Photovoltaic cluster power prediction method based on step-by-step spatial feature clustering and improved graph attention network

Abstract

The invention discloses a photovoltaic cluster power prediction method based on step-by-step spatial feature clustering and an improved graph attention network, and relates to the photovoltaic field, and the method comprises the following steps: constructing a distributed photovoltaic cluster data set, the method comprises the following steps: performing primary clustering division by taking the physical characteristics of a photovoltaic module as characteristics to be input into an affinity propagation AP algorithm, then performing secondary clustering division by taking a solar altitude angle sequence as characteristics to be input into the AP algorithm, and finally dividing a photovoltaic cluster into a plurality of sub-clusters; for each photovoltaic sub-cluster, sorting and merging historical power and historical meteorological time sequence data, and inputting the historical power and historical meteorological time sequence data into a GAT-Encoder-Decoder deep learning model for training; reasoning and outputting a day-ahead power prediction result of each power station in the sub-cluster; and accumulating the prediction results of all the power stations to obtain a power prediction result of the whole photovoltaic cluster. According to the invention, clustering calculation is carried out through step-by-step spatial features so as to obtain a sub-cluster division result which can better reflect the spatial feature state of the photovoltaic power station, and the correlation of the output of the photovoltaic power station in the sub-cluster is improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaics, specifically a photovoltaic cluster power prediction method based on hierarchical spatial feature clustering and an improved graph attention network. Background Technology

[0002] With the widespread use of fossil fuels such as oil and natural gas, the greenhouse effect and environmental degradation are spreading globally, and the high dependence on fossil fuels has led to a global energy crisis. By the end of 2024, China's cumulative grid-connected distributed photovoltaic (PV) capacity had reached 374.78 GW. In the future, an even higher proportion of distributed PV power plants will be connected to the grid. However, due to the lack of professional meteorological monitoring equipment for distributed PV power plant clusters and the lack of sufficient historical data for the large number of newly built distributed PV power plants each year, it is difficult to establish high-precision prediction models. For new energy service companies and local power grids, the rapid development of distributed PV has significantly increased the difficulty of power system supply and demand balance scheduling and management. Therefore, there is an urgent need to construct a high-precision method for predicting the power output of distributed PV clusters. Distributed PV cluster power prediction refers to the process of predicting the future power output of multiple geographically dispersed PV power plants or household PV systems connected to different nodes through data modeling and algorithmic inference. Its core objective is to achieve accurate prediction of the group's PV power under complex spatial distribution and uncertain meteorological conditions, in order to serve grid dispatch and new energy utilization.

[0003] Research on power prediction methods for distributed photovoltaic (PV) clusters has significant application value in contemporary society. Daily power prediction for large-scale distributed PV power plants helps to rationally address the threats to the power grid posed by the randomness and uncertainty of PV power generation, thereby improving grid stability. It also lays a solid foundation for further integration of distributed PV clusters into the grid, contributing to a more effective response to the challenges posed by the randomness and uncertainty of PV power generation.

