Distributed photovoltaic power prediction method and system based on graph neural network

Abstract

The present invention belongs to the field of data processing technology. A distributed photovoltaic power prediction method and system based on a graph neural network is provided. According to the irradiation data of distributed photovoltaics in a set period before the current moment, spatial Kriging interpolation is used to initialize the irradiation data, and the initialized irradiation data is enhanced based on a graph diffusion convolutional network to obtain enhanced irradiation data; the graph convolution model is coupled with the spatial Transformer module using a residual structure to capture spatial features; the enhanced irradiation data and the spatial features are used as inputs, and a trend gate recursive unit network is used to obtain the distributed photovoltaic power prediction results for a set period after the current moment. The prediction method proposed in the present invention improves the accuracy of prediction from three aspects: data, model, and parameters, and provides a practical demand-based solution for high-precision short-term prediction of distributed photovoltaic power stations.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a distributed photovoltaic power prediction method and system based on graph neural network. Background Art

[0002] The statements in this section merely provide background art related to the present invention and do not necessarily constitute prior art.

[0003] Strengthening power forecasting for intermittent and non-dispatchable energy sources is a key factor in increasing the penetration of renewable energy in the power grid. Existing centralized photovoltaic forecasting methods can be roughly divided into two categories: direct methods and data-driven methods, and the research and application of these methods have reached a fairly mature level. In contrast, distributed photovoltaic power generation forecasting specifically targets multi-location or regional photovoltaic power generation forecasting of distributed photovoltaic power stations. In recent years, research on distributed photovoltaic power generation forecasting has often been inspired by centralized forecasting methods; however, there are still significant differences in data conditions, predictability of objects, and applicable forecasting models. Specifically, distributed photovoltaic power generation forecasting faces challenges such as missing irradiation data and insufficient exploration of spatial correlation by forecasting models, making it difficult to improve the accuracy of distributed photovoltaic forecasting.

[0004] Distributed photovoltaic power stations have the characteristics of small individual capacity, scattered installation locations, and large numbers, and often show a trend of local high penetration and high density. The cost of deploying meteorological instruments for small and scattered distributed photovoltaic power stations is too high. Therefore, most distributed photovoltaic power stations rely on numerical weather forecasts (NWP) from nearby central photovoltaic power stations, which often lack high-resolution key meteorological data for the entire region. Therefore, it is a key research challenge to effectively obtain high-resolution NWP data such as solar irradiance by using public data resources and using statistical methods with strong nonlinear fitting capabilities. In recent years, Kriging interpolation and Gaussian process regression have become prominent techniques for geographic interpolation of meteorological data. These methods provide fine-grained and high-resolution spatiotemporal data interpolation by considering temporal trends and spatial correlations. However, they have limitations in computational cost and processing large-scale data sets, resulting in suboptimal interpolation performance and lack of economic feasibility.

[0005] Deep neural networks have become the main means of short-term photovoltaic power generation forecasting. For example, convolutional neural networks and long short-term memory networks are used to enhance the model's fitting ability. Attention models are often integrated into various neural network units to improve the model's generalization ability and correlation mining ability. However, the above methods only focus on the temporal autocorrelation of photovoltaic sequences, while ignoring the spatial correlation between adjacent power station sequences in distributed photovoltaic power generation forecasting scenarios. In recent years, scholars have turned their attention to using the spatiotemporal correlation between adjacent sites to improve the accuracy of short-term photovoltaic forecasting. The forecasting method takes into account information from surrounding photovoltaic power stations, but its focus on the spatial dimension is mostly limited to data preprocessing or the selection of input features, resulting in generally low forecast accuracy. Summary of the invention

[0006] In order to address the shortcomings of the prior art, the present invention provides a distributed photovoltaic power prediction method and system based on graph neural network, which improves the accuracy of prediction from three aspects: data, model and parameters, and provides a practical demand-based solution for high-precision short-term prediction of distributed photovoltaic power stations.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a distributed photovoltaic power prediction method based on graph neural network.

