighted heterogeneous graph includes: treating each power station as a power station node, connecting each power station node with each corresponding time node to form an edge, connecting each time node with the corresponding environment node to form an edge, and the weight of the edge is 1, thereby obtaining multiple subgraphs; calculating the average distance between all power stations, if the distance between the power station nodes in two subgraphs is less than the average distance, there is an edge between the two power station nodes, and the weight of the edge is the distance between the two nodes, otherwise there is no edge, thereby obtaining an undirected weighted heterogeneous graph.

3. The method for short-term prediction of distributed photovoltaic power generation based on big data according to claim 1 is characterized in that: The process of obtaining the original node features of all nodes includes: The original features of the power station node include the geographical location features, equipment parameter features, maintenance status features and power station layout features of the power station; the original features of the time node include timestamp features, astronomical features, seasonal identification, holiday identification, power demand features and historical performance features; the original features of the environment node include meteorological observation data, extreme weather identification, meteorological forecast information, lighting condition features, and air quality features; The original features corresponding to each node are encoded and concatenated to obtain the original node features of all nodes.

4. The method for short-term prediction of distributed photovoltaic power generation based on big data according to claim 1 is characterized in that: The importance score is calculated as: Among them, I ii represents the importance score of node i, N represents the number of nodes in the graph, and a ij Represents the element in the i-th row and j-th column of the adjacency matrix A of the graph. If there is an edge between node i and node j, then a ij =a ji =1, d(i,j) represents the shortest path length from node i to node j, σ st represents the number of shortest paths between node s and node t, σ st (i) represents the number of nodes i passed through in the shortest path.

5. The method for short-term prediction of distributed photovoltaic power generation based on big data according to claim 1 is characterized in that: The formula for obtaining the new time series features is: in, represents the output feature vector of the current layer at time t, I ii represents the importance score of node i, K represents the width of the convolution kernel, represents the weight vector of the l-th layer convolution kernel at the delayed k position, d (l) represents the expansion rate of the convolution kernel of the lth layer, Indicates that node i is in td (l) The time series characteristics at time k, b (l) represents the bias term of the lth layer, Represents the output feature vector of the previous layer at time t, and ReLU() represents the ReLU activation function.

6. The method for short-term prediction of distributed photovoltaic power generation based on big data according to claim 1 is characterized in that: In step S5, the process of processing the connectivity feature and the first updated node feature using the multi-head path graph attention mechanism includes: For power station nodes with loops, a loop is randomly selected as the input path, and the connectivity features and the first updated node features of the power station nodes on the path are processed using the multi-head path graph attention mechanism to obtain the third updated node features of the power station nodes; The calculation formula of the multi-head path graph attention mechanism is: Among them, h″ i represents the third updated node feature of node i, concat represents the concatenation operation, P represents the input path, softmax represents the normalization function, and W a represents the learnable weight matrix, represents the feature matrix after linear transformation of the first updated node feature of node i in path P under the first attention head, represents the feature matrix after linear transformation of the first updated node feature of node j in path P under the first attention head, represents the feature matrix after linear transformation of the first updated node feature of node j in path P, C ij represents the element in the i-th row and j-th column of the connection matrix, dh represents the feature dimension in the calculation of the multi-head path graph attention mechanism, H represents the total number of attention heads, and W O represents the output weight matrix, represents the feature matrix after linear transformation of the first updated node feature of node i in path P under the Hth attention head, Represents the feature matrix after linear transformation of the first updated node feature of node j in path P under the Hth attention head.

Images