[0004] Currently, there are many related technologies in China for distributed photovoltaic (PV) power prediction. Patent CN202511087272.4 effectively handles complex temporal features for distributed PV power prediction through multi-scale processing and result fusion, without relying on high-precision meteorological observations or high-cost external meteorological data, thus improving the accuracy and applicability of distributed PV output prediction. Patent CN202511241274.4 proposes a joint optimization clustering method for distributed PV power prediction, solving the technical problem of low accuracy in existing power prediction methods and achieving the technical effect of improving prediction accuracy. However, research on distributed PV power prediction still faces many challenges. First, common clustering algorithms are relatively sensitive to outliers and require manually preset cluster numbers, which negatively impacts the application scenarios of distributed PV clusters. Simultaneously, commonly used clustering methods that directly utilize the geographical location of PV power plants lack an intuitive theoretical basis and struggle to guarantee the output correlation of sub-clusters. Secondly, for distributed photovoltaic (PV) clusters, a common method is to divide the clusters into sub-clusters, select a benchmark power station, and then amplify its power prediction results with certain weights to obtain the prediction results for the entire cluster. However, when this upscaling method is applied to distributed PV clusters with a wide spatial distribution, the output of the benchmark power station often exhibits inconsistent power fluctuation characteristics. This inconsistency is amplified by the upscaling method, leading to a decrease in the cluster prediction accuracy. Therefore, introducing better cluster division methods and power prediction techniques can further improve the power prediction performance of distributed PV clusters. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic cluster power prediction method based on hierarchical spatial feature clustering and improved graph attention network to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A photovoltaic cluster power prediction method based on hierarchical spatial feature clustering and improved graph attention network includes the following steps. S1. Construct a distributed photovoltaic cluster dataset, extract physical features such as installation tilt angle and tracking method, and calculate the solar altitude angle using the geographical coordinates of the photovoltaic power station; S2. First, input the photovoltaic module features into the affinity propagation AP algorithm for primary cluster analysis, and then use the solar altitude angle sequence as the feature input into the affinity propagation AP algorithm for cluster analysis to divide the photovoltaic cluster into several subsets. S3. Build a GAT-Encoder-Decoder deep learning model. For each distributed photovoltaic sub-cluster, after sorting and merging the historical power and historical meteorological time series data, input them into the GAT-Encoder-Decoder deep learning model for training. After training is completed, save the prediction model with the same number of sub-clusters. S4. For each distributed photovoltaic sub-cluster, the collected meteorological time series data is input into the prediction model trained in the previous step, and the day-ahead power prediction results of each power station in the sub-cluster are inferred and output; the prediction results of all power stations are summed to obtain the power prediction result of the entire distributed photovoltaic cluster.

[0007] As a preferred embodiment of the present invention: if photovoltaic power data uses per-unit values, the cluster power prediction results... Based on the installed capacity of the photovoltaic power station Weighted summation: , This is the photovoltaic power prediction result.

[0008] As a preferred embodiment of the present invention: step S1 specifically involves: the declination angle represents the angle between the line connecting the center of the Sun and the center of the Earth and the equatorial plane, the declination angle The calculation formula is: in, Ordinal numbers indicate the number of days in a year for a given day; Solar altitude angle It refers to the angle between the direct sunlight from the center of the sun and the local horizontal plane, and the calculation formula is as follows: in, The latitude of the power station. Let be the hour angle, and its calculation formula is as follows: in, Local true solar time; Solar radiation and They are directly proportional, obtained by iterating through the time steps. Time series.

[0009] As a preferred embodiment of the present invention: step S2 specifically involves: the off-diagonal elements of the similarity matrix S of the affinity propagation (AP) algorithm being points. With point The negative of the distance, using Euclidean distance, is calculated using the following formula: in, and Is the index as i and k The data points. Additionally, the diagonal elements. This is called preference; After obtaining the input matrix, the information propagated between data points consists of the responsibility matrix R and the availability matrix A. The elements of these two matrices... and The following iterative formula is used to obtain: At the beginning of the iteration, A is set to 0; The output is E=R+A, where the position k of the maximum value in the i-th row of E indicates that the cluster center to which point i belongs is point k.

