Distributed photovoltaic power prediction method, device and equipment, readable storage medium

Abstract

The application relates to a distributed photovoltaic power prediction method and device, computer equipment and a computer readable storage medium. The method comprises the following steps: collecting output data and meteorological data of a distributed photovoltaic power station group comprising a plurality of photovoltaic power stations, constructing an original node matrix for representing the multi-dimensional characteristics of the output of each photovoltaic power station and the corresponding meteorological data according to the two types of data, carrying out node embedding processing on the original node matrix to obtain a node embedding matrix with a lower dimension, and finally inputting the node embedding matrix into a trained graph neural network model to output the power prediction result of the distributed photovoltaic power station group. The method can effectively improve the accuracy and efficiency of power prediction of the distributed photovoltaic power station group under complex meteorological scenarios.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power prediction technology, and in particular to a distributed photovoltaic power prediction method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] With the development of photovoltaic power prediction technology, distributed photovoltaic power prediction technology has emerged. This technology focuses on the output prediction of distributed photovoltaic power plants, and its core is to achieve accurate prediction by extracting time-series features and building dynamic models.

[0003] Traditional technologies primarily rely on historical power output data and observations from a single meteorological station, treating each photovoltaic node as an independent entity for forecasting. They address the power output forecasting needs of distributed photovoltaic systems through only simple time-series analysis and basic modeling.

[0004] However, current traditional methods neglect the coupling effect of spatial correlation between photovoltaic nodes and fail to combine multiple factors such as meteorological changes and geographical location for comprehensive prediction, resulting in significant prediction errors in complex meteorological scenarios. Summary of the Invention

[0005] Therefore, it is necessary to provide a distributed photovoltaic power prediction method, device, computer equipment, computer-readable storage medium, and computer program product that can accurately match the diverse needs of actual dispatching under complex meteorological scenarios to address the above-mentioned technical problems.

[0006] In a first aspect, this application provides a method for predicting distributed photovoltaic power, including:

[0007] Acquire power output data and meteorological data corresponding to a distributed photovoltaic power station cluster; a distributed photovoltaic power station cluster includes multiple photovoltaic power stations.

[0008] An original node matrix is ​​constructed based on power output data and meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0009] Perform node embedding processing on the original node matrix to obtain the node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0010] The node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0011] In one embodiment, the graph neural network model includes a graph learning layer, a temporal feature extraction layer, a spatial feature extraction layer, and an output layer connected in sequence; the node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group, including:

[0012] The node embedding matrix is ​​input into the graph learning layer to output an asymmetric adjacency matrix;

[0013] The asymmetric adjacency matrix is ​​input into the temporal feature extraction layer to output the temporal feature matrix;

[0014] The asymmetric adjacency matrix is ​​input into the spatial feature extraction layer to output the spatial feature matrix;

[0015] The temporal and spatial feature matrices are input into the output layer to output the power prediction results of the distributed photovoltaic power station group.

[0016] In one embodiment, the node embedding matrix is ​​input to the graph learning layer to output an asymmetric adjacency matrix, including:

[0017] Based on the node embedding matrix, the first intermediate matrix and the second intermediate matrix are obtained respectively through learnable parameters;

[0018] Calculate the first product of the transpose of the first intermediate matrix and the second intermediate matrix, and the second product of the transpose of the second intermediate matrix and the first intermediate matrix. Subtract the second product from the first product to obtain the difference matrix.

[0019] The difference matrix is ​​processed by an activation function to obtain an asymmetric adjacency matrix.

[0020] In one embodiment, the asymmetric adjacency matrix is ​​input to the temporal feature extraction layer to output a temporal feature matrix, including:

[0021] Multi-scale temporal features are extracted from the input temporal sequence based on the node embedding matrix. The multi-scale temporal features obtained by convolution with different dilation rates are concatenated to obtain a multi-scale feature matrix.

[0022] The multi-scale feature matrix is ​​averaged and pooled to obtain the global temporal feature matrix.

[0023] Based on the global temporal feature matrix and the asymmetric adjacency matrix, a dynamic adjacency matrix is ​​generated through dynamic convolution and activation functions;

[0024] The hidden layer state is updated based on the multi-scale feature matrix, the asymmetric adjacency matrix, the dynamic adjacency matrix, and the hidden state of the previous time step.

[0025] After performing residual connections on the hidden layer states, the temporal feature matrix is ​​output.

[0026] In one embodiment, the spatial feature extraction layer consists of two hybrid skip propagation layers; inputting an asymmetric adjacency matrix into the spatial feature extraction layer to output a spatial feature matrix includes:

[0027] Based on the asymmetric adjacency matrix, degree normalization is performed, and the spatial features after propagation are obtained through the first hybrid hop propagation layer.

[0028] Based on spatial features combined with a learnable weight matrix, the spatial features after propagation are filtered through a second hybrid jump propagation layer to obtain a filtered spatial feature matrix.

[0029] In one embodiment, the temporal feature matrix and the spatial feature matrix are input to the output layer to output the power prediction results of the distributed photovoltaic power station group, including:

[0030] The temporal feature matrix and the spatial feature matrix are added together, and the result is normalized.

[0031] The processing results are input into the fully connected layer, and the power prediction results of the distributed photovoltaic power station group are output.

[0032] In one embodiment, the method further includes:

[0033] Obtain historical power output data and historical meteorological data corresponding to distributed photovoltaic power station groups;

[0034] Based on historical power output data and historical meteorological data, a training set and a test set are constructed;

[0035] The pre-built graph neural network model is trained using the training set to obtain the initial graph neural network model;

[0036] The initial graph neural network model was evaluated using a test set to obtain the evaluation results;

[0037] If the evaluation results do not meet the preset standards, the parameters of the initial graph neural network model are adjusted according to the evaluation results, and the prediction model is retrained using the training set until the preset standards are met.

