Ultra-short-term photovoltaic power prediction method in complex meteorological environment

Abstract

The invention discloses an ultra-short-term photovoltaic power prediction method in a complex meteorological environment, and belongs to the technical field of extreme weather power prediction, and the prediction method comprises the following specific steps: I, selecting historical power data and historical meteorological data of a photovoltaic power station in a Saggolacking region, deleting missing data and error data in a historical data set, and carrying out the prediction of the historical power data and the historical meteorological data; dividing a training set, a verification set and a test set, and finally performing normalization processing on the data; iI, decomposing the complex meteorological factors by using SMD to obtain a plurality of IMFs and trend terms; iII, judging whether each characteristic time sequence is related or not according to the obtained trend term, and taking a historical value of a meteorological factor as a node characteristic; compared with the prior art, comprehensive graph structure information can be captured, the quality of dynamic graph feature information extraction is improved, and the prediction precision is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of extreme weather power prediction, and in particular to an ultra-short-term photovoltaic power prediction method under complex meteorological environments. Background Art

[0002] In the context of developing a new power system with new energy as the main supplier, expanding the proportion of clean energy consumption represented by solar energy and building a low-carbon, safe and efficient energy system are one of the important measures to achieve my country's "dual carbon" strategic goals. With the rapid development of the photovoltaic industry, my country's cumulative photovoltaic installed capacity has continued to grow. By the end of September 2023, the cumulative installed capacity of photovoltaics reached 520 million kilowatts, and the installed capacity showed a rapid growth trend, with an annual growth rate of more than 50%. China's northwest desert, Gobi and desert areas are rich in solar energy resources. Photovoltaic development in these areas can better form economies of scale and reduce costs in land, infrastructure, operation and maintenance. The advancement of large-scale photovoltaic base construction will bring significant growth to the development of new energy. It is estimated that the new installed capacity of wind power and photovoltaic power generation in the country will exceed 200 million kilowatts in 2023. However, most of the "Shagohuang" areas are above 1,000 meters above sea level, with harsh climatic conditions, strong winds and sand, complex meteorological factors, and many dust days, thunderstorms, and sandstorms, which make the output of photovoltaic power stations extremely random and volatile, bringing severe challenges to photovoltaic power prediction. Therefore, accurate power prediction technology is urgently needed to provide strong support for the dispatching and operation of active distribution networks and photovoltaic consumption in the "Shagohuang" area.

[0003] At present, the research on photovoltaic power prediction technology is becoming more and more mature, and most of them choose artificial neural network methods for prediction. Considering the complex changes of meteorological factors and the different degrees of influence on power, many studies will first perform feature screening on the data when selecting input data. The main methods used in the existing technology are: by performing correlation analysis on characteristic factors, factors with strong correlation are selected as input; a method based on adaptive reconstruction algorithm to optimize features to reduce data dimension; Morgan-Stone algebra and deterministic independence screening (MS-SIS) method is used to filter meteorological features and select characteristic factors with strong influence on power as input; Spearman rank correlation coefficient and grey correlation analysis are combined to screen various meteorological features; Pearson correlation coefficient is used to identify meteorological variables that significantly affect the power generation of photovoltaic power stations. However, due to the changeable climate in the "Shagohuang" area and the proneness to complex weather, the required input data is obtained only through simple feature screening, without considering the mutual coupling between meteorological factors, and it is impossible to accurately characterize the factors that affect power in complex environments.

[0004] In recent years, graph structures have attracted the attention of relevant scholars due to their advantages in intuitively representing the dependencies between nodes. In the field of power prediction, graph modeling is mostly performed for distributed stations. The main methods used in existing technologies are: considering the output and location information of each wind turbine to construct a weighted directed graph; decomposing the output sequence of each photovoltaic power station, and combining the sub-modal components of the fluctuations of each site into multiple graph structures; dynamically constructing a graph adjacency matrix based on the correlation coefficient and distance between sites to extract spatial features. Most of them only focus on the site location information and output information, and do not consider meteorological factors.

