A power prediction method for a gas power station

Abstract

The present invention discloses a method for predicting the power of a gas power station, specifically as follows: Step 1: Collect the historical data of the gas power station as sample data; Step 2: Preprocess the sample data; Step 3: Construct the preprocessed data into a spatio-temporal graph H; Step 4: Design an adaptive and scalable spatio-temporal graph convolutional network model; Step 5: Use the preprocessed data and the adjacency matrix as the training set to train the adaptive and scalable spatio-temporal graph convolutional network model; Step 6: Use the trained adaptive and scalable spatio-temporal graph convolutional network model to predict the power of the gas power station. The method proposed by the invention shows higher accuracy, better stability, stronger trend tracking ability and better outlier handling ability when predicting the gas power generation.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas power stations, and particularly relates to a method for predicting the power of a gas power station. Background Art

[0002] Gas power stations play an important role in modern power systems. Accurate prediction of their power output is of great significance for the stable operation and optimal dispatching of power systems. However, the power output of gas power stations is affected by various complex factors, including weather conditions, gas supply, coal mine production, energy demand, etc. There are complex non-linear relationships and spatio-temporal dependencies among these factors, which make the number of gas power stations and nodes grow dynamically, and it is difficult to predict their power.

[0003] Traditional methods for predicting the power of gas power stations mainly rely on time series models and machine learning models. These methods include autoregressive integrated moving average model (ARIMA), support vector regression (SVR), etc. Time series models (such as long short-term memory recurrent neural network (LSTM)) predict by analyzing the time dependence of historical data, but their main assumption is the linear relationship of data, and it is difficult to capture the complex non-linear characteristics in the data. In addition, these models often ignore the spatial correlation of the power output of gas power stations, resulting in limited prediction accuracy; at the same time, most of these models adjust model parameters manually, and their adaptability and scalability for power prediction in the case of dynamic growth of the number of gas power stations and nodes are poor.

[0004] In recent years, with the development of deep learning technology, graph convolutional neural network (GCN) and spatio-temporal graph convolutional neural network (ST-GCN) have been widely used in fields such as traffic flow prediction and power load prediction. GCN can effectively capture the spatial dependence relationship in data through convolutional operations on the graph structure; ST-GCN can capture both the time characteristics and spatial characteristics of data by combining time convolution and graph convolution. Although GCN and ST-GCN perform well in traffic flow and power load prediction, there are no examples of using GCN and ST-GCN for prediction in the field of gas power station power prediction. In addition, the power output of gas power stations is not only affected by external factors such as weather and gas supply, but also has significant spatio-temporal heterogeneity and dynamics. Existing methods still have limitations in dealing with these unique spatio-temporal heterogeneity and dynamics. Summary of the Invention

[0005] Object of the Invention: To solve the problems existing in the above-mentioned prior art, the present invention provides a method for predicting the power of a gas power station.

[0006] Technical Solution: The present invention provides a method for predicting the power of a gas power station, which specifically includes the following steps:

[0007] Step 1: Collect the historical data of the gas power station as sample data;

[0008] Step 2: Preprocess the sample data;

[0009] Step 3: Construct a spatio-temporal graph H from the preprocessed data; H = (V, E, A), where V represents the set of nodes, the nodes represent the gas power stations, E represents the set of edges, the edges are the connecting edges between nodes, and A is the adjacency matrix;

[0010] Step 4: Design an adaptive and scalable spatio-temporal graph convolutional network model;

[0011] Step 5: Use the preprocessed data and the adjacency matrix as the training set to train the adaptive and scalable spatio-temporal graph convolutional network model;

[0012] Step 6: Use the trained adaptive and scalable spatio-temporal graph convolutional network model to predict the power of the gas power station.

[0013] Further, the sample data in Step 1 includes historical power, weather data related to power generation, and gas supply data.

[0014] Further, the data preprocessing in Step 2 is specifically to first perform data cleaning on the data and then perform normalization processing. The data cleaning is specifically: using the forward filling or backward filling method to process the missing power data; then using the Z-score method to detect and process outliers; converting the timestamp to the datetime type.

[0015] Further, in Step 3, a Gaussian kernel function is used to construct the adjacency matrix, and the expression of the adjacency matrix A is:

[0016]

[0017] where i and j respectively represent nodes i and j in the set of nodes; p i and p j respectively represent the geographical locations of nodes i and j, d(i, j) is the correlation coefficient between the historical powers of nodes i and j, and α and κ are parameters controlling the sparsity of the adjacency matrix.

[0018] Further, the adaptive and scalable spatio-temporal graph convolutional network model is configured to perform the following operations:

[0019] Use spatio-temporal convolutional blocks to capture the spatio-temporal features of the gas power station, and improve the feature extraction ability by stacking L spatio-temporal convolutional blocks; then use an adaptive module to adjust the model parameters; then use an extensible module to expand according to the increase in the gas power generation data volume and the number of nodes, and finally output the predicted power of the gas power station through a fully connected layer.

[0020] Furthermore, each spatio-temporal convolutional block uses graph convolution operations to extract the features of each node and its neighborhood in the gas power station, uses one-dimensional convolution operations in temporal convolution to extract temporal features, and simultaneously uses gated linear units in temporal convolution to capture the temporal dynamic behavior of the gas power station. The structure of the spatio-temporal convolutional block is as follows:

[0021]

[0022] Among them, represents the output feature of the (l + 1)-th spatio-temporal convolutional block, TemporalConv(.) represents the temporal convolution operation, σ(.) represents the activation function, GraphConv(.) represents the graph convolution operation, and H (l) represents the output feature of the l-th graph convolution, and Z (l) represents the output feature of the l-th temporal convolution, represents the bias of the one-dimensional convolution in the l-th temporal convolution, represents the normalized adjacency matrix, where J represents the degree matrix, and Θ (l) represents the weight matrix of the l-th graph convolution; l = 0, 1, 2,..., L - 1, X is the sample data;

[0023] The structure of the graph convolution is as follows:

[0024