[0038] Secondly, this application also provides a training model device for distributed photovoltaic power generation prediction, comprising:

[0039] The acquisition module is used to acquire power output data and meteorological data corresponding to the distributed photovoltaic power station group; the distributed photovoltaic power station group includes multiple photovoltaic power stations;

[0040] The module is used to construct the original node matrix based on power output data and meteorological data; the original node matrix is ​​used to represent the multi-dimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multi-dimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions.

[0041] The processing module is used to perform node embedding processing on the original node matrix to obtain a node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0042] The output module is used to input the node embedding matrix into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0043] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0044] Acquire power output data and meteorological data corresponding to a distributed photovoltaic power station cluster; a distributed photovoltaic power station cluster includes multiple photovoltaic power stations.

[0045] An original node matrix is ​​constructed based on power output data and meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0046] Perform node embedding processing on the original node matrix to obtain the node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0047] The node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0049] Acquire power output data and meteorological data corresponding to a distributed photovoltaic power station cluster; a distributed photovoltaic power station cluster includes multiple photovoltaic power stations.

[0050] An original node matrix is ​​constructed based on power output data and meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0051] Perform node embedding processing on the original node matrix to obtain the node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0052] The node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0053] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0054] Acquire power output data and meteorological data corresponding to a distributed photovoltaic power station cluster; a distributed photovoltaic power station cluster includes multiple photovoltaic power stations.

[0055] An original node matrix is ​​constructed based on power output data and meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0056] Perform node embedding processing on the original node matrix to obtain the node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0057] The node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0058] The aforementioned distributed photovoltaic (PV) power prediction method, apparatus, computer equipment, computer-readable storage medium, and computer program product first collect power output data and meteorological data from a distributed PV power station cluster containing multiple PV power stations. Based on these two types of data, an original node matrix is ​​constructed to characterize the multi-dimensional features of each PV power station's output and corresponding meteorological conditions. Then, node embedding processing is performed on the original node matrix to obtain a lower-dimensional node embedding matrix. Finally, the node embedding matrix is ​​input into a trained graph neural network model to output the power prediction result for the distributed PV power station cluster. Node embedding processing achieves dimensionality reduction and efficient characterization of the multi-dimensional features of PV power stations. Simultaneously, by leveraging the graph neural network model to mine the internal correlation characteristics of the distributed PV power station cluster, the mapping relationship between the power station cluster's output and meteorological factors can be accurately captured. This reduces data redundancy while effectively improving the accuracy and efficiency of power prediction for distributed PV power station clusters, adapting to the power prediction needs of multi-power station collaborative scenarios. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1This is a flowchart illustrating a distributed photovoltaic power prediction method in one embodiment;

[0061] Figure 2 This is a flowchart illustrating the steps of a distributed photovoltaic power prediction method in one embodiment;

[0062] Figure 3 This is a schematic diagram of a distributed photovoltaic power prediction method in another embodiment;

[0063] Figure 4 This is a structural block diagram of a distributed photovoltaic power prediction device in one embodiment;

[0064] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0066] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0067] In one embodiment, such as Figure 1 As shown, a distributed photovoltaic power prediction method is provided. This embodiment illustrates the method's application to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0068] Step 102: Obtain the power output data and meteorological data corresponding to the distributed photovoltaic power station group; the distributed photovoltaic power station group includes multiple photovoltaic power stations;

[0069] Among them, distributed photovoltaic power station clusters refer to a collection of multiple photovoltaic power stations that are geographically dispersed, electrically interconnected, or operate independently; output data includes relevant data reflecting the power generation capacity, such as the historical actual power generation, installed capacity, and operating status of each photovoltaic power station; meteorological data includes meteorological elements that affect photovoltaic output, such as solar irradiance, ambient temperature, wind speed, humidity, air pressure, and weather forecast data.

[0070] Optionally, the power output data and meteorological data of each photovoltaic power station can be collected and summarized in real time or periodically through power station data acquisition terminals, SCADA systems, meteorological monitoring stations or third-party meteorological data interfaces to form a multi-source time-series dataset for the entire power station group, providing raw data support for subsequent feature construction.

[0071] Step 104: Construct an original node matrix based on power output data and meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0072] The original node matrix is ​​a matrix structure constructed with each photovoltaic power station in the distributed photovoltaic power station group as a node and the corresponding power output data and meteorological data of each power station as feature dimensions. The multi-dimensional feature refers to the multi-dimensional feature vector formed by integrating the power output information of the photovoltaic power station and the corresponding meteorological element information, which is used to comprehensively characterize the power generation status and external environmental conditions of a single photovoltaic power station.

[0073] Optionally, during the construction process, each photovoltaic power station is treated as an independent node. The output characteristics of the corresponding node, such as historical output and real-time output, are spliced ​​with meteorological characteristics such as irradiance, temperature, humidity, and wind speed to form a multi-dimensional feature vector of the corresponding node. Then, the multi-dimensional feature vectors of all photovoltaic power stations are combined and arranged in a preset order to construct the original node matrix covering the entire distributed photovoltaic power station group.