[0005] In the power forecasting link, the Transformer model has excellent performance in solving various tasks of time series prediction, and the model has gradually been applied to power forecasting. The main methods used in the existing technology are: using a two-layer L-Transformer model to capture time dependencies while focusing on key feature quantities to achieve photovoltaic power prediction; using a multi-scale time block sparse autoencoder Transformer model for power prediction, embedding historical power data into the network through sparse autoencoder features, and using a multi-head self-attention mechanism to capture and identify sequence dependency patterns, and modeling long-term sequence dependencies. However, the original Transformer model designed for natural language processing overemphasizes global semantics, does not capture graph structure information comprehensively, and is difficult to accurately extract features, resulting in low prediction accuracy; therefore, it is particularly important to invent an ultra-short-term photovoltaic power prediction method under complex meteorological environments. Summary of the invention

[0006] The purpose of the present invention is to solve the defects in the prior art and to propose an ultra-short-term photovoltaic power prediction method under complex meteorological environment.

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

[0008] An ultra-short-term photovoltaic power prediction method under complex meteorological environment, the specific steps of the prediction method are as follows:

[0009] Ⅰ. Select the historical power data and historical meteorological data of photovoltaic power stations in the "Shagohuang" area, delete the missing data and erroneous data in the historical data set, and then divide the data into training set, validation set and test set, and finally normalize the data;

[0010] Ⅱ. Use SMD to decompose complex meteorological factors to obtain multiple IMF and trend items;

[0011] III. Determine whether each characteristic time series is related based on the obtained trend items, and use the historical values ​​of meteorological factors as node characteristics;

[0012] IV. Based on the obtained IMF between each node, an edge feature matrix is ​​established to characterize the intrinsic relationship between feature time series, and a dynamic graph that changes over time is constructed;

[0013] V. Construct a hybrid neural network, input the dynamic graph data into the self-attention module to adaptively capture the dependencies between nodes, and pass the obtained data into the Trans-GNN module. Alternately use the Transformer layer and the GNN layer to extract graph feature information, and then calculate the output of the improved ISAM-Trans-GNN.

[0014] VI. Calculate the network prediction error and use the Adam optimizer to optimize the prediction model parameters;

[0015] VII. Repeat the extraction of graph feature information and the optimization of ISAM-Trans-GNN network parameters. If the maximum number of iterations is reached, the iteration is terminated and the improved ISAM-Trans-GNN network parameters are output;

[0016] VIII. Use the trained improved ISAM-Trans-GNN network to perform ultra-short-term photovoltaic power forecasting and obtain the predicted power for the period to be predicted.

[0017] As a further solution of the present invention, the specific steps of deleting missing data and erroneous data in the historical data set in step I, then dividing the data into a training set, a validation set and a test set, and finally normalizing the data are as follows:

[0018] S1.1: Set the time resolution to 15 minutes, and collect the power generation data of the photovoltaic base in the "Shagohuang" area in the past year according to the set time resolution to construct the corresponding historical data set;

[0019] S1.2 Use anomaly detection methods and data mining technology to delete and correct inconsistent errors, unreasonable values, missing values, and duplicate records in historical data sets, normalize the cleaned data, and divide it into training set, validation set, and test set in a ratio of 8:1:1. The specific calculation formula for normalization is as follows:

[0020]

[0021] In the formula, represents normalized power; p min and p max Represent the maximum and minimum values ​​of the power sequence respectively.

