Wind power cluster data fusion method and system based on heterogeneous space-time diagram convolutional network

Abstract

The invention discloses a wind power cluster data fusion method and system based on a heterogeneous space-time diagram convolutional network, and the method comprises the steps: representing each wind power station as a node, and dynamically adjusting the size of an adjacent window of each station data through calculating the influence weight between different data; carrying out the fusion and reconstruction of the time sequence multi-modal features through a relational graph convolution network; then, on the basis of the design of integrating a two-dimensional convolutional neural network, filters with different sizes and a multi-order graph convolutional network, further enhancing the representation capability of cluster data prior information from time and space dimensions; and finally, embedding a specially designed lightweight graph convolution attention module, and carrying out hierarchical modeling on spatial dependence of a neighborhood field and local space fields of different data modes. According to the method, more information among different data can be captured in a scene where modals are seriously lost, spatial and temporal changes of data among stations can be effectively learned, and spatial and temporal correlation among graph nodes can be dynamically adjusted.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and relates to a method and system for wind farm data fusion, in particular to a method and system for wind power cluster data fusion based on a heterogeneous spatiotemporal graph convolutional network. Background Technique

[0002] Currently, the research on wind power cluster data fusion mainly focuses on two aspects: one is multi-source data fusion based on traditional statistical methods, and the other is multi-source data fusion based on machine learning and deep learning. Traditional statistical methods usually fuse different types of data through methods such as weighted average, interpolation, and regression. However, when facing the high-dimensionality and complexity of data, it is difficult to fully extract the internal relationships between data. In the application of machine learning and deep learning, although there have been many studies in recent years attempting to use deep neural networks (such as CNN, RNN, etc.) to process wind power data, these methods usually focus on the processing of single-type data and do not consider the fusion problem of the static geographical location information of wind farms and the spatio-temporal correlation of dynamic power-meteorological data.

[0003] The spatiotemporal graph convolutional network (STGCN) is a deep learning method that combines graph convolutional networks and spatio-temporal characteristics, and has made significant progress in the processing of time series data and graph-structured data in recent years. This method can automatically learn the spatial relationships between data nodes through graph convolutional operations. At the same time, compared with traditional convolutions, graph convolutions can more effectively process non-Euclidean space data, which is beneficial to improving the fusion quality of the geographical location map of wind power cluster sites and power data. Summary of the Invention

[0004] The present invention aims to improve the dynamic spatio-temporal correlation of existing multi-source data fusion methods for wind power clusters, and proposes a method and system for wind power cluster data fusion based on a heterogeneous spatiotemporal graph convolutional network (HSTGCN).

[0005] The technical solution adopted by the method of the present invention is: a method for wind power cluster data fusion based on a heterogeneous spatiotemporal graph convolutional network, including the following steps: Step 1: Obtain N the original data of wind farm sites, including power, wind speed, wind direction, and temperature, and perform normalization processing; Step 2: Obtain a heterogeneous graph , where M represents the number of data patterns in the heterogeneous graph, and V and E are the sets of nodes and edges respectively; the matrix ={ } is used to describe the spatial proximity of any two nodes with connecting edges , if , for and , then , and there is an edge between the two nodes, otherwise, , the connection between them does not exist; is the threshold for eliminating weak connections; Step 3: Use the prior feature matrix of the Pearson correlation coefficient as the node embedding matrix, and calculate the feature matrix of T the N station data in a rolling manner ; where T is the step size required to construct the heterogeneous graph; Step 4: Add to the embedding space, and set the combined adaptive adjacency matrix to to obtain the final combined adjacency matrix A ; Step 5: Input the power, wind speed, wind direction, and temperature time series data of N wind farms into the heterogeneous spatio-temporal graph convolutional network to output the fusion result; The heterogeneous spatio-temporal graph convolutional network includes a dynamic heterogeneous graph constructor , a time correlation network based on CNN and a space correlation network based on GCN ; ; ; In the formula, is the power value at the q-1 time step before time t; is t the sampling data of the i th station at the q-1 time step before time ; In the formula, refers to the data dimension transformation layer of the single-scale convolutional network; is the time feature exploration layer of the multi-scale convolutional network; is the multi-scale feature splicing layer; , respectively represent and shared parameter sets of ; is the power value at the q-1 time step before time t; ; ​ is the initial spatial feature exploration layer, is the neighborhood attention weight learning layer, is the spatial attention weight learning layer; , respectively represent and 's shared parameter set, ; is the final output time feature three-dimensional tensor of all nodes; is the dynamic heterogeneous graph constructor.