[0074] Step 106: Perform node embedding processing on the original node matrix to obtain the node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0075] Among them, node embedding processing refers to the feature mapping and feature compression of the high-dimensional sparse original node features, and the feature processing process of converting the high-dimensional feature vector into a low-dimensional dense vector; the node embedding matrix refers to the low-dimensional matrix obtained after node embedding dimensionality reduction processing, which is used to represent the core feature information of each photovoltaic power station node. It retains the key feature information in the original node matrix, while the feature dimension is significantly reduced.

[0076] Optionally, a feature reduction algorithm or a graph node embedding algorithm can be used to perform a nonlinear mapping on the original node matrix. While retaining the correlation features and key attribute information between the nodes of each photovoltaic power station, redundant features are eliminated and the feature dimension is compressed, thereby transforming the high-dimensional original node matrix into a low-dimensional dense node embedding matrix, achieving efficient representation of the features of the power station nodes.

[0077] Step 108: Input the node embedding matrix into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0078] Among them, the well-trained graph neural network model refers to a graph neural network model that has been trained, verified and optimized with sufficient sample data (including photovoltaic power plant output, meteorological related samples and corresponding labels), with fixed model parameters and prediction accuracy and generalization ability meeting the preset requirements for distributed photovoltaic power prediction; the node embedding matrix refers to a low-dimensional dense matrix obtained after node embedding processing, which retains the key features of each photovoltaic power plant node and has a lower dimension than the original node matrix; the power prediction result refers to the power generation prediction data of the distributed photovoltaic power plant group within a preset time period output by the graph neural network model based on the input node embedding matrix, which may include the total power prediction value of the power plant group and the power prediction value of each individual photovoltaic power plant.

[0079] Optionally, in the specific implementation, the pre-trained graph neural network model is first deployed to ensure that the model is in a stable running state and can receive input data and perform calculations normally. Then, the obtained node embedding matrix is ​​used as the input data of the model. This node embedding matrix has integrated the core features of all photovoltaic power stations in the distributed photovoltaic power station group and the potential correlation information between nodes. The graph neural network model deeply explores the collaborative correlation rules between the nodes of each photovoltaic power station through core processes such as internal graph convolution operation and node feature aggregation. At the same time, it fits the mapping relationship between the low-dimensional features of nodes and photovoltaic power output to complete feature extraction and prediction calculation. Finally, the pre-trained graph neural network model outputs the power prediction results of the distributed photovoltaic power station group, providing accurate data support for subsequent work such as power generation scheduling, power control, and grid connection adaptation of the distributed photovoltaic power station group, and realizing efficient power prediction in multi-power station collaborative scenarios.

[0080] The aforementioned distributed photovoltaic (PV) power prediction method first collects output data and meteorological data from a distributed PV power station cluster containing multiple PV power plants. Based on these two types of data, an original node matrix is ​​constructed to characterize the multi-dimensional features of each PV power plant's output and corresponding meteorological conditions. Then, node embedding processing is performed on the original node matrix to obtain a lower-dimensional node embedding matrix. Finally, the node embedding matrix is ​​input into a trained graph neural network model to output the power prediction result for the distributed PV power station cluster. Node embedding processing achieves dimensionality reduction and efficient characterization of the multi-dimensional features of PV power plants. Simultaneously, by leveraging the graph neural network model to mine the correlation characteristics within the distributed PV power station cluster, the method can accurately capture the mapping relationship between the power station cluster's output and meteorological factors. This reduces data redundancy while effectively improving the accuracy and efficiency of power prediction for distributed PV power station clusters, adapting to the power prediction needs of multi-power station collaborative scenarios.

[0081] In one embodiment, the graph neural network model includes a graph learning layer, a temporal feature extraction layer, a spatial feature extraction layer, and an output layer connected in sequence; the node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group, including:

[0082] The node embedding matrix is ​​input into the graph learning layer to output an asymmetric adjacency matrix;

[0083] The asymmetric adjacency matrix is ​​input into the temporal feature extraction layer to output the temporal feature matrix;

[0084] The asymmetric adjacency matrix is ​​input into the spatial feature extraction layer to output the spatial feature matrix;

[0085] The temporal and spatial feature matrices are input into the output layer to output the power prediction results of the distributed photovoltaic power station group.

[0086] Specifically, the graph neural network model consists of a graph learning layer, a temporal feature extraction layer, a spatial feature extraction layer, and an output layer connected in sequence. When performing power prediction, the node embedding matrix is ​​first input into the graph learning layer to obtain an asymmetric adjacency matrix. Then, the asymmetric adjacency matrix is ​​input into the temporal feature extraction layer and the spatial feature extraction layer respectively, and the temporal feature matrix and the spatial feature matrix are output accordingly. Finally, the two types of feature matrices are input into the output layer, and after fusion processing, the power prediction result of the distributed photovoltaic power station group is output.

[0087] In the above embodiments, the hierarchical model structure can extract the temporal variation patterns and spatial correlation characteristics of photovoltaic power station groups in a targeted manner. The asymmetric adjacency matrix better reflects the complex non-equivalent influence relationships between actual distributed photovoltaic power stations. The independent extraction and fusion of temporal and spatial features can fully explore multi-dimensional correlation information, effectively improve the model's fitting accuracy and prediction reliability of photovoltaic power, and make the feature learning process clearer and the model more interpretable.

[0088] In one embodiment, the node embedding matrix is ​​input to the graph learning layer to output an asymmetric adjacency matrix, including:

[0089] Based on the node embedding matrix, the first intermediate matrix and the second intermediate matrix are obtained respectively through learnable parameters;

[0090] Calculate the first product of the transpose of the first intermediate matrix and the second intermediate matrix, and the second product of the transpose of the second intermediate matrix and the first intermediate matrix. Subtract the second product from the first product to obtain the difference matrix.