[0022] As a further solution of the present invention, the specific steps of using SMD to decompose complex meteorological factors in step II to obtain multiple IMFs and trend items are as follows:

[0023] S2.1: For a given time series f, set h 1,0 =f initializes the algorithm and calculates h 1,0 The local maximum and minimum values ​​of the cubic spline interpolation are used to obtain the corresponding upper and lower envelopes, and then the point-by-point average value m is obtained through the upper and lower envelopes. 1,1 , solve for h 1,1 =h 1,0 -m 1,1 ;

[0024] S2.2: Repeat the calculation of the upper and lower envelopes, the point-by-point averages, and the solution until h 1,k =h 1,k-1 -m 1,k When the envelope of is symmetric about the time axis and the difference between its local maximum and the number of zero crossings is equal to -1, 0 or 1, h 1,k Defined as f1;

[0025] S2.3: Based on the above steps S2.1-S2.2, the next IMF component is searched again until the remainder is a monotonic or slowly oscillating function, and the process stops to obtain multiple groups of IMF components and trend terms f(t)=f1(t)+...+f n (t)+R n (t), t∈R, apply SSO to each IMF component separately, and apply the corresponding threshold parameter μ>0 to the output to obtain a set of multiple clusters G k ,k=1,...,n;

[0026] S2.4: By calculating each cluster to obtain the frequency estimate of the sub-signal, the extracted frequency estimate is used as input, and the initial IMF is further decomposed using the same SSO parameters to obtain the super-resolution IMF components and trend terms.

[0027] As a further solution of the present invention, the specific steps of constructing the dynamic graph that changes over time in step IV are as follows:

[0028] S3.1: Select different meteorological factor time series as the node matrix of the graph structure, where each node contains L x Historical values;

[0029] S3.2: The residual trend item obtained by SMD decomposition is used to determine whether each characteristic factor is highly correlated, and whether to establish an edge is determined based on the judgment result. Then, the connectivity between nodes is represented by the graph adjacency matrix;

[0030] S3.3: The edge feature matrix determined by the correlation between the multiple super-resolution IMF components obtained by SMD decomposition and the initial time series is defined as ε = (e ijk)∈R N×N×K ,in e ijk Represents the time series of meteorological factors x i With x j The correlation of the kth IMF component between .

[0031] As a further solution of the present invention, the specific steps of inputting the dynamic graph data into the self-attention module in step V to adaptively capture the dependencies between nodes, and the obtained data is passed to the Trans-GNN module, and the Transformer layer and the GNN layer are alternately used to extract the graph feature information are as follows:

[0032] S4.1: An RNN layer is used as the position embedding layer on the node matrix Xt, and the position encoding is used to represent the order of the sequence, which contains the time information h(Xt). t )=RNN(X t ), and the hidden feature h(X t ) is passed to the next layer;

[0033] S4.2: Multiply the input data with the matrix Wq, matrix Wk and matrix Wv respectively to obtain the query vector q, key vector k and value vector v, then query the scaled dot product with all keys and divide by To output the corresponding matrix where d k Represents the dimension of the input Key;

[0034] S4.3: Multiple attention units are stacked side by side, and the projected attention structure produces a dv-dimensional output through dimension splicing. At the same time, the multi-sensor head self-attention operation is modified to To aggregate the information of nodes, edges and graph structures, where α i (i=0,...,K+1) represents the trainable parameters about the three types of information: nodes, edges, and graph structures, and then the feature mapping is completed by projection through the fully connected layer.

[0035] As a further solution of the present invention, the first item of the node feature is a primitive self-attention layer, representing the intra-time series relationship contained in the node; the inter-sequence relationship of the edge feature has K items, D imfk is the covariance matrix of the kth IMF of all time series; the last item A of the adjacency matrix is ​​the adjacency matrix of the dynamic graph.

[0036] As a further solution of the present invention, the constructed hybrid neural network specifically uses a Trans-GNN hybrid model that improves the Self-Attention mechanism, alternately uses Transformer layers and GNN layers to improve each other, and uses the Transformer layer to aggregate information of more related nodes to improve the message passing mechanism of GNN.