[0006] Preferably, in step 3, ; ; In the formula, represents N Pearson correlation coefficient between the measured data of any two stations among i, j stations; , are the sampling power values of the i-th and j-th wind farms at time t respectively, and respectively represent the average values of the two groups of data, .

[0007] Preferably, in step 4, .

[0008] Preferably, in step 5, uses a F convolution filter with a size of 1×1 to operate on the one-dimensional sequence data of each node, and splices the obtained convolution outputs into an image with a size of q × F ; then, changes the one-dimensional sequence of each node into a group of two-dimensional images with a length of q ; in addition, N groups of consecutive images of N nodes will be fed into the next CNN layer in parallel, and capture time features along the time axis through convolution kernels of different scales; taking the number of nodes N as the channel size and using the 2D image as the time data of the node input to obtain the 3D data tensor , .

[0009] Preferably, in step 5, uses N groups of convolution filters to operate on each channel simultaneously, and uses a group of convolution filters with different sizes to extract multi-scale time features of each channel in parallel; Consisting of a series of single-scale convolution operations, where each convolution scale is applied along the image, and the size of each node is q × F ; ; In the formula, represents a convolution filter of size λ × λ × N where the n th convolution filter is performed on the th channel of the input data tensor n to output the n th temporal feature map ; λ is the configuration parameter of the convolution filter.

[0010] Preferably, in step 5, the learned features of each node are used as the output of this layer and regarded as the initial node features, and are input into the dynamic heterogeneous graph in units of q time steps, and the spatial features are further explored through the spatial correlation network based on GCN; The concatenated temporal feature tensors of each station and all stations are respectively: ; ; In the formula, truncates the multi-scale features to the same size, is the final output temporal feature three-dimensional tensor of all nodes with the same dimension as .

[0011] Preferably, in step 5, a multi-order neighborhood GCN is used to form to extract the spatial correlation output by the temporal correlation network based on CNN; The multi-order neighborhood GCN has a special parameter, that is, the k -order neighborhood of the target node, defined as , where represents the minimum number of edges moving in the same direction from m to for the pattern , is the i th node, is the set of nodes of the pattern m , is the number of all wind farm stations under the pattern m; meanwhile, from the adjacency matrix Calculate the spatial proximity between the target node and its k -th order neighborhood; Described by equations (1) to (5): (1) (2) (3) (4) (5) In the formula, represents the -th order spatial feature tensor captured from all nodes at , and its dimension is the same as k ; represents a diagonalizable degree matrix that measures the sum of the spatial proximities of the -th order neighborhood related to the target node, and k is the diagonal element of ; is used to normalize to avoid weight explosion or vanishing; the initial spatial feature block is the -th order spatial feature tensor including all q time points; k The first channel of represents the current input of all nodes.

[0012] Preferably, in step 5, the neighborhood attention weight learning layer is designed to guide the GCN to learn the positions of focusing or forgetting along the channel dimension of K ; in order to obtain the attention weights of each layer , this layer first performs average pooling operations q and max pooling operations F in parallel for each channel along the height N and width dimensions to integrate the feature tensor, discard redundant features and obtain a comprehensive representation of the spatial information; then, after the multi-layer perceptron and sigmoid activation operation , the two pooled feature vectors are element-wise summed to assign neighborhood attention weights; The neighborhood attention learning mechanism of is: ; In addition, the initial spatial feature block is aggregated according to the neighborhood attention weights: ; In the formula, refers to the refined spatial feature tensor weighted by neighborhood attention. is the k th channel tensor of

[0013] Preferably, in step 5, the feature information on the width F is aggregated through max pooling operation and average pooling operation, the two resulting pooling features are concatenated, and then convolution is performed to output the spatial attention weight ; ; The spatial attention learning mechanism of is: ; In the formula, represents the convolution operation, is the final output of the neighborhood and local spatial weights.