[0009] A distributed photovoltaic power prediction method based on graph neural network includes the following processes:

[0010] According to the irradiation data of distributed photovoltaics in the set period before the current moment, the irradiation data is initialized by using spatial Kriging interpolation, and the initialized irradiation data is enhanced based on the graph diffusion convolutional network to obtain enhanced irradiation data;

[0011] According to the irradiation data of distributed photovoltaics in the set period before the current moment, the graph convolution model is coupled with the spatial Transformer module using the residual structure to capture spatial features;

[0012] The enhanced irradiation data and the spatial features are used as inputs, and a trend gate recursive unit network is used to obtain a distributed photovoltaic power prediction result for a set period after the current moment.

[0013] As a further limitation of the first aspect of the present invention, a hyperparameter optimization feedback improvement strategy suitable for a swarm intelligence algorithm is adopted, and the particle population density is quantified based on an information entropy function, so as to enhance the optimization effect of the algorithm by adding disturbances in a timely manner.

[0014] In a second aspect, the present invention provides a distributed photovoltaic power prediction system based on graph neural network.

[0015] A distributed photovoltaic power prediction system based on graph neural network, comprising:

[0016] The data enhancement unit is configured to: initialize the irradiation data using spatial Kriging interpolation according to the irradiation data of the distributed photovoltaic in the set period before the current moment, and enhance the initialized irradiation data based on the graph diffusion convolutional network to obtain enhanced irradiation data;

[0017] The feature extraction unit is configured to: according to the irradiation data of distributed photovoltaics in a set period before the current moment, use the residual structure to couple the graph convolution model with the spatial Transformer module to capture the spatial features;

[0018] The power prediction unit is configured to: take the enhanced irradiation data and the spatial features as input, adopt a trend gate recursive unit network, and obtain a distributed photovoltaic power prediction result for a set period after the current moment.

[0019] As a further limitation of the first aspect of the present invention, it also includes a hyperparameter optimization unit, which is configured to: adopt a hyperparameter optimization feedback improvement strategy suitable for a swarm intelligence algorithm, quantify the particle population density based on an information entropy function, and add disturbances in a timely manner to enhance the optimization effect of the algorithm.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. This invention innovatively proposes a meteorological data enhancement method based on Kriging-graph diffusion convolution, which achieves low-cost acquisition of high-resolution irradiation data and constructs a data foundation for subsequent distributed photovoltaic power prediction accuracy improvement.

[0022] 2. This invention innovatively proposes a novel spatiotemporal graph transformer network model, innovatively applies the spatiotemporal graph network to distributed photovoltaic power prediction, and uses graph convolution and attention mechanism to effectively extract spatial information between adjacent photovoltaic power stations, thus achieving high-precision short-term power prediction for multiple photovoltaic power stations.

[0023] 3. The present invention innovatively proposes a feedback improvement strategy based on information entropy function to quantify group diversity suitable for swarm intelligence algorithm, which effectively evaluates the current state of the population to guide the timely addition of disturbances, thereby improving optimization performance and prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0025] Figure 1A schematic diagram of the process of the Kriging-graph diffusion convolution data enhancement method provided in Example 1 of the present invention;

[0026] Figure 2 A multi-module structure diagram of a network model of a time-space graph converter provided in Example 1 of the present invention;

[0027] Figure 3 A schematic diagram of a trend gate recursive unit network provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0028] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0030] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0031] Embodiment 1:

[0032] This implementation proposes a distributed photovoltaic power prediction method based on graph neural network. First, a Kriging-graph diffusion convolution spatial data enhancement method is designed. The spatial Kriging interpolation method is used to initialize the irradiation data. Then, data enhancement is implemented based on the graph diffusion convolution network to infer the data of unsampled locations in the graph structure to achieve high-resolution coverage of key meteorological irradiation data; then, a new spatiotemporal graph network model is proposed for distributed photovoltaic power prediction. The network model uses a residual structure to couple the graph convolution model with the Transformer model to capture spatial features, and uses an improved trend gate recurrent unit network to extract temporal autocorrelation, effectively solving the prediction lag problem; finally, a hyperparameter optimization feedback improvement strategy suitable for swarm intelligence algorithm is proposed. The particle population density is quantified based on the information entropy function, so as to enhance the optimization effect of the algorithm by adding disturbances in a timely manner.