[0037] As a further solution of the present invention, the specific steps of using the trained improved ISAM-Trans-GNN network to perform ultra-short-term photovoltaic power prediction in step VIII are as follows:

[0038] S5.1: Input the dynamic graph information, initialize the weights of the improved ISAM-Trans-GNN neural network, set the maximum number of iterations K = 100, calculate the network prediction error, and use the Adam optimizer to optimize the model parameters;

[0039] S5.2: Determine whether k is greater than 100. If so, the iteration is terminated and the improved ISAM-Trans-GNN network parameters are output. Otherwise, set k=k+1 and return to the step of improving the ISAM-Trans-GNN network. Then, use the trained network model to predict the centralized photovoltaic power.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The ultra-short-term photovoltaic power prediction method under complex meteorological environment first pre-processes the data, decomposes the time series by super empirical mode decomposition, introduces SSO to further extract the initially obtained IMF components, takes the historical values ​​of meteorological factors as node features, establishes an edge feature matrix according to the IMF between nodes, constructs a dynamic graph that changes over time, constructs a hybrid neural network, inputs the dynamic graph data into the self-attention module to capture the dependency between nodes, and passes the obtained data into the Trans-GNN module to calculate the output result of the improved ISAM-Trans-GNN, and calculates the network error. Adam is used to optimize the prediction model parameters, and the ultra-short-term photovoltaic power prediction is performed with the help of the trained improved ISAM-Trans-GNN network to obtain the predicted power for the time period to be predicted. Compared with the prior art, the present invention can capture comprehensive graph structure information, improve the quality of extracting dynamic graph feature information, and further improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0043] Figure 1A flowchart of an ultra-short-term photovoltaic power prediction method under a complex meteorological environment proposed by the present invention;

[0044] Figure 2 A schematic diagram of the logic framework of an ultra-short-term photovoltaic power prediction method under a complex meteorological environment proposed by the present invention;

[0045] Figure 3 This is a T-time graph structure modeling block diagram of an ultra-short-term photovoltaic power prediction method under a complex meteorological environment proposed by the present invention;

[0046] Figure 4 This is a flow chart of the improved ISAM-Trans-GNN prediction algorithm for the ultra-short-term photovoltaic power prediction method under complex meteorological environments proposed by the present invention. DETAILED DESCRIPTION

[0047] Reference Figure 1-4 , an ultra-short-term photovoltaic power prediction method under complex meteorological environment, the specific steps of the prediction method are as follows:

[0048] The historical power data and historical meteorological data of photovoltaic power stations in the "Shagohuang" area were selected, and the missing data and erroneous data in the historical data set were deleted. Then the data were divided into training set, validation set and test set, and finally the data were normalized.

[0049] Specifically, the time resolution is set to 15 minutes, and the power generation data of the photovoltaic base in the "Shagohuang" area in the past year are collected according to the set time resolution to construct the corresponding historical data set. The inconsistent errors, unreasonable values, missing values ​​and duplicate records in the historical data set are deleted and corrected through anomaly detection methods and data mining techniques. The data obtained after cleaning are normalized and divided into training set, validation set and test set in the ratio of 8:1:1.

[0050] It should be further explained that the specific calculation formula for normalization processing is as follows:

[0051]

[0052] In the formula, represents normalized power; p min and p max Represent the maximum and minimum values ​​of the power sequence respectively.

[0053] SMD is used to decompose complex meteorological factors to obtain multiple IMFs and trend items.

[0054] Specifically, refer to Figure 2 It can be seen that for a given time series f, setting h 1,0=f initializes the algorithm and calculates h 1,0 The local maximum and minimum values ​​of the cubic spline interpolation are used to obtain the corresponding upper and lower envelopes, and then the point-by-point average value m is obtained through the upper and lower envelopes. 1,1 , solve for h 1,1 =h 1,0 -m 1,1 , repeatedly calculate the upper and lower envelopes, the point-by-point averages, and solve until h 1,k =h 1,k-1 -m 1,k When the envelope of is symmetric about the time axis and the difference between its local maximum and the number of zero crossings is equal to -1, 0 or 1, h 1,k Defined as f1, based on the above steps, the next IMF component is searched again until the remainder is a monotonic or slowly oscillating function, and the process stops to obtain multiple groups of IMF components and trend terms f(t) = f1(t) + ... + f n (t)+R n (t), t∈R, apply SSO to each IMF component separately, and apply the corresponding threshold parameter μ>0 to the output to obtain a set of multiple clusters G k , k=1,...,n, by calculating each cluster to obtain the frequency estimation value of the sub-signal, the extracted frequency estimation value is used as input, and the initial IMF is further decomposed using the same SSO parameters to obtain the super-resolution IMF components and trend terms.