[0014] The technical solution adopted by the system of the present invention is: a wind power cluster data fusion system based on a heterogeneous spatio-temporal graph convolutional network, including: one or more processors; a storage device for storing one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the wind power cluster data fusion method based on the heterogeneous spatio-temporal graph convolutional network.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The present invention constructs a data fusion model based on HSTGCN. By constructing a multi-station data heterogeneous graph network, the modeling of spatio-temporal dynamic features is realized under a unified framework, which not only retains the initial spatio-temporal characteristics of each station, but also maintains the independence of different data sources; a hybrid architecture of a multi-scale convolutional neural network and a high-order graph convolutional network is adopted to effectively enhance the interaction ability between multi-source heterogeneous data; a lightweight graph attention mechanism is introduced to realize the hierarchical representation of multi-dimensional spatial correlations between stations, providing rich historical training data for the subsequent wind power cluster prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The following uses examples and specific implementation manners to further illustrate the technical solution of the present invention. Additionally, some drawings are also used in the process of explaining the technical solution. For those skilled in the art, other drawings and the intention of the present invention can also be obtained based on these drawings without creative efforts.

[0017] Figure 1 Schematic flow diagram of the method according to the embodiment of the present invention.

[0018] Figure 2 Schematic diagram of the dynamic heterogeneous graph according to the embodiment of the present invention.

[0019] Figure 3 Structural diagram of the CNN time model according to the embodiment of the present invention.

[0020] Figure 4 Schematic diagram of the multi - order prior information aggregation according to the embodiment of the present invention.

[0021] Figure 5 Schematic diagram of the lightweight GCN graph convolution attention module according to the embodiment of the present invention.

[0022] Figure 6 Comparison graph of RMSE curves of different methods.

[0023] Figure 7 Comparison graph of cluster prediction results before and after data fusion.

[0024] Figure 8 Bar graph of cluster prediction errors before and after data fusion. Detailed implementation manner

[0025] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and detailed implementation manners and specific operations are given, but the protection scope of the present invention is not limited to the following embodiments.

[0026] As Figure 1 shown, a wind power cluster data fusion method based on a heterogeneous spatio - temporal graph convolutional network provided by this example includes the following steps: Step 1, obtain 5 pieces of original data of the stations, including power, wind speed, wind direction and temperature, and the data acquisition time resolution is 15 min; Step 2, data pre - processing, perform max - min normalization on the original data, and the formula is: (1) In the formula, and are the maximum and minimum values of the original data x respectively, and represents the normalized data.

[0027] Step 3, obtain the heterogeneous graph ; As Figure 2 shown. Represent this heterogeneous graph as , where M represents the number of data patterns in the heterogeneous graph, and V and E are the sets of nodes and edges respectively. The nodes in V are described as , where is the set of nodes of pattern m, is the i-th node. Similarly, , where is the set of edges of pattern m, the edges in can be described as , indicating that there is an edge between nodes and .

[0028] Matrix = } is used to describe the spatial proximity of any two nodes with connected edges , if , for and , then , and there is an edge between the two nodes, otherwise, , the connection between the two does not exist; is the threshold for eliminating weak connections.

[0029] Step 4: Use the prior feature matrix of the Pearson correlation coefficient as the node embedding matrix to reduce the irregular steps caused by random initialization during the model training process. The Pearson correlation coefficient has good computational efficiency and can well match the training cost of the subsequent adaptive adjacency matrix.

[0030] In one implementation, considering the dynamic change of the adjacent matrix, calculate the feature matrix of the T length of N substation data in a rolling manner , as shown in Eqs. (2) and (3).