[0033] Specifically, the distributed photovoltaic power prediction method includes:

[0034] S1: Kriging-Graph Diffusion Convolution Data Augmentation.

[0035] Figure 1The complete Kriging-graph diffusion convolution irradiance data enhancement method is outlined. First, Kriging interpolation is used to initialize the graph signal at the desired target resolution, that is, to initialize the meteorological irradiance data of the unsampled stations. At each node of the target resolution, based on the meteorological irradiance data of all known locations obtained from the centralized station (complete station), in order to avoid the over-smoothing effect in the graph diffusion convolution data enhancement process, the Kriging method is used to use the distance between stations to predict the value of the unsampled location.

[0036] S1.1: Kriging initialization.

[0037] In actual distributed photovoltaic prediction scenarios, the available irradiation data is usually limited, often with only spatial geographic information but lacking time delay and complete auxiliary meteorological data. In order to overcome this limitation, the Kriging method is used to obtain the initial distributed site irradiation value by combining geographic location and distance information, where the unbiased weighted average of neighboring sites is used to estimate the value of the unsampled site:

[0038] (1);

[0039] In the formula, and It is a data point With unobserved points Between and data points and The semivariogram function between and Respectively represent i and j The weight of each data point, N represents the number of data points.

[0040] S1.2: Graph Dilated Convolution Data Augmentation.

[0041] This implementation introduces a double-layer graph diffusion convolution combined with a residual network, and uses a random sampling subgraph method for model training. Its core operation is to randomly sample a subset of nodes to obtain the feature matrix X of the sampled data and construct the corresponding adjacency matrix W. The mathematical expression of the network structure of the double-layer graph diffusion convolution is:

[0042] (2);

[0043] (3);

[0044] In the formula, and They are the forward transfer matrix and the backward transfer matrix, which can be expressed as the normalized matrix of the weight matrix , for an undirected graph, the adjacency matrix is ​​symmetric, = , in the distributed photovoltaic scenario, the model is an undirected graph. K is the total number of diffuse convolution orders, is the diffusion convolution order, and the convolution process in the diffusion convolution is approximated by Chebyshev polynomials. , and is the learning parameter of the first layer, which controls the transformation of each node. represents the transformation part, Represented as a residual operation, W is the adjacency matrix, is the output of the second layer, is the output of the first layer.

[0045] S1.3: Gaussian method to construct edge weights.

[0046] For the adjacency matrix W ,The irradiance of neighboring PV plants is affected by geographical location, cloud cover and wind speed. The Euclidean distance between PV sites and the statistical correlation between auxiliary variables are used to assign edge weights in the graph. Considering that the information may be combined with the observed correlation characteristics and distance characteristics, if time-varying meteorological auxiliary data are available, two or more time-varying variables and any combination of static variables and auxiliary variables can be used to define edge weights. The edge weights are calculated using a Gaussian kernel and are expressed in mathematical form as follows:

[0047] (4);

[0048] In the formula, and Represents adjacent nodes and The difference between the static variable and the time-varying variable and the auxiliary variable, and The relative value of controls the relative importance of static variables and time-varying variables.

[0049] S2: Spatiotemporal graph transformer network distributed photovoltaic prediction model.

[0050] like Figure 2 As shown in the figure, it outlines the architecture of the spatiotemporal graph transformer network model (STGTN). The STGTN model consists of a spatial network and a temporal network. The spatial network and the temporal network are tightly coupled. The spatial network uses a graph convolutional model (GCN) and a spatial transformer model (Transformer) to extract spatial features, and the temporal network uses an enhanced trend gate recurrent unit network (TGRU) to process the time dimension.