[0055] It should be further explained that SSO is applied to each IMF component separately and is defined as:

[0056]

[0057] where h is the admissible window function,

[0058]

[0059] Parameter δ,a>0 and satisfies h a is a positive number. After applying an appropriate threshold parameter μ>0 to the output, a set G of n non-empty and non-intersecting clusters is obtained. k ,k=1,...,n. By calculating

[0060]

[0061] Get the frequency estimate of the sub-signal

[0062]

[0063] Finally, the extracted frequency is used as input, and the same SSO parameters are used to further decompose the initial IMF to obtain the super-resolution IMF components and trend terms. The formula is as follows:

[0064]

[0065] The correlation of each characteristic time series is determined based on the obtained trend items, and the historical values ​​of meteorological factors are used as node characteristics.

[0066] Based on the obtained IMF between nodes, an edge feature matrix is ​​established to characterize the intrinsic relationship between feature time series, and a dynamic graph that changes over time is constructed.

[0067] Specifically, refer to Figure 2 as well as Figure 3 It can be seen that the time series of different meteorological factors are selected as the node matrix of the graph structure, where each node contains the L x The residual trend item obtained by SMD decomposition is used to determine whether each characteristic factor is highly correlated, and whether to establish an edge is determined based on the judgment result. Then, the connectivity between nodes is represented by the graph adjacency matrix. The edge feature matrix determined by the correlation between multiple super-resolution IMF components obtained by SMD decomposition and the initial time series is defined as ε = (e ijk )∈R N×N×K ,in e ijk Represents the time series of meteorological factors x i With x j The correlation of the kth IMF component between .

[0068] It should be further explained that the graph adjacency matrix is ​​used to represent the connectivity between nodes, and its specific expression is as follows:

[0069]

[0070] Among them, ρ ij Represents the characteristic time series x i With x j The correlation of the residuals between

[0071]

[0072] A hybrid neural network is constructed, and the dynamic graph data is input into the self-attention module to adaptively capture the dependencies between nodes. The obtained data is passed to the Trans-GNN module, and the Transformer layer and the GNN layer are used alternately to extract graph feature information. Then, the output results of the improved ISAM-Trans-GNN are calculated.

[0073] Specifically, refer to Figure 4 It can be seen that an RNN layer is used as the position embedding layer on the node matrix Xt, and the position encoding is used to represent the order of the sequence, which contains the time information h(X t )=RNN(X t ), and the hidden feature h(X t ) is passed to the next layer, where the input data is multiplied by the matrix Wq, matrix Wk, and matrix Wv respectively to obtain the query vector q, key vector k, and value vector v. The query is then scaled dot product with all Keys and divided by To output the corresponding matrix where d k Represents the dimension of the input Key, stacks multiple attention units side by side, and the projected attention structure produces a dv-dimensional output through dimension splicing. At the same time, the multi-sensor head self-attention operation is modified to To aggregate the information of nodes, edges and graph structures, where α i (i=0,...,K+1) represents the trainable parameters about the three types of information: nodes, edges, and graph structures, and then the feature mapping is completed by projection through the fully connected layer.

[0074] It should be noted that the specific calculation formulas for the query vector q, key vector k and value vector v are as follows:

[0075] q=W q *x

[0076] k=W k *x

[0077] v=W v *x.

[0078] It is necessary to further explain that the first item of the node feature is a primitive self-attention layer, representing the intra-time series relationship contained in the node; the inter-sequence relationship of the edge feature has K items, D imfk is the covariance matrix of the kth IMF of all time series; the last item A of the adjacency matrix is ​​the adjacency matrix of the dynamic graph.