[0031] (2) (3) In the formula, represents the Pearson correlation coefficient between the measured data of any two substations among the N substations, and respectively represent the average values of the two sets of data. T is the step size required to construct the heterogeneous graph. , are the sampled power values of the i-th and j-th wind farm substations at time t respectively.

[0032] Step 5: Take Added to the embedding space, and set the combined adaptive adjacency matrix as . Meanwhile, a unidirectional learning method is referred to in the graph construction process, which can describe the actual directed spatial relationship between different stations, that is, the change of one wind farm will affect the state of another adjacent station. Obtain the final combined adjacency matrix A ; In view of the randomness and interference existing in τ , a corresponding diagonal correction and sparsification process is required, which can not only retain the actual self-influence of each node, but also separate the intricate mapping relationships of each data pattern on the heterogeneous graph according to the data pattern type. Then, after setting the relevant threshold A to eliminate the weak connections in the definition of the adjacency matrix, the final combined adjacency matrix (4) where .

[0033] Step 6, input the time series data of power, wind speed, wind direction and temperature of five wind farm stations into the heterogeneous spatio-temporal graph convolutional network, and output the fusion result; The heterogeneous spatio-temporal graph convolutional network includes a dynamic heterogeneous graph constructor , a time-related network based on CNN and a space-related network based on GCN ; ; ; In the formula, is the power value at the q - 1 time step before time t; is t the sampling data of the i th station at the q - 1 time step before ; In the formula, refers to the data dimension transformation layer of the single-scale convolutional network; is the time feature exploration layer of the multi-scale convolutional network; is the multi-scale feature splicing layer; , respectively represent and the shared parameter sets of ; is the power value at the q - 1 time step before time t; ; In the formula, is the initial space feature exploration layer, is the neighborhood attention weight learning layer, is the spatial attention weight learning layer; , respectively represent and the shared parameter sets of ; is the final output time feature three-dimensional tensor of all nodes; is the dynamic heterogeneous graph constructor.

[0034] In one implementation, as Figure 3 shown. The internal structure of time-related modeling includes a set of specially designed convolutional network layers, as shown in Equation (5).

[0035] (5) In the formula, refers to the data dimension transformation layer of the single-scale convolutional network; is the time feature exploration layer of the multi-scale convolutional network; is the multi-scale feature splicing layer. , respectively represent and the shared parameter sets of .

[0036] In one implementation, a single-scale convolutional network is used for data dimension transformation, using a F convolutional filter of size 1×1 to operate on the one-dimensional sequence data of each node, and splicing the obtained convolutional output into an q × F sized image. Then, the one-dimensional sequence of each node is changed into a set of two-dimensional images with a length of q . In addition, N the N sets of consecutive images of nodes will be fed into the next CNN layer in parallel and capture time features along the time axis through convolutional kernels of different scales. Due to the design of the 1×1 filter size, this layer can not only enhance the dimension of the original data to assist the multi-scale CNN in capturing time features from a visual perspective, but also retain historical time characteristics. So far, the 3D data tensor , taking the number of nodes N as the channel size and using the 2D image as the time data of the node input, is obtained as shown in Equation (6).

[0037] (6) In one implementation, a multi-scale convolutional network is used for time feature exploration. At In terms of implementation, since N The group of images corresponds to N nodes, so the N channels of the multi-scale convolution operation need to prevent information confusion when capturing time features. Therefore, the present invention uses N groups of convolution filters to operate on each channel simultaneously, which is similar to the depth convolution stage of depthwise separable convolution, to avoid N performing traditional one-to-one time modeling for

[0038] Considering that time patterns with different time granularities can reflect information at different scales, a set of convolution filters of different sizes is given to extract multi-scale time features of each channel in parallel. The set of convolution filters is configured as {2×2, 3×3, 5×5, 6×6, 7×7}, which can better balance short-term and long-term features. Therefore, The function of q × F . The iterative calculation is shown in Equation (7).