[0010] As a preferred embodiment of the present invention: step S3 specifically involves: constructing a GAT layer, firstly generating a feature tensor from the distributed photovoltaic cluster dataset to form a unified spatiotemporal graph, and then using a weight matrix... As a shared, learnable linear transformation, it is applied to the feature vector of each node in the graph: ; in For the input feature dimension, To output feature dimensions, Let i be the initial feature vector of node i; then, the GAT layer transforms the feature vectors of the two nodes. and The input is fed into a single-layer feedforward neural network to compute the unnormalized attention score. To quantify the importance of the features of node j to node i, a single-layer feedforward neural network consists of a weight vector. After parameterization, the concatenated feature vectors are multiplied by 'a', and the LeakyReLU nonlinear activation function is applied. in This indicates concatenation; GAT uses masked attention and attention scores. The calculation is performed only on node j, which is a direct neighbor of node i. Then, the softmax function is applied to all neighbors of a node to reduce the original attention score. Normalized to a probability distribution: The obtained coefficients These are normalized attention weights, and for each node i, their sum is 1; The principle and process of the GAT layer are as follows: Figure 2 As shown; like Figure 3As shown, for a single GAT layer, stacking them together yields an encoder / decoder, which is then used to build a GAT-Encoder-Decoder prediction model. The GAT encoder traverses each GAT layer, and the GAT module of each layer calculates the attention coefficient between each node and all its spatiotemporal neighbors on the entire unified spatiotemporal graph. It also weights and aggregates the features of neighboring nodes and updates the feature representation of the node. The decoder works on the causal graph, performing the same GAT processing flow as the encoder, and gradually generating the hidden state representation for each future time step. Finally, the dataset of each sub-cluster generated in step S2 is input into the GAT-Encoder-Decoder model for training, and the trained model is saved. Each sub-cluster will obtain a corresponding prediction model.

[0011] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention takes into account the differences in physical attributes and time-series output under the spatial distribution of photovoltaic power stations. Through hierarchical spatial characteristic clustering analysis, it obtains sub-cluster division results that better reflect the spatial characteristic state of photovoltaic clusters, thereby improving the correlation of photovoltaic power station output within sub-clusters.

[0012] (2) The present invention is based on a deep learning model of graph attention network (GAT) and encoder-decoder. This model captures the spatial dependence and temporal features within a distributed photovoltaic power station cluster by using a unified spatiotemporal graph. It can learn the relative importance between different neighboring sites and effectively alleviate the problem of local difference statistical amplification in the prediction of distributed photovoltaic clusters under conventional upscaling methods. Attached Figure Description

[0013] Figure 1 This is a flowchart of the method of the present invention.

[0014] Figure 2 This is a flowchart illustrating the principle of the GAT layer.

[0015] Figure 3 This is a diagram of the Encoder-Decoder model architecture built based on the GAT layer in this invention.

[0016] Figure 4 This is a heatmap showing the output correlation within the sub-cluster in this invention.

[0017] Figure 5 This is a comparison chart of power prediction results using different methods in the embodiments of the present invention.

[0018] Figure 6 This is a sample of the prediction results for a single power station in sub-cluster I in this embodiment of the invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 In this embodiment of the invention, a photovoltaic cluster power prediction method based on hierarchical spatial feature clustering and improved graph attention network includes the following steps: S1. Extract physical characteristics such as installation tilt angle and tracking method, and calculate the solar altitude angle using the geographical coordinates of the photovoltaic power station.

[0021] When calculating spacetime angles, declination is the angle between the line connecting the Sun and the Earth's center and the equatorial plane, and it changes daily over time. The calculation formula is: Where d is the day ordinal number, representing the number of days in a year for a given day.

[0022] Solar altitude angle This refers to the angle between the direct sunlight from the center of the sun and the local horizontal plane. The calculation formula is as follows: in, The latitude of the power station. Let be the hour angle, and its calculation formula is as follows: in, This is the local true solar time.

[0023] Astronomical radiation refers to the amount of solar radiation reaching the upper boundary of Earth's atmosphere, determined by the Sun's astronomical position. This variable is related to... They are directly proportional. This can be obtained by iterating through the time steps using the above formula. Time series.

[0024] S2. Based on the results of S1, construct the similarity matrix S input to the AP algorithm. Its off-diagonal elements are the negatives of the distances between points i and k, calculated using Euclidean distance: The diagonal element s(i,k) is called the "preference". The number of sample points (i.e., the number of clusters) determined by AP is affected by the input preference. Different values ​​will produce different numbers of clusters. When the preference is the median of the input similarity, a moderate number of clusters will be produced. When the preference is the minimum of the input similarity, fewer clusters will be produced. By controlling the preference value, the total number of sub-regions N will be controlled within the range of 3 to 5 according to the power grid planning requirements.