[0079] In addition, it should be noted that the constructed hybrid neural network is specifically a Trans-GNN hybrid model that uses an improved Self-Attention mechanism. Transformer layers and GNN layers are used alternately to improve each other, and the Transformer layer is used to aggregate information of more related nodes to improve the message passing mechanism of GNN.

[0080] Calculate the network prediction error and use the Adam optimizer to optimize the prediction model parameters.

[0081] Repeat the extraction of graph feature information and the optimization of ISAM-Trans-GNN network parameters. If the maximum number of iterations is reached, the iteration is terminated and the improved ISAM-Trans-GNN network parameters are output.

[0082] The trained improved ISAM-Trans-GNN network is used to perform ultra-short-term photovoltaic power forecasting to obtain the predicted power for the period to be predicted.

[0083] Specifically, the dynamic graph information is input, and then the improved ISAM-Trans-GNN neural network weights are initialized. The maximum number of iterations K is set to 100, and then the network prediction error is calculated. The Adam optimizer is used to optimize the model parameters, and it is determined whether k is greater than 100. If it is greater than 100, the iteration is terminated and the improved ISAM-Trans-GNN network parameters are output. Otherwise, k is set to k+1, and the step of improving the ISAM-Trans-GNN network is returned. Then, the trained network model is used to predict the centralized photovoltaic power.

Claims

1. A method for ultra-short-term photovoltaic power prediction under complex meteorological environment, characterized in that: The specific steps of the prediction method are as follows: Ⅰ. Select the historical power data and historical meteorological data of photovoltaic power stations in the "Shagohuang" area, delete the missing data and erroneous data in the historical data set, and then divide the data into training set, validation set and test set, and finally normalize the data; Ⅱ. Use SMD to decompose complex meteorological factors to obtain multiple IMF and trend items; III. Determine whether each characteristic time series is related based on the obtained trend items, and use the historical values ​​of meteorological factors as node characteristics; IV. Based on the obtained IMF between each node, an edge feature matrix is ​​established to characterize the intrinsic relationship between feature time series, and a dynamic graph that changes over time is constructed; V. Construct a hybrid neural network, input the dynamic graph data into the self-attention module to adaptively capture the dependencies between nodes, and pass the obtained data into the Trans-GNN module. Alternately use the Transformer layer and the GNN layer to extract graph feature information, and then calculate the output of the improved ISAM-Trans-GNN. VI. Calculate the network prediction error and use the Adam optimizer to optimize the prediction model parameters; VII. Repeat the extraction of graph feature information and the optimization of ISAM-Trans-GNN network parameters. If the maximum number of iterations is reached, the iteration is terminated and the improved ISAM-Trans-GNN network parameters are output; VIII. Use the trained improved ISAM-Trans-GNN network to perform ultra-short-term photovoltaic power forecasting and obtain the predicted power for the period to be predicted.

2. The ultra-short-term photovoltaic power prediction method under complex meteorological environment according to claim 1 is characterized in that: The specific steps of deleting missing data and erroneous data in the historical data set described in step I, then dividing the training set, validation set and test set, and finally normalizing the data are as follows: S1.1: Set the time resolution to 15 minutes, and collect the power generation data of the photovoltaic base in the "Shagohuang" area in the past year according to the set time resolution to construct the corresponding historical data set; S1.2 Use anomaly detection methods and data mining technology to delete and correct inconsistent errors, unreasonable values, missing values, and duplicate records in historical data sets, normalize the cleaned data, and divide it into training set, validation set, and test set in a ratio of 8:1:

1. The specific calculation formula for normalization is as follows: In the formula, represents normalized power; p min and p max Represent the maximum and minimum values ​​of the power sequence respectively.