[0039] (7) In the formula, represents a convolution filter of size λ × λ × N , where the n th filter operates on the th channel of the input data tensor n to output the n th time feature map . Among them, λ is the configuration parameter of the convolution filter, λ = 2, 3, 5, 6, 7.

[0040] In one implementation, multi-scale feature splicing is performed, and the learned feature of each node is as the output of this layer and is regarded as the initial node feature, and is sent to the dynamic heterogeneous graph in q time steps to further explore the spatial features through the GCN layer. According to the outputs of the above 5 convolution filters, the concatenated time feature tensors of each station and all stations can be described by Equations (8) and (8).

[0041] (8) (9) In the formula, can truncate the multi-scale features to the same size. is the final output time feature three-dimensional tensor of all nodes with the same dimension as The three-dimensional tensor of the final output time features of all nodes with the same dimension as .

[0042] In one implementation, an initial spatial feature is obtained. To fully capture the spatial features of different locations, a GCN with multi-order neighborhoods is used to form Extract the spatial correlation output by the time model. Because the prior information comes not only from the nodes directly associated with the target, but also from the nodes that are far away from the target. In addition, even if the heterogeneous graph contains various data patterns, the proposed GCN can process the entire graph in the same way as processing a homogeneous graph network. Compared with the traditional GCN, the multi-order GCN has a special parameter, that is, the k order neighborhood of the target node, defined as , where represents the minimum number of edges moving in the same direction from m to for pattern The minimum number of edges moving in the same direction from m to for pattern . is the i th node, is the node set of pattern m , is the number of all wind farm stations under pattern m. At the same time, the spatial proximity between the target node and its order neighborhood is calculated by the adjacency matrix k . The adjacency matrix A is a typical first-order neighborhood matrix, and each associated node is a first-order neighborhood of the target node. In addition, the initial spatial correlation modeling can be described by equations (10) to (14).

[0043] (10) (11) (12) (13) (15) In the formula, represents the order spatial feature tensor of all nodes captured in from k , and its dimension is the same as . represents a diagonalizable degree matrix, which measures the sum of the spatial proximities of the k order neighborhoods associated with the target node, and is the diagonal element. For normalization to avoid weight explosion or vanishing. The initial spatial feature block consists of q all k order spatial feature tensors at Figure 4 time points. As shown, the first channel of this module represents the current input of all nodes.

[0044] In one implementation, to capture neighborhood attention weights and extract more useful spatial features from the initial spatial feature block, the neighborhood attention weight learning layer aims to guide the GCN to learn the positions to focus on or forget along the channel dimension K as shown in Figure 5 (a). To obtain the attention weights for each layer of this layer, the layer first performs average pooling operations q and max pooling operations F in parallel for each channel along the height N , width and depth dimensions to integrate the feature tensors, discard redundant features and obtain a comprehensive representation of the spatial information. Then, after a multi-layer perceptron (MLP) and sigmoid activation operation , the two pooled feature vectors are element-wise summed to assign neighborhood attention weights. The neighborhood attention learning mechanism of

[0045] (15) In addition, the initial spatial feature block can be aggregated according to the neighborhood attention weights as shown in equation (16).

[0046] (16) where refers to the refined spatial feature tensor weighted by neighborhood attention. is the k th channel tensor of

[0047] In one implementation, to capture spatial attention weights and further refine and emphasize the spatial features of each data pattern from the output, the spatial attention weight learning layer is deployed along the order of the spatial dimensions of to learn what to focus on or suppress, as shown in Figure 5As shown in (b). Due to the characteristics of the heterogeneous graph, the spatial features of different data patterns should be extracted separately. Different from the previous layer, in order to automatically obtain the spatial attention weights covering all data patterns, this layer aggregates the feature information on the width F of through the max-pooling operation and the average-pooling operation, concatenates the two pooled features obtained, and then performs convolution to output the spatial attention weights . The spatial attention learning mechanism of

[0048] is shown in Eqs. (17) and (18). (18) In the formula, represents the convolution operation, and is the final output of the neighborhood and local spatial weights.