[0025] After obtaining the input matrix, the information propagated between data points (affinity) consists of the responsibility matrix R and the availability matrix A, which are obtained through the following iterative formula: At the start of the iteration, A is set to 0.

[0026] The algorithm outputs E = R + A. Here, the position k of the maximum value in the i-th row of E indicates that point i belongs to cluster center k.

[0027] S3. Build the GAT-Encoder-Decoder deep learning model and train it to obtain the prediction model for each sub-cluster.

[0028] Construct the GAT layer. First, generate a feature tensor from the distributed photovoltaic cluster dataset to form a unified spatiotemporal graph, which is composed of a weight matrix. (in For the input feature dimension, The output feature dimension is used as a shared, learnable linear transformation applied to the feature vector of each node in the graph: in Let i be the initial feature vector of node i. Then, the GAT layer computes an unnormalized attention score. The importance of the features of node j to node i is quantified. This is achieved by transforming the feature vectors of the two nodes. and This is implemented by inputting into a single-layer feedforward neural network, which consists of a weight vector. Parameterization is achieved by concatenating the feature vectors, performing a dot product with 'a', and then applying the LeakyReLU non-linear activation function. in This indicates concatenation. To inject graph structure into the model, GAT employs masked attention and attention scores. Only for node j in the direct neighborhood of node i (i.e. ,in The calculation is performed through a self-loop (including node i itself). Then, to convert the original attention score... To normalize to a probability distribution, the softmax function needs to be applied to all neighbors of a node: The obtained coefficients These are normalized attention weights. For each node i, their sum is 1.

[0029] For a single GAT layer, stacking them together yields an encoder / decoder, which is then used to build a GAT-Encoder-Decoder prediction model, such as... Figure 2 As shown, the GAT encoder traverses each GAT layer. Each layer's GAT module calculates the attention coefficients between each node and all its spatiotemporal neighbors across the entire unified spatiotemporal graph, weights and aggregates the features of neighboring nodes, and updates the node's feature representation. The decoder operates on the causal graph, where it performs the exact same GAT processing flow as the encoder, progressively generating hidden state representations for each future time step.

[0030] The dataset of each sub-cluster generated by S2 is input into the GAT-Encoder-Decoder model for training. The trained model is saved, and each sub-cluster will obtain a corresponding prediction model.

[0031] S4. Input the meteorological dataset of the sub-cluster for the day to be predicted into the prediction model saved in S3, and infer the photovoltaic power prediction result of each power station in the sub-cluster on the target date. The cluster power prediction results are obtained by summing them up. If photovoltaic power data uses per-unit values, it needs to be based on the installed capacity of the photovoltaic power plant. Weighted summation: This invention underwent a simulation comparison experiment on a cluster of 30 distributed photovoltaic power stations in Shanghai. The prediction accuracy results of different prediction methods were recorded and compared. Experimental method 1 used the hierarchical spatial characteristic cluster partitioning method and GAT-Encoder-Decoder prediction model proposed in this invention. Experimental method 2 used a geographical location-based cluster partitioning and a Long Short-Term Memory (LSTM) network model as the benchmark method. During the training phase, historical power data and meteorological data from February 10, 2024 to August 30, 2024 were used, divided into a 70% training set, a 15% test set, and a 15% validation set.

[0032] During the cluster partitioning phase, the method of this invention ultimately outputs three sub-clusters, as shown in Table 1.

[0033] Calculate the output correlation of the member power plants for the three resulting sub-clusters, such as... Figure 4 As shown in the figure; the mean correlation values ​​are shown in Table 2.

[0034] The results show that the sub-cluster partitioning method used in this invention can ensure that all distributed photovoltaic sub-clusters maintain a high output correlation.