[0091] The difference matrix is ​​processed by an activation function to obtain an asymmetric adjacency matrix.

[0092] Specifically, a graph learning layer is constructed to calculate an asymmetric adjacency matrix and capture the hidden relationships between time series data.

[0093] The asymmetric adjacency matrix is:

[0094]

[0095] in, For the intermediate matrix, The output node embedding matrix, This is the scaling factor. For learnable parameters, The sigmoid activation function is used, and the intermediate matrix is ​​used. The generation method is similar.

[0096] In the above embodiments, by introducing learnable parameters to adaptively generate intermediate matrices and using matrix product differences to construct adjacency relationships, the asymmetric and directional spatial correlation characteristics between distributed photovoltaic power stations can be effectively learned and reflected, avoiding the simplification and distortion of actual correlation relationships by symmetric adjacency matrices. At the same time, the activation function is combined to ensure the numerical rationality and feature expression capability of the output matrix, thereby improving the accuracy of subsequent temporal and spatial feature extraction.

[0097] In one embodiment, the asymmetric adjacency matrix is ​​input to the temporal feature extraction layer to output a temporal feature matrix, including:

[0098] Multi-scale temporal features are extracted from the input temporal sequence based on the node embedding matrix. The multi-scale temporal features obtained by convolution with different dilation rates are concatenated to obtain a multi-scale feature matrix.

[0099] The multi-scale feature matrix is ​​averaged and pooled to obtain the global temporal feature matrix.

[0100] Based on the global temporal feature matrix and the asymmetric adjacency matrix, a dynamic adjacency matrix is ​​generated through dynamic convolution and activation functions;

[0101] The hidden layer state is updated based on the multi-scale feature matrix, the asymmetric adjacency matrix, the dynamic adjacency matrix, and the hidden state of the previous time step.

[0102] After performing residual connections on the hidden layer states, the temporal feature matrix is ​​output.

[0103] Specifically, the input sequence is processed through an expanded initial layer. Perform multi-scale temporal convolution:

[0104]

[0105] in, The total number of time steps. This refers to batch size in deep learning. This indicates a concatenation operation along the feature dimension, integrating multi-scale temporal features extracted by convolutions with different dilation rates into a unified feature matrix. Thus, short-term, medium-term, and long-term temporal patterns can be represented simultaneously in the same feature space; Indicates the expansion rate One-dimensional temporal convolution.

[0106] Generate a dynamic adjacency matrix based on the current input and historical states. :

[0107]

[0108] in, The dynamic adjacency matrix represents the global temporal characteristics. This is a dynamic intermediate matrix.

[0109] Update the hidden layer state using a bidirectional gated loop unit:

[0110]

[0111] in, For the corresponding learnable parameters, This is the hidden layer state. This represents a dynamic graph convolution operation. These are the update door and the reset door, respectively. These are candidate states for the hidden layer. This represents the current hidden layer state.

[0112] Adding residual connections yields time features:

[0113]

[0114] in, It is a learnable residual matrix.

[0115] In the above embodiments, multi-scale dilated convolution can effectively capture the temporal variation of photovoltaic power output at different time granularities, average pooling can extract global temporal information, dynamic adjacency matrix can be combined to realize dynamic adaptation of temporal features and spatial correlation, and residual connection can alleviate the gradient vanishing problem when extracting deep features. This not only enhances the model's ability to fit complex temporal fluctuations, but also improves the stability and robustness of temporal feature extraction.

[0116] In one embodiment, the spatial feature extraction layer consists of two hybrid skip propagation layers; inputting an asymmetric adjacency matrix into the spatial feature extraction layer to output a spatial feature matrix includes:

[0117] Based on the asymmetric adjacency matrix, degree normalization is performed, and the spatial features after propagation are obtained through the first hybrid hop propagation layer.

[0118] Based on spatial features combined with a learnable weight matrix, the spatial features after propagation are filtered through a second hybrid jump propagation layer to obtain a filtered spatial feature matrix.

[0119] Specifically, the spatial feature extraction layer consists of two hybrid jump propagation layers, which process the inflow and outflow information through each node respectively. The net inflow information is obtained by adding the outputs of the two hybrid jump propagation layers.

[0120] The established hybrid jump propagation layer consists of two steps, and the information propagation steps are as follows:

[0121]

[0122] in, is a hyperparameter (range 0.1-0.9) used to control the fusion ratio of node information and neighboring node information in graph convolution operations. It is usually determined through grid search or cross-validation. is the propagation depth, equal to the number of hybrid jump propagation layers. It indicates the range of spatial information propagated on the map.

[0123] The information selection steps are as follows:

[0124]

[0125] in, The normalized graph adjacency matrix, This is the inverse of the degree matrix, used for degree normalization. for A 3D identity matrix is ​​used to preserve information about the nodes themselves; This is a learnable weight matrix.

[0126] In the above embodiments, the structure of the dual hybrid jump propagation layer is adopted and combined with degree normalization and learnable weight screening. On the basis of making full use of the asymmetric adjacency matrix to characterize the spatial relationship between photovoltaic power stations, more accurate and efficient spatial feature propagation can be achieved. At the same time, redundant spatial information is effectively filtered, key spatial relationship features are strengthened, and the expressive power and distinguishability of spatial features are improved.

[0127] In one embodiment, the temporal feature matrix and the spatial feature matrix are input to the output layer to output the power prediction results of the distributed photovoltaic power station group, including:

[0128] The temporal feature matrix and the spatial feature matrix are added together, and the result is normalized.