[0049] The heterogeneous spatio-temporal graph convolutional network of this embodiment is a trained network, and the training process is as shown in Figure 1 , which will not be elaborated here.

[0050] This embodiment also provides a wind power cluster data fusion system based on a heterogeneous spatio-temporal graph convolutional network, including: One or more processors; A storage device for storing one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the wind power cluster data fusion method based on the heterogeneous spatio-temporal graph convolutional network.

[0051] The following further elaborates on the present invention through specific experiments.

[0052] To evaluate the cluster multi-source data fusion effect of the present invention, the power, wind speed, wind direction, and temperature data of 35,040 sampling time points in a year of five cluster wind farms in a certain area are used as a data set for wind power cluster data fusion and prediction. The sampling frequency is 15 min / time, and the training set and test set are divided according to 8:2. A data set H of 4 data patterns from 5 stations is obtained for subsequent comparative experiments of HSTGCN, STGCN, and CNN-LSTM.

[0053] The present invention uses the root mean square error and percentage error indicators to respectively test the generalization ability and data fusion effect of the HSTGCN network model.

[0054] The root mean square error (RMSE) is used to measure the deviation between the predicted value and the true value: ; In the formula,n is the test set data volume, i is the power moment value, is the model prediction value, is the actual power value of the test set.

[0055] The calculation formula for percentage error (PE): ; In the formula, is the percentage error of the predicted value of the cluster wind farm at the i moment, is the predicted value of the power of the cluster wind farm by the HSTGCN network at the i moment, is the actual power value of the cluster wind farm at the i-th moment of the system.

[0056] (1) Generalization performance analysis Use 80% of the dataset H for training respectively. For the trained HSTGCN, STGCN, and CNN-LSTM networks, randomly select 10 groups of test data from the 20% test dataset to conduct generalization performance analysis on the networks. The root mean square errors of the HSTGCN, STGCN, and CNN-LSTM networks in the test data are shown in Table 1.

[0057] Figure 6 Based on the experimental data in Table 1, conduct visual reproduction, and intuitively reveal the RMSE distribution laws of the HSTGCN, STGCN, and CNN-LSTM networks on 10 groups of test sets through curve comparison.

[0058] Table 1 Root mean square error table of different network models for predicting dataset H

[0059] From Table 1 and Figure 6 it can be seen that CNN-LSTM only fuses and extracts the time series features of the data, resulting in the largest error; at the same time, although STGCN fuses the spatio-temporal features of the data, due to the lack of construction of the dynamic heterogeneous graph of multi-source data, its stability is poor during the prediction process, resulting in error distribution; compared with the STGCN and CNN-LSTM networks, HSTGCN not only fuses multi-source data, but also models the dynamic spatio-temporal correlation of the data. Therefore, the error of the cluster prediction result after fusing wind power data is the smallest and stable. The average RMSE value of the cluster prediction using dataset H is reduced by 2.16% and 3.31%, and it has good generalization performance for wind power cluster data fusion.

[0060] (2) Data fusion quality analysis Perform HSTGCN network control verification on the fused data and the unfused data. For the first group, build the HSTGCN network and train it using the dataset H for fusion. For the second group, only use the SCADA power data, also build the HSTGCN network and train it using the power data in the dataset H, and select 50 groups of consecutive sampling points to obtain the prediction results of the cluster wind farm station as Figure 7 shown.

[0061] Further calculate the percentage errors of the prediction results of the cluster wind farm stations in the datasets H and K as Figure 8 shown, and the average prediction percentage errors of these 50 groups are shown in Table 2.

[0062] Table 2 Average prediction percentage error table of different network models

[0063] From Figure 7 , 8 and Table 2, it can be seen that compared with directly performing wind power cluster prediction using only SCADA power data, when using the fused data of SCADA, NWP, and geographical location information for wind power cluster prediction, the average percentage error of the wind farm station cluster prediction has decreased by 2.33%; it shows that the fusion method proposed in this paper can effectively extract the dynamic spatio-temporal characteristics of the data between the stations, thereby improving the cluster prediction accuracy.