[0035] After the forecast phase is completed, the power forecast results for the target photovoltaic cluster as a whole on the forecast date are as follows: Figure 5 As shown in Table 3, the prediction accuracy is as follows.

[0036] The results show that the method proposed in this invention has the smallest prediction error; both the cluster partitioning based on hierarchical spatial feature clustering and the GAT-Encoder-Decoder model can improve the accuracy of power prediction for distributed photovoltaic clusters.

[0037] In addition to the overall cluster scale, the prediction results for individual power plants are shown using power plants 2 and 17 in sub-cluster I, which are divided by the method of this invention, as examples. Figure 6 As shown in Table 4, the corresponding prediction accuracy is as follows.

[0038] This result shows that although there are significant differences in power output fluctuations between power plant 2 and power plant 17 in the same sub-cluster, the prediction accuracy of the method proposed in this invention remains at a relatively good level at the scale of a single power plant. This indicates that the method captures the power output fluctuation characteristics of different power plants well and can be applied to both cluster-scale and single-station-scale prediction.

[0039] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0040] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A photovoltaic cluster power prediction method based on hierarchical spatial feature clustering and improved graph attention network, characterized in that, The method comprises the following steps: S1, constructing a distributed photovoltaic cluster dataset, extracting physical characteristics of the distributed photovoltaic cluster, and calculating a solar elevation angle through geographic coordinates of the photovoltaic power station; S2, inputting the physical characteristic set into an affinity propagation (AP) algorithm for primary clustering analysis, and inputting a solar elevation angle sequence as a feature into the AP algorithm for clustering analysis, so as to divide the photovoltaic cluster into a plurality of subsets; S3, building a GAT-Encoder-Decoder deep learning model, for each photovoltaic subset, after historical power and historical meteorological time series data are sorted and combined, the data are input into the GAT-Encoder-Decoder deep learning model for training, and the trained prediction model is saved, and the number of prediction models is the same as the number of subsets; S4, for each photovoltaic subset, meteorological time series data are sorted and input into the prediction model trained in the previous step, and a day-ahead power prediction result of each power station in the subset is output by inference; and the prediction results of all power stations are accumulated to obtain a power prediction result of the entire distributed photovoltaic cluster. Wherein, at the beginning of iteration, A is set to 0; 2. The photovoltaic cluster power prediction method based on hierarchical spatial feature clustering and improved graph attention network according to claim 1, characterized in that, If the photovoltaic power data is using a unit, the cluster power prediction result is based on the installed capacity of the photovoltaic power station weighted sum: , is a photovoltaic power prediction result. 3.The photovoltaic cluster power forecasting method based on hierarchical spatial feature clustering and improved graph attention network according to claim 1, wherein, The step S1 is specifically: the declination angle represents the angle between the sun-Earth center connecting line and the equatorial plane, and the calculation formula of the declination angle is: ​ wherein is the day of the year, indicating the day number in a year; Solar altitude angle is the angle between the direct sunlight from the sun center and the local horizontal plane, and the calculation formula is as follows: wherein is the latitude of the power plant, is the hour angle, which is calculated as follows: wherein, is local true solar time; Solar radiation and They are directly proportional, obtained by iterating through the time steps. Time series. 4.The photovoltaic cluster power forecasting method based on hierarchical spatial feature clustering and improved graph attention network according to claim 1, wherein, The step S2 is specifically: non-diagonal elements of the similarity matrix S of the affinity propagation AP algorithm is a point is a point The reciprocal of the distance between the point and the point is calculated using the Euclidean distance, and the formula is: wherein, with is the index of i with k data points; diagonal elements are called preferences; After obtaining the input matrix, the information propagated between data points is composed of the responsibility matrix R and the availability matrix A, whose elements With are obtained by the following iterative formula: Finally, the dataset of each subset generated in step S2 is input into the GAT-Encoder-Decoder model for training, and the trained model is saved, and each subset will obtain a corresponding prediction model. The output result is E = R + A, where the first i row maximum value position k representative point i The cluster center to which it belongs is point k . 