[0129] The processing results are input into the fully connected layer, and the power prediction results of the distributed photovoltaic power station group are output.

[0130] Specifically, the constructed graph neural network model includes a graph learning layer, a temporal feature extraction layer, a spatial feature extraction layer, and an output layer connected in sequence.

[0131] Feature fusion specifically involves adding temporal and spatial features and then performing layer normalization:

[0132]

[0133] in, Presentation layer normalization technique.

[0134] The features are mapped to distributed photovoltaic power generation through fully connected layers at multiple time scales, and the prediction results are output:

[0135]

[0136] in, Time scale The predicted output sequence, They are time scales The learnable weight matrix and bias vector.

[0137] To optimize the multi-scale prediction model for distributed photovoltaic power, a loss function is defined as follows:

[0138]

[0139] in, Time scale Actual output data The regularization coefficient is . This is the regularization term for the model parameters.

[0140] Calculate the loss value of the loss function, and based on this loss value, update the parameters of each layer of the graph neural network model through backpropagation. Iterate and optimize until the loss value converges, thus completing the training of the multi-scale prediction model.

[0141] In the above embodiments, the lightweight fusion of temporal and spatial features is achieved by direct addition. Combined with normalization processing, the feature value distribution can be stabilized and the dimensional differences can be eliminated. Then, the fully connected layer completes efficient feature mapping and prediction output. While ensuring prediction accuracy, the model calculation process is simplified and the prediction efficiency is improved. This allows the temporal patterns and spatial correlation information to fully coordinate, further improving the accuracy and stability of photovoltaic power prediction results.

[0142] In one embodiment, the method further includes:

[0143] Obtain historical power output data and historical meteorological data corresponding to distributed photovoltaic power station groups;

[0144] Based on historical power output data and historical meteorological data, a training set and a test set are constructed;

[0145] The pre-built graph neural network model is trained using the training set to obtain the initial graph neural network model;

[0146] The initial graph neural network model was evaluated using a test set to obtain the evaluation results;

[0147] If the evaluation results do not meet the preset standards, the parameters of the initial graph neural network model are adjusted according to the evaluation results, and the prediction model is retrained using the training set until the preset standards are met.

[0148] Specifically, the dataset obtained by S1 is divided into training set and test set according to time series.

[0149] The trained graph neural network model is evaluated using a test set, with the root mean square error (RMSE) as the evaluation metric, to determine the model's prediction accuracy.

[0150] If the prediction accuracy fails to meet the preset standard, the model parameters are adjusted based on the evaluation feedback, and the graph neural network model is retrained until the prediction accuracy meets the requirements.

[0151] In the above embodiments, the model training and optimization process achieves standardized training and objective evaluation of the model by dividing the training set and the test set. Combined with the iterative parameter adjustment mechanism, it can effectively improve the model's fitting ability and generalization ability, avoid the problem of overfitting or underfitting, and ensure that the graph neural network model used for prediction has stable and reliable prediction performance, providing a reliable model foundation for power prediction of distributed photovoltaic power station groups.

[0152] In one embodiment, such as Figure 2 The diagram illustrates a distributed photovoltaic power prediction method in a specific embodiment, comprising the following steps:

[0153] S1: Obtain historical power output data and meteorological data of distributed photovoltaic power station groups, construct datasets and perform data preprocessing;

[0154] Data preprocessing includes missing value handling, outlier detection, and feature normalization. Based on this, the original node matrix is ​​constructed. ,in For the number of photovoltaic power plants, The feature dimension includes multi-dimensional features such as historical power generation, power change rate, irradiance, temperature, and humidity. Subsequently, node embedding is performed through linear transformation and activation functions, mapping the original node matrix to a low-dimensional embedding space.

[0155] (1)

[0156] in, The output node embedding matrix, The learnable weight matrix for the node embedding layer. For embedded dimensions, It is a linear rectification activation function.

[0157] S2: Construct a graph learning layer to calculate an asymmetric adjacency matrix and capture the hidden relationships between time series data;

[0158] The asymmetric adjacency matrix is:

[0159] (2)

[0160] in, For the intermediate matrix, This is the scaling factor. For learnable parameters, The sigmoid activation function is used, and the intermediate matrix is ​​used. The generation method is similar.

[0161] S3: Construct a dynamic recurrent network module, and extract temporal features using a temporal feature extraction layer;

[0162] S3.1: Input sequence through an expanded initial layer Perform multi-scale temporal convolution:

[0163] (3)

[0164] in, The total number of time steps. This refers to batch size in deep learning. This indicates a concatenation operation along the feature dimension, integrating multi-scale temporal features extracted by convolutions with different dilation rates into a unified feature matrix. Thus, short-term, medium-term, and long-term temporal patterns can be represented simultaneously in the same feature space; Indicates the expansion rate One-dimensional temporal convolution.

[0165] S3.2: Generate a dynamic adjacency matrix based on the current input and historical states. :

[0166] (4)

[0167] in, The dynamic adjacency matrix represents the global temporal characteristics. This is a dynamic intermediate matrix.

[0168] S3.3: Update the hidden layer state via a bidirectional gated loop unit:

[0169] (5)

[0170] in, For the corresponding learnable parameters, This is the hidden layer state. This represents a dynamic graph convolution operation. These are the update door and the reset door, respectively. These are candidate states for the hidden layer. This represents the current hidden layer state.

[0171] S3.4: Adding residual connections yields time characteristics:

[0172] (6)

[0173] in, It is a learnable residual matrix.