[0051] S2.1: GCN model.

[0052] Unlike traditional convolutional neural networks (CNNs) which are limited to processing Euclidean space, graph neural networks (GCNs) can more flexibly process the graph structure formed by distributed photovoltaic topologies, thereby capturing spatial dependencies more accurately.

[0053] Given an adjacency matrix A and the feature matrix X , the GCN model constructs a filter in the Fourier frequency domain, which operates on the nodes of the graph, captures the spatial features between nodes through the first-order neighborhood of the nodes, and then constructs a two-layer GCN model by superposition, which is expressed in mathematical form as follows:

[0054] (5);

[0055] In the formula, represents the adjacency matrix with self-connection, is the identity matrix, for The degree matrix of is the output of the second layer of the graph convolutional model, and is the weight parameter of the first and second layers, and the nonlinear activation function is used and Express.

[0056] Specifically, the input is the original data, which is input into the spatial Transformer model after graph convolution, ReLU layer and layer normalization of the GCN model.

[0057] S2.2: Spatial Transformer model.

[0058] In order to extract spatial information more effectively, inspired by the successful application of spatial transformers in long-sequence time prediction, a spatial Transformer module is designed. The spatial Transformer module can be regarded as the transposition of the traditional Transformer input data, treating each time series as an independent feature vector, and using a self-attention mechanism to capture the key information in the sequence. At the same time, the spatial dependency between features is also modeled. The spatial Transformer module includes two key components: a scaled dot product attention mechanism and a position encoding. The position encoding is usually added to the input embedding to model the position index sequence, which includes:

[0059] (6);

[0060] In the formula, n is the position index, f is the dimension index, dmodelis the feature hidden dimension, and each time series is regarded as an independent feature vector. The scaling formula of the multi-head self-attention network of the spatial transformer is:

[0061] (7);

[0062] (8);

[0063] In the formula, query Q, key K, and value V are the three basic elements in the self-attention mechanism, which are obtained by non-shared linear transformation of the original input. are the weight matrices of query Q, key K, and value V respectively, is the input feature matrix, , among which N represent N A position index sequence, T represents time; , which expresses the attention of different time series, represents the field of real numbers, Represents the vector dimension in Transformer, is a scaling factor equal to the hidden dimension of the model, , h represent h To stabilize the training process, we use multi-head attention and h groups of learned projections, each of which implements a different attention function. is the normalized exponential function.

[0064] Specifically, the spatial Transformer model includes a multi-head attention layer and a feedforward network, and the output of the feedforward network is used as the spatial feature.

[0065] S2.3: Enhanced trend-gated recurrent unit network.

[0066] Although GRUs is a mature time series forecasting technology, it faces inherent challenges in accurately capturing the trend characteristics of non-stationary series. This limitation leads to significant lag in the predicted photovoltaic power generation data, affecting the accuracy of the forecasting task. The TGRU model introduces a trend gate, integrates the network's sensitivity to differential data, and captures trend information more effectively, thereby improving its performance in predicting non-stationary series by more accurately modeling trend changes.

[0067] The network structure is as follows Figure 3 As shown, its mathematical expression is as follows:

[0068] (9);

[0069] In the formula, Indicates time tThe feature input; It means the hidden layer at time t Output: is the update gate, which controls the degree to which the past hidden state is retained; To reset the gate, adjust the influence of the past hidden state on the current time step; It is the trend gate, which controls the degree of trend information update; is the first-order difference of the input feature, and its value is , represents the trend information between sequences, and the initial state is 0; Represents a new memory cell, indicating a candidate update value; is the sigmoid activation function; ⊙ represents the dot product; , , and Respectively represent the weight matrix; , , and They represent the bias vectors, is a non-linear activation function.