5.The photovoltaic cluster power forecasting method based on hierarchical spatial feature clustering and improved graph attention network according to claim 1, wherein, The step S3 is specifically: constructing a GAT layer, first generating a feature tensor from a distributed photovoltaic cluster dataset to form a unified space-time graph, and applying a weight matrix as a shared, learnable linear transformation to the feature vector of each node in the graph: ; wherein is the input feature dimension, is the output feature dimension, is the initial feature vector of node i; The GAT layer then inputs the transformed feature vectors of the two nodes and to a single-layer feedforward neural network to compute an unnormalized attention score quantifying the importance of the features of node j to node i, parameterized by a weight vector , which is concatenated with a and dot-producted, and a LeakyReLU nonlinear activation function is applied: where denotes concatenation, GAT employs masked attention, attention scores are computed only for nodes j in the direct neighborhood of node i, then, a softmax function is applied to all neighbors of a node to normalize the raw attention scores to a probability distribution: The obtained coefficients are normalized attention weights, the sum of which for each node i is 1 ; For the constructed single GAT layer, the GAT-Encoder-Decoder model is obtained after stacking, and the GAT-Encoder-Decoder prediction model is further built, the GAT encoder traverses each GAT layer, the GAT module of each layer calculates the attention coefficient between each node and all spatiotemporal neighbors on the entire unified spatiotemporal graph, aggregates the features of the neighbor nodes by weighting, and updates the feature representation of the node, and the decoder works on the causal graph, the decoder performs the same GAT processing procedure as the encoder, and gradually generates the hidden state representation of each future time step. 6.The photovoltaic cluster power forecasting method based on hierarchical spatial feature clustering and improved graph attention network according to claim 5, wherein, The physical characteristics extracted in the step S1 include an installation inclination of the photovoltaic module and a tracking mode. 7.The photovoltaic cluster power forecasting method based on hierarchical spatial feature clustering and improved graph attention network according to claim 1, wherein, The step S1 further comprises data preprocessing on the historical power and meteorological time series data, and the preprocessing includes sliding window slicing, abnormal data elimination, standardization processing and time format conversion. 8.The photovoltaic cluster power forecasting method based on hierarchical spatial feature clustering and improved graph attention network according to claim 1, wherein, The attention weight calculation in the GAT layer adopts a multi-head attention mechanism, wherein the number of learned projections h is set to 4 to 8 groups. 9.The photovoltaic cluster power forecasting method based on hierarchical spatial feature clustering and improved graph attention network according to claim 5, wherein, The prediction result output in the step S4 is further subjected to precision evaluation by a verification module, the evaluation indexes include a root mean square error and a mean absolute error, and the model parameters are dynamically adjusted according to the evaluation result. 10.The photovoltaic cluster power forecasting method based on hierarchical spatial feature clustering and improved graph attention network according to claim 1, wherein, The method comprises the following steps: S1, constructing a distributed photovoltaic cluster dataset, extracting physical characteristics of the distributed photovoltaic cluster, and calculating a solar elevation angle through geographic coordinates of the photovoltaic power station; S2, inputting the physical characteristic set into an affinity propagation (AP) algorithm for primary clustering analysis, and inputting a solar elevation angle sequence as a feature into the AP algorithm for clustering analysis, so as to divide the photovoltaic cluster into a plurality of subsets; S3, building a GAT-Encoder-Decoder deep learning model, for each photovoltaic subset, after historical power and historical meteorological time series data are sorted and combined, the data are input into the GAT-Encoder-Decoder deep learning model for training, and the trained prediction model is saved, and the number of prediction models is the same as the number of subsets; S4, for each photovoltaic subset, meteorological time series data are sorted and input into the prediction model trained in the previous step, and a day-ahead power prediction result of each power station in the subset is output by inference; and the prediction results of all power stations are accumulated to obtain a power prediction result of the entire distributed photovoltaic cluster.

Images