[0174] S4: Construct a directed graph convolution module and a spatial feature extraction layer to extract spatial features;

[0175] S4.1: The spatial feature extraction layer of the directed graph convolution module consists of two hybrid jump propagation layers, which process the inflow and outflow information through each node respectively. The net inflow information is obtained by adding the outputs of the two hybrid jump propagation layers.

[0176] S4.2: The established hybrid jump propagation layer includes two steps, and the information propagation steps are as follows:

[0177] (7)

[0178] in, is a hyperparameter (range 0.1-0.9) used to control the fusion ratio of node information and neighboring node information in graph convolution operations. It is usually determined through grid search or cross-validation. is the propagation depth, equal to the number of hybrid jump propagation layers. It indicates the range of spatial information propagated on the map.

[0179] The information selection steps are as follows:

[0180] (8)

[0181] in, The normalized graph adjacency matrix, This is the inverse of the degree matrix, used for degree normalization. for A 3D identity matrix is ​​used to preserve information about the nodes themselves; This is a learnable weight matrix.

[0182] S5: Construct a graph neural network model of dynamic directed graph network, fuse temporal and spatial features, and output multi-scale prediction results;

[0183] S5.1: The constructed graph neural network model includes a graph learning layer, a temporal feature extraction layer (dynamic recurrent network module), a spatial feature extraction layer (directed graph convolution module), and an output layer connected in sequence. The principle of the graph neural network model based on S2-S4 is as follows: Figure 3 As shown.

[0184] S5.2: Feature fusion specifically involves adding temporal and spatial features and then performing layer normalization.

[0185] (9)

[0186] in, Presentation layer normalization technique.

[0187] S5.3: Features are mapped to distributed photovoltaic power generation through fully connected layers at multiple time scales, and prediction results are output.

[0188] (10)

[0189] in, Time scale The predicted output sequence, They are time scales The learnable weight matrix and bias vector.

[0190] S5.4: To optimize the multi-scale prediction model for distributed photovoltaic power, the loss function is defined as follows:

[0191] (11)

[0192] in, Time scale Actual output data The regularization coefficient is . This is the regularization term for the model parameters.

[0193] Calculate the loss value of the loss function, and based on this loss value, update the parameters of each layer of the graph neural network model through backpropagation. Iterate and optimize until the loss value converges, thus completing the training of the multi-scale prediction model.

[0194] S6: Use the dataset to train a graph neural network model to predict the power of distributed photovoltaic power station clusters.

[0195] S6.1: Divide the dataset obtained in S1 into a training set and a test set according to the time series.

[0196] S6.2: Use the test set to evaluate the trained graph neural network model. The evaluation metric is the root mean square error (RMSE) to determine the model's prediction accuracy.

[0197] S6.3: If the prediction accuracy fails to meet the preset standard, adjust the model parameters based on the evaluation feedback and retrain the graph neural network model until the prediction accuracy meets the requirements.

[0198] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0199] Based on the same inventive concept, this application also provides a distributed photovoltaic power prediction device for implementing the distributed photovoltaic power prediction method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the distributed photovoltaic power prediction device provided below can be found in the limitations of the distributed photovoltaic power prediction method described above, and will not be repeated here.

[0200] In one exemplary embodiment, such as Figure 4 As shown, a distributed photovoltaic power prediction device is provided, including: an acquisition module 402, a construction module 404, a processing module 406, and an output module 408, wherein:

[0201] The acquisition module 402 is used to acquire the output data and meteorological data corresponding to the distributed photovoltaic power station group; the distributed photovoltaic power station group includes multiple photovoltaic power stations;

[0202] Module 404 is used to construct an original node matrix based on power output data and meteorological data; the original node matrix is ​​used to represent the multi-dimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multi-dimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0203] Processing module 406 is used to perform node embedding processing on the original node matrix to obtain a node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0204] Output module 408 is used to input the node embedding matrix into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0205] In one embodiment, the output module 408 is used to input a node embedding matrix into the trained graph neural network model, which includes a graph learning layer, a temporal feature extraction layer, a spatial feature extraction layer, and an output layer connected in sequence, to output the power prediction results of the distributed photovoltaic power station group, including:

[0206] The node embedding matrix is ​​input into the graph learning layer to output an asymmetric adjacency matrix;

[0207] The asymmetric adjacency matrix is ​​input into the temporal feature extraction layer to output the temporal feature matrix;

[0208] The asymmetric adjacency matrix is ​​input into the spatial feature extraction layer to output the spatial feature matrix;

[0209] The temporal and spatial feature matrices are input into the output layer to output the power prediction results of the distributed photovoltaic power station group.

[0210] In one embodiment, the graph neural network model includes a learning module 4081 for inputting a node embedding matrix into a graph learning layer to output an asymmetric adjacency matrix, including:

[0211] Based on the node embedding matrix, the first intermediate matrix and the second intermediate matrix are obtained respectively through learnable parameters;

[0212] Calculate the first product of the transpose of the first intermediate matrix and the second intermediate matrix, and the second product of the transpose of the second intermediate matrix and the first intermediate matrix. Subtract the second product from the first product to obtain the difference matrix.

[0213] The difference matrix is ​​processed by an activation function to obtain an asymmetric adjacency matrix.

[0214] In one embodiment, the graph neural network model includes a dynamic recurrent network module 4082 for inputting an asymmetric adjacency matrix to a temporal feature extraction layer to output a temporal feature matrix, including:

[0215] Multi-scale temporal features are extracted from the input temporal sequence based on the node embedding matrix. The multi-scale temporal features obtained by convolution with different dilation rates are concatenated to obtain a multi-scale feature matrix.