[0070] Specifically, Figure 2 As shown in the figure, the enhanced meteorological irradiance data and spatial features are input into the enhanced trend gate recurrent unit network (TGRU model) for fusion prediction to obtain short-term prediction results.

[0071] S3: Use improved swarm intelligence algorithm to optimize hyperparameters.

[0072] Existing model parameter optimization methods focus more on swarm intelligence algorithms to output interpretable parameters to improve prediction performance. Swarm intelligence algorithms originate from the study of social behavior of biological groups. They form swarm intelligence through interaction mechanisms to solve complex optimization problems, but premature convergence often occurs during the optimization process. Existing research is mostly limited to studying different perturbation strategies, while ignoring the critical timing of introducing perturbations. This oversight may cause the algorithm to fail to converge or the perturbation to fail to effectively produce the desired effect. Therefore, the present invention designs an improved feedback strategy based on the information entropy function, which aims to quantify group diversity and add perturbations in a timely manner to get rid of local optimality. This implementation method mainly adjusts the correlation threshold ( Ts ), hidden layer dimensions ( Hd ), learning rate ( Lr ) and the input window dimensions ( T ) and other hyperparameters for optimization.

[0073] Assume that n The initial state of a group of individuals can be represented by the state set Indicates that Indicates i The calculation of information entropy quantifies the distribution of individual states, indicating the diversity of the population. Calculated using the following formula:

[0074] (10);

[0075] In the formula, is a particle in the population The probability of is a small constant, usually chosen very small to avoid instabilities in logarithmic calculations.

[0076] Combined with theoretical analysis, the calculation of information entropy in a group has the characteristics of not being affected by factors such as individual dimension, search space size, and time. Its value is only related to the size of the group, and is usually positively correlated with the number of individuals in the group. In the middle of the iteration, the information entropy drops "cliff-like", and the corresponding population diversity drops sharply. This decline causes the population to converge prematurely, causing it to fall into a local minimum. The change in information entropy can be used as an effective indicator to evaluate the current state of the group and help determine whether the group is trapped in a local optimum. When using information entropy to determine different states between individuals, the existing errors should be considered. If the distance between individuals is less than the specified error (experienced as 10−6), the individuals are considered to have the same state in the system. During the iteration process, the first-order derivative of the information entropy is calculated by differential method. If the first-order derivative is less than the initial information entropy divided by the size of the group, it is considered that the group has the possibility of falling into a local optimum. After effectively identifying the local state, an adaptive variance perturbation is added to each position. The mathematical expression is as follows:

[0077] (11);

[0078] In the formula, express and The difference, Expressed as the entropy value of the initialized population, express t The population entropy value at the moment, Ht-1 represents the population entropy value at time t-1, N represents the population size, To update the rules, For the current t The state of the moment, express t+1 The state of the moment, are algorithm parameters, The representative role is i The disturbance of an individual.

[0079] Embodiment 2:

[0080] This implementation provides a distributed photovoltaic power prediction system based on a graph neural network, including:

[0081] The data enhancement unit is configured to: initialize the irradiation data using spatial Kriging interpolation according to the irradiation data of the distributed photovoltaic in the set period before the current moment, and enhance the initialized irradiation data based on the graph diffusion convolutional network to obtain enhanced irradiation data;

[0082] The feature extraction unit is configured to: according to the irradiation data of distributed photovoltaics in a set period before the current moment, use the residual structure to couple the graph convolution model with the spatial Transformer module to capture the spatial features;

[0083] The power prediction unit is configured to: take the enhanced irradiation data and the spatial features as input, adopt a trend gate recursive unit network, and obtain a distributed photovoltaic power prediction result for a set period after the current moment.

[0084] Optionally, a hyperparameter optimization unit may be included, which is configured to: adopt a hyperparameter optimization feedback improvement strategy suitable for swarm intelligence algorithms, quantify the particle population density based on the information entropy function, and add disturbances in a timely manner to enhance the optimization effect of the algorithm.