[0064] It should be understood that the described embodiments above are part of the embodiments of the present invention, rather than all of the embodiments. In addition, the technical features in each embodiment or individual embodiment provided by the present invention can be combined with each other arbitrarily to form a feasible technical solution. This combination is not restricted by the order of steps and / or the structural composition mode, but must be based on what can be achieved by those of ordinary skill in the art. When the combination of technical solutions appears to be contradictory or unable to be implemented, it should be considered that this combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0065] It should be understood that the above description of the preferred embodiment is relatively detailed, and it should not be considered as a limitation to the protection scope of the patent of the present invention. Under the inspiration of the present invention, those of ordinary skill in the art can also make substitutions or deformations without departing from the protection scope defined by the claims of the present invention, and all fall within the protection scope of the present invention. The scope of protection requested by the present invention shall be subject to the appended claims.

Claims

1. A wind power cluster data fusion method based on a heterogeneous spatio-temporal graph convolutional network, characterized in that Including the following steps: Step 1: Obtain N the original data of a wind farm station, including power, wind speed, wind direction and temperature, and perform normalization processing; Step 2: Obtain the heterogeneous graph , where M represents the number of data patterns in the heterogeneous graph, and V and E are the sets of nodes and edges respectively; the matrix = } is used to describe the spatial proximity of any two nodes with a connecting edge , if , for and , then , and there is an edge between the two nodes, otherwise, , the connection between the two does not exist; is the threshold for eliminating weak connections; Step 3: Use the prior feature matrix of the Pearson correlation coefficient as the node embedding matrix, and calculate the feature matrix of the T of the N field station data in a rolling manner ; where T is the step size required to construct the heterogeneous graph; Step 4: Add to the embedding space and set the combined adaptive adjacency matrix to , obtaining the final combined adjacency matrix A ; Step 5: Input the time series data of the power, wind speed, wind direction, and temperature of N wind farm stations into the heterogeneous spatio-temporal graph convolutional network, and output the fusion result; The heterogeneous spatio-temporal graph convolutional network includes a dynamic heterogeneous graph constructor , a time-related network based on CNN and a space-related network based on GCN ; ; ; In the formula, is the power value at the (q - 1)-th time step before time t; is t the sampling data of the i -th station at the (q - 1)-th time step before time ； In the formula, refers to the data dimension transformation layer of the single-scale convolutional network; is the temporal feature exploration layer of the multi-scale convolutional network; is the multi-scale feature concatenation layer; 、 respectively represent and the shared parameter sets of ; is the power value at the q - 1 time steps before the t-th moment; ； In the formula, is the initial spatial feature exploration layer, is the neighborhood attention weight learning layer, is the spatial attention weight learning layer; , respectively represent and the shared parameter sets of, ; is the final output time feature three-dimensional tensor of all nodes.

2. The method for fusing wind power cluster data based on a heterogeneous spatio-temporal graph convolutional network according to claim 1, wherein: In step 3, ； ； In the formula, represents N the Pearson correlation coefficient between the measured data of any two of the i, j sites; , are the sampled power values of the i-th and j-th wind farms at time t, and represent the average values of the two sets of data respectively, .

3. The method for wind power cluster data fusion based on a heterogeneous spatio-temporal graph convolutional network according to claim 1, wherein: In step 4, .

4. The wind power cluster data fusion method based on the heterogeneous spatio-temporal graph convolutional network according to claim 1, wherein: In step 5, a 1×1 sized F convolution filter operates on the one-dimensional sequence data of each node, and the obtained convolution outputs are concatenated into an image of q × F size; then, the one-dimensional sequence of each node is changed into a group of two-dimensional images with a length of q ; in addition, N groups of consecutive images of N nodes will be fed into the next CNN layer in parallel and capture temporal features along the time axis through convolutional kernels of different scales; Taking the number of nodes N as the channel size and using the 2D image as the time data of the node input to obtain a 3D data tensor , .