[0216] The multi-scale feature matrix is ​​averaged and pooled to obtain the global temporal feature matrix.

[0217] Based on the global temporal feature matrix and the asymmetric adjacency matrix, a dynamic adjacency matrix is ​​generated through dynamic convolution and activation functions;

[0218] The hidden layer state is updated based on the multi-scale feature matrix, the asymmetric adjacency matrix, the dynamic adjacency matrix, and the hidden state of the previous time step.

[0219] After performing residual connections on the hidden layer states, the temporal feature matrix is ​​output.

[0220] In one embodiment, the graph neural network model includes a directed graph convolutional module 4083, and the spatial feature extraction layer consists of two hybrid skip propagation layers; inputting an asymmetric adjacency matrix to the spatial feature extraction layer to output a spatial feature matrix includes:

[0221] Based on the asymmetric adjacency matrix, degree normalization is performed, and the spatial features after propagation are obtained through the first hybrid hop propagation layer.

[0222] Based on spatial features combined with a learnable weight matrix, the spatial features after propagation are filtered through a second hybrid jump propagation layer to obtain a filtered spatial feature matrix.

[0223] In one embodiment, the output module 408 is further configured to input the time-series feature matrix and the spatial feature matrix to the output layer to output the power prediction results of the distributed photovoltaic power station group, including:

[0224] The temporal feature matrix and the spatial feature matrix are added together, and the result is normalized.

[0225] The processing results are input into the fully connected layer, and the power prediction results of the distributed photovoltaic power station group are output.

[0226] In one embodiment, the acquisition module 402 is further configured to:

[0227] Obtain historical power output data and historical meteorological data corresponding to distributed photovoltaic power station groups;

[0228] Based on historical power output data and historical meteorological data, a training set and a test set are constructed;

[0229] The pre-built graph neural network model is trained using the training set to obtain the initial graph neural network model;

[0230] The initial graph neural network model was evaluated using a test set to obtain the evaluation results;

[0231] If the evaluation results do not meet the preset standards, the parameters of the initial graph neural network model are adjusted according to the evaluation results, and the prediction model is retrained using the training set until the preset standards are met.

[0232] Each module in the aforementioned distributed photovoltaic power prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0233] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores photovoltaic-related data (including but not limited to node embedding data, time-series characteristic data, dynamic adjacency matrix parameters, and historical operating data). The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a distributed photovoltaic power prediction method.

[0234] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0235] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0236] Acquire power output data and meteorological data corresponding to a distributed photovoltaic power station cluster; a distributed photovoltaic power station cluster includes multiple photovoltaic power stations.

[0237] An original node matrix is ​​constructed based on power output data and meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0238] Perform node embedding processing on the original node matrix to obtain the node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0239] The node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0240] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0241] The node embedding matrix is ​​input into the graph learning layer to output an asymmetric adjacency matrix;

[0242] The asymmetric adjacency matrix is ​​input into the temporal feature extraction layer to output the temporal feature matrix;

[0243] The asymmetric adjacency matrix is ​​input into the spatial feature extraction layer to output the spatial feature matrix;

[0244] The temporal and spatial feature matrices are input into the output layer to output the power prediction results of the distributed photovoltaic power station group.

[0245] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0246] Based on the node embedding matrix, the first intermediate matrix and the second intermediate matrix are obtained respectively through learnable parameters;

[0247] Calculate the first product of the transpose of the first intermediate matrix and the second intermediate matrix, and the second product of the transpose of the second intermediate matrix and the first intermediate matrix. Subtract the second product from the first product to obtain the difference matrix.

[0248] The difference matrix is ​​processed by an activation function to obtain an asymmetric adjacency matrix.

[0249] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0250] Multi-scale temporal features are extracted from the input temporal sequence based on the node embedding matrix. The multi-scale temporal features obtained by convolution with different dilation rates are concatenated to obtain a multi-scale feature matrix.

[0251] The multi-scale feature matrix is ​​averaged and pooled to obtain the global temporal feature matrix.

[0252] Based on the global temporal feature matrix and the asymmetric adjacency matrix, a dynamic adjacency matrix is ​​generated through dynamic convolution and activation functions;

[0253] The hidden layer state is updated based on the multi-scale feature matrix, the asymmetric adjacency matrix, the dynamic adjacency matrix, and the hidden state of the previous time step.

[0254] After performing residual connections on the hidden layer states, the temporal feature matrix is ​​output.

[0255] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0256] Based on the asymmetric adjacency matrix, degree normalization is performed, and the spatial features after propagation are obtained through the first hybrid hop propagation layer.

[0257] Based on spatial features combined with a learnable weight matrix, the spatial features after propagation are filtered through a second hybrid jump propagation layer to obtain a filtered spatial feature matrix.

[0258] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0259] The temporal feature matrix and the spatial feature matrix are added together, and the result is normalized.

[0260] The processing results are input into the fully connected layer, and the power prediction results of the distributed photovoltaic power station group are output.

[0261] In one embodiment, the method further includes:

[0262] Obtain historical power output data and historical meteorological data corresponding to distributed photovoltaic power station groups;

[0263] Based on historical power output data and historical meteorological data, a training set and a test set are constructed;

[0264] The pre-built graph neural network model is trained using the training set to obtain the initial graph neural network model;

[0265] The initial graph neural network model was evaluated using a test set to obtain the evaluation results;

[0266] If the evaluation results do not meet the preset standards, the parameters of the initial graph neural network model are adjusted according to the evaluation results, and the prediction model is retrained using the training set until the preset standards are met.