[0085] It can be understood that the above-mentioned units can be separately or completely combined into one or several other units to constitute, or some of the units can be further divided into multiple functionally smaller units to constitute, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above-mentioned units are divided based on logical functions. In practical applications, the function of one unit can also be implemented by multiple units, or the functions of multiple units can be implemented by one unit. In other embodiments of the present application, the distributed photovoltaic power prediction system of the graph neural network can also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented by the collaboration of multiple units.

[0086] According to another embodiment of the present application, the system described in this embodiment can be constructed, and the distributed photovoltaic power prediction method of the graph neural network of the embodiment of the present application can be implemented by running a computer program (including program code) capable of executing the steps involved in the corresponding method described in Example 1 on a general computing device such as a computer including processing elements and storage elements such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM). The computer program can be recorded on, for example, a computer-readable recording medium, and loaded into the above-mentioned computing device through the computer-readable recording medium and run therein.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A distributed photovoltaic power prediction method based on graph neural network, characterized in that: The process includes: According to the irradiation data of distributed photovoltaics in the set period before the current moment, the irradiation data is initialized by using spatial Kriging interpolation, and the initialized irradiation data is enhanced based on the graph diffusion convolutional network to obtain enhanced irradiation data; The Kriging method is used to obtain the initial distributed site irradiance values, including: using the unbiased weighted average of neighboring sites to estimate the value of the unsampled site: ； In the formula, and It is a data point With unobserved points Between and data points and The semivariogram function between and Represent the weights of the i-th and j-th data points respectively, and N represents the number of data points; The graph diffusion convolutional network includes: ； ； in, and are the forward transfer matrix and the backward transfer matrix respectively, K is the total number of diffuse convolution orders, is the diffuse convolution order, , W is the adjacency matrix; the convolution process in the graph diffusion convolution is approximated by Chebyshev polynomials. , X refers to or , and is the learning parameter of the first layer, which controls the transformation of each node. represents the transformation part, Represented as a residual operation, W is the adjacency matrix, is the output of the second layer, is the output of the first layer; According to the irradiation data of distributed photovoltaics in the set period before the current moment, the graph convolution model is coupled with the spatial Transformer module using the residual structure to capture spatial features; The graph convolution model includes: Given the adjacency matrix A and the input feature matrix X, a filter is constructed in the Fourier frequency domain. The nodes of the graph are operated through the filter. The spatial features between nodes are captured through the first-order neighborhood of the nodes. The graph convolution model is constructed by superimposing the spatial features of each node. The output of the second layer of the graph convolution model is for: ； In the formula, represents the adjacency matrix with self-connection, is the identity matrix, for The degree matrix of is the output of the second layer of the graph convolutional model, and is the weight parameter of the first and second layers, and the nonlinear activation function is used and Express; The spatial Transformer module includes: Each time series is regarded as an independent feature vector, and a self-attention mechanism is used to capture the key information in the sequence. At the same time, the spatial dependency between features is modeled, including a scaled dot product attention mechanism and position encoding. The position encoding is added to the input embedding to model the position index sequence. The position index sequence includes: ； In the formula, n is the position index, f is the dimension index, dmodel is the feature hidden dimension, and each time series is regarded as an independent feature vector. The scaling formula of the self-attention network is: ； ； In the formula, query Q, key K, and value V are the three basic elements in the self-attention mechanism, which are obtained by non-shared linear transformation of the original input. are the weight matrices of query Q, key K, and value V respectively, is the scale factor, , h represents the number of projections learned, using multi-head attention and h The projections learned by the groups are each realized with different attention functions; the enhanced irradiation data and the spatial features are used as input, and a trend gate recursive unit network is used to obtain the distributed photovoltaic power prediction result for a set period after the current moment; The trend gate recurrent unit network comprises: ； In the formula, Indicates time t The feature input; It means the hidden layer at time t Output: is the update gate, which controls the degree to which the past hidden state is retained; To reset the gate, adjust the influence of the past hidden state on the current time step; It is the trend gate, which controls the degree of trend information update; is the first-order difference of the input feature, and its value is , represents the trend information between sequences, and the initial state is 0; Represents a new memory cell, indicating a candidate update value; is the sigmoid activation function; ⊙ represents the dot product; , , and Respectively represent the weight matrix; , , and They represent the bias vectors, is a non-linear activation function.