3. The ultra-short-term photovoltaic power prediction method under complex meteorological environment according to claim 2 is characterized in that: The specific steps of using SMD to decompose complex meteorological factors in step II to obtain multiple IMF and trend items are as follows: S2.1: For a given time series f, set h 1,0 =f initializes the algorithm and calculates h 1,0 The local maximum and minimum values ​​of the cubic spline interpolation are used to obtain the corresponding upper and lower envelopes, and then the point-by-point average value m is obtained through the upper and lower envelopes. 1,1 , solve for h 1,1 =h 1,0 -m 1,1 ; S2.2: Repeat the calculation of the upper and lower envelopes, the point-by-point averages, and the solution until h 1,k =h 1,k-1 -m 1,k When the envelope of is symmetric about the time axis and the difference between its local maximum and the number of zero crossings is equal to -1, 0 or 1, h 1,k Defined as f1; S2.3: Based on the above steps S2.1-S2.2, the next IMF component is searched again until the remainder is a monotonic or slowly oscillating function, and the process stops to obtain multiple groups of IMF components and trend terms f(t)=f1(t)+...+f n (t)+R n (t), t∈R, apply SSO to each IMF component separately, and apply the corresponding threshold parameter μ>0 to the output to obtain a set of multiple clusters G k ,k=1,...,n; S2.4: By calculating each cluster to obtain the frequency estimate of the sub-signal, the extracted frequency estimate is used as input, and the initial IMF is further decomposed using the same SSO parameters to obtain the super-resolution IMF components and trend terms.

4. The method for ultra-short-term photovoltaic power prediction under complex meteorological environment according to claim 3 is characterized in that: The specific steps of constructing the dynamic graph that changes over time described in step IV are as follows: S3.1: Select different meteorological factor time series as the node matrix of the graph structure, where each node contains L x Historical values; S3.2: The residual trend item obtained by SMD decomposition is used to determine whether each characteristic factor is highly correlated, and whether to establish an edge is determined based on the judgment result. Then, the connectivity between nodes is represented by the graph adjacency matrix; S3.3: The edge feature matrix determined by the correlation between the multiple super-resolution IMF components obtained by SMD decomposition and the initial time series is defined as ε = (e ijk )∈R N×N×K ,in e ijk Represents the time series of meteorological factors x i With x j The correlation of the kth IMF component between .

5. The ultra-short-term photovoltaic power prediction method under complex meteorological environment according to claim 1 is characterized in that: In step V, the dynamic graph data is input into the self-attention module to adaptively capture the dependencies between nodes, and the obtained data is passed to the Trans-GNN module. The specific steps of alternately using the Transformer layer and the GNN layer to extract the graph feature information are as follows: S4.1: An RNN layer is used as the position embedding layer on the node matrix Xt, and the position encoding is used to represent the order of the sequence, which contains the time information h(Xt). t )=RNN(X t ), and the hidden feature h(X t ) is passed to the next layer; S4.2: Multiply the input data with the matrix Wq, matrix Wk and matrix Wv respectively to obtain the query vector q, key vector k and value vector v, then query the scaled dot product with all keys and divide by To output the corresponding matrix where d k Represents the dimension of the input Key; S4.3: Multiple attention units are stacked side by side, and the projected attention structure produces a dv-dimensional output through dimension splicing. At the same time, the multi-sensor head self-attention operation is modified to To aggregate the information of nodes, edges and graph structures, where α i (i=0,...,K+1) represents the trainable parameters about the three types of information: nodes, edges, and graph structures, and then the feature mapping is completed by projection through the fully connected layer.

6. The ultra-short-term photovoltaic power prediction method under complex meteorological environment according to claim 1, characterized in that: The specific steps of using the trained improved ISAM-Trans-GNN network to perform ultra-short-term photovoltaic power forecasting as described in step VIII are as follows: S5.1: Input the dynamic graph information, initialize the weights of the improved ISAM-Trans-GNN neural network, set the maximum number of iterations K = 100, calculate the network prediction error, and use the Adam optimizer to optimize the model parameters; S5.2: Determine whether k is greater than 100. If so, the iteration is terminated and the improved ISAM-Trans-GNN network parameters are output. Otherwise, set k=k+1 and return to the step of improving the ISAM-Trans-GNN network. Then, use the trained network model to predict the centralized photovoltaic power.

Images