[0267] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0268] Acquire power output data and meteorological data corresponding to a distributed photovoltaic power station cluster; a distributed photovoltaic power station cluster includes multiple photovoltaic power stations.

[0269] An original node matrix is ​​constructed based on power output data and meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0270] Perform node embedding processing on the original node matrix to obtain the node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0271] The node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0272] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0273] Acquire power output data and meteorological data corresponding to a distributed photovoltaic power station cluster; a distributed photovoltaic power station cluster includes multiple photovoltaic power stations.

[0274] An original node matrix is ​​constructed based on power output data and meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions;

[0275] Perform node embedding processing on the original node matrix to obtain the node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix.

[0276] The node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

[0277] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0278] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0279] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0280] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting distributed photovoltaic power, characterized in that, The method includes: Acquire power output data and meteorological data corresponding to the distributed photovoltaic power station group; the distributed photovoltaic power station group includes multiple photovoltaic power stations. An original node matrix is ​​constructed based on the power output data and the meteorological data; the original node matrix is ​​used to represent the multidimensional characteristics of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional characteristics are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions; The original node matrix is ​​subjected to node embedding processing to obtain a node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix. The node embedding matrix is ​​input into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

2. The method according to claim 1, characterized in that, The graph neural network model includes a graph learning layer, a temporal feature extraction layer, a spatial feature extraction layer, and an output layer connected in sequence; the step of inputting the node embedding matrix into the trained graph neural network model to output the power prediction result of the distributed photovoltaic power station group includes: The node embedding matrix is ​​input into the graph learning layer to output an asymmetric adjacency matrix; The asymmetric adjacency matrix is ​​input into the temporal feature extraction layer to output the temporal feature matrix; The asymmetric adjacency matrix is ​​input into the spatial feature extraction layer to output the spatial feature matrix; The temporal feature matrix and the spatial feature matrix are input to the output layer to output the power prediction results of the distributed photovoltaic power station group.

3. The method according to claim 2, characterized in that, The step of inputting the node embedding matrix into the graph learning layer to output an asymmetric adjacency matrix includes: Based on the node embedding matrix, the first intermediate matrix and the second intermediate matrix are obtained respectively by using learnable parameters; Calculate the first product of the transpose of the first intermediate matrix and the second intermediate matrix, and the second product of the transpose of the second intermediate matrix and the first intermediate matrix. Subtract the second product from the first product to obtain the difference matrix. The difference matrix is ​​processed by an activation function to obtain an asymmetric adjacency matrix.

4. The method according to claim 2, characterized in that, The step of inputting the asymmetric adjacency matrix into the temporal feature extraction layer to output a temporal feature matrix includes: Based on the node embedding matrix, multi-scale temporal features are extracted from the input temporal sequence. The multi-scale temporal features obtained by convolution with different dilation rates are concatenated to obtain a multi-scale feature matrix. The multi-scale feature matrix is ​​subjected to average pooling to obtain the global temporal feature matrix; Based on the global temporal feature matrix and the asymmetric adjacency matrix, a dynamic adjacency matrix is ​​generated through dynamic convolution and activation functions; The hidden layer state is updated based on the multi-scale feature matrix, the asymmetric adjacency matrix, the dynamic adjacency matrix, and the hidden state at the previous time step. After performing residual connections on the hidden layer states, a temporal feature matrix is ​​output.

5. The method according to claim 2, characterized in that, The spatial feature extraction layer consists of two hybrid skip propagation layers; the step of inputting the asymmetric adjacency matrix into the spatial feature extraction layer to output the spatial feature matrix includes: Based on the asymmetric adjacency matrix, degree normalization is performed, and the transmitted spatial features are obtained through the first hybrid hop propagation layer. Based on the spatial features combined with the learnable weight matrix, the transmitted spatial features are filtered through the second hybrid jump propagation layer to obtain the filtered spatial feature matrix.

6. The method according to claim 2, characterized in that, The step of inputting the temporal feature matrix and the spatial feature matrix into the output layer to output the power prediction results of the distributed photovoltaic power station group includes: The temporal feature matrix is ​​added to the spatial feature matrix, and the result is normalized. The processing results are input into the fully connected layer, and the power prediction results of the distributed photovoltaic power station group are output.

7. The method according to claim 1, characterized in that, The method further includes: Obtain the historical power output data and historical meteorological data corresponding to the distributed photovoltaic power station group; Based on the historical power output data and the historical meteorological data, a training set and a test set are constructed; The pre-built graph neural network model is trained using the training set to obtain an initial graph neural network model; The initial graph neural network model is evaluated using the test set to obtain evaluation results; If the evaluation result does not meet the preset standard, the parameters of the initial graph neural network model are adjusted according to the evaluation result, and the prediction model is retrained using the training set until the preset standard is met.

8. A training model device for distributed photovoltaic power generation prediction, characterized in that, The device includes: The acquisition module is used to acquire power output data and meteorological data corresponding to the distributed photovoltaic power station group; the distributed photovoltaic power station group includes multiple photovoltaic power stations; A construction module is used to construct an original node matrix based on the power output data and the meteorological data; the original node matrix is ​​used to represent the multidimensional features of each photovoltaic power station in the distributed photovoltaic power station group; the multidimensional features are used to represent the power output of the photovoltaic power station and the corresponding meteorological conditions; The processing module is used to perform node embedding processing on the original node matrix to obtain a node embedding matrix; the dimension of the node embedding matrix is ​​smaller than that of the original node matrix. The output module is used to input the node embedding matrix into the trained graph neural network model to output the power prediction results of the distributed photovoltaic power station group.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images