2. The distributed photovoltaic power prediction method based on graph neural network according to claim 1, characterized in that: Adjacency Matrix W The edge weight ,include: ； In the formula, and Represents adjacent nodes and The difference between the static variable and the time-varying variable and the auxiliary variable, and are the weight coefficients of static variables and time-varying variables respectively.

3. The distributed photovoltaic power prediction method based on graph neural network according to any one of claims 1 to 2, characterized in that: The hyperparameter optimization feedback improvement strategy is adopted to quantify the particle population density based on the information entropy function, and the hyperparameter optimization effect of the graph convolution model is enhanced by adding disturbances in a timely manner, including: Assume that the initial state of a population consisting of n individuals can be represented by the state set ,in Indicates i The information entropy function of a population is the state of each individual. for: ； In the formula, is a particle in the population The probability of is a constant; If the distance between individuals is less than the specified error, the individuals are considered to have the same state in the system. During the iteration process, the first-order derivative of the information entropy is calculated using the differential method. If the derivative is less than the initial information entropy divided by the size of the group, it is considered that the group has the possibility of falling into a local optimum. Adaptive perturbations are added, including: ； In the formula, express and The difference, Expressed as the entropy value of the initialized population, express t The population entropy value at the moment, Ht-1 represents the population entropy value at time t-1, N represents the population size, is the algorithm-specific update rule, For the current t The state of the moment, express t+1 The state of the moment, are algorithm parameters, The representative role is i The disturbance of an individual.

4. A distributed photovoltaic power prediction system based on a graph neural network for implementing the distributed photovoltaic power prediction method based on a graph neural network as described in any one of claims 1 to 3, characterized in that: include: The data enhancement unit is configured to: initialize the irradiation data using spatial Kriging interpolation according to the irradiation data of the distributed photovoltaic in the set period before the current moment, and enhance the initialized irradiation data based on the graph diffusion convolution network to obtain enhanced irradiation data; The feature extraction unit is configured to: according to the irradiation data of distributed photovoltaics in a set period before the current moment, use the residual structure to couple the graph convolution model with the spatial Transformer module to capture the spatial features; The power prediction unit is configured to: take the enhanced irradiation data and the spatial features as input, adopt a trend gate recursive unit network, and obtain a distributed photovoltaic power prediction result for a set period after the current moment.

5. The distributed photovoltaic power prediction system based on graph neural network according to claim 4, characterized in that: It also includes a hyperparameter optimization unit, which is configured to: adopt a hyperparameter optimization feedback improvement strategy, quantify the particle population density based on the information entropy function, and enhance the hyperparameter optimization effect of the graph convolution model by adding disturbances in a timely manner, including: Assume that the initial state of a population consisting of n individuals can be represented by the state set ,in Indicates i The information entropy function of a population is the state of each individual. for: ； In the formula, is a particle in the population The probability of n is the sum of the current states, is a constant; If the distance between individuals is less than the specified error, the individuals are considered to have the same state in the system. During the iteration process, the first-order derivative of entropy is calculated using the differential method. If the derivative is less than the initial entropy divided by the size of the group, the group is considered to have the possibility of falling into a local optimum. Adaptive perturbations are added to each position, including: ； In the formula, express and The difference, Expressed as the entropy value of the initialized population, express t The population entropy value at the moment, H t-1 represents the population entropy value at time t-1, N represents the population size, is the algorithm-specific update rule, For the current t The state of the moment, express t+1 The state of the moment, are algorithm parameters, The representative role is i The disturbance of an individual.

Images