5. The method for fusing wind power cluster data based on a heterogeneous spatio-temporal graph convolutional network according to claim 4, wherein: In step 5, Use N Group convolution filters to operate on each channel simultaneously, and adopt a group of convolution filters with different sizes to extract multi-scale temporal features of each channel in parallel; consisting of a series of single-scale convolutional operations, where each convolutional scale is applied along the image, and the size of each node is q × F ; ； In the formula, represents a convolution filter of size λ × λ × N , where the n -th convolution filter operates on the -th channel of the input data tensor to output the n -th temporal feature map n ; ; λ are the configuration parameters of the convolution filter.

6. The method for fusing wind power cluster data based on a heterogeneous spatio-temporal graph convolutional network according to claim 5, wherein: In step 5, the learned features of each node are used as the output of this layer and regarded as the initial node features, and q input into the dynamic heterogeneous graph in units of time steps to further explore the spatial features through the GCN-based spatial correlation network; The concatenated time feature tensors of each station and all stations are respectively: ； ； In the formula, Truncate the multi-scale features to the same size, is the final output temporal feature three-dimensional tensor of all nodes with the same dimension as The above translation is based on the best understanding of the context. If there are any specific requirements or corrections needed, please feel free to let me know. The final output temporal feature three-dimensional tensor of all nodes with the same dimension as 7. The method for fusing wind power cluster data based on a heterogeneous spatio-temporal graph convolutional network according to claim 6, characterized in that: In step 5, a multi-order neighborhood GCN is used to form Extract a CNN-based time-related network The output spatial correlation; The multi - order neighborhood GCN has a special parameter, namely the k - order neighborhood of the target node, which is defined as , where represents the minimum number of edges moving in the same direction from m to in pattern . is the i - th node, is the set of nodes in pattern m ; is the total number of all wind farm stations under pattern m; meanwhile, the spatial proximity between the target node and its - order neighborhood is calculated by the adjacency matrix k . Described by formulas (1) to (5): （1） （2） （3） （4） （5） In the formula, represents all nodes' captured from at k order spatial feature tensors, whose dimensions are the same as ; represents a diagonalizable degree matrix that measures the sum of spatial proximities of the k order neighborhood related to the target node, and is the diagonal element of ; is used to normalize to avoid weight explosion or vanishing; the initial spatial feature block is all q time points including k order spatial feature tensors; 's first channel represents the current input of all nodes.

8. The method for fusing wind power cluster data based on a heterogeneous spatio-temporal graph convolutional network according to claim 7, wherein: In step 5, the neighborhood attention weight learning layer aims to guide the GCN along the channel dimension K to learn the positions to focus on or forget; to obtain the attention weights for each layer of this layer, the layer first performs average pooling operations q and maximum pooling operations F in parallel for each channel along the height N , width and depth dimensions to integrate the feature tensors, discard redundant features and obtain a comprehensive representation of the spatial information; then, after a multi-layer perceptron and sigmoid activation operations , the two pooled feature vectors are element-wise summed to assign neighborhood attention weights; The neighborhood attention learning mechanism is as follows: ； In addition, the initial spatial feature block is aggregated according to the neighborhood attention weights: ； wherein, refers to the refined spatial feature tensor weighted by neighborhood attention, is the k th channel tensor.

9. The method for fusing wind power cluster data based on a heterogeneous spatio-temporal graph convolutional network according to claim 8, wherein: In step 5, aggregate the feature information on the width F through max pooling operation and average pooling operation, connect the two resulting pooling features, and then perform convolution to output the spatial attention weights ; ; The spatial attention learning mechanism is as follows: ； ； In the formula, represents a convolution operation, is the final output of the neighborhood and local spatial weights.

10. A wind power cluster data fusion system based on a heterogeneous spatio-temporal graph convolutional network, characterized in that, Including: One or more processors; A storage device for storing one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the wind power cluster data fusion method based on the heterogeneous spatio-temporal graph convolutional network according to any one of claims 1 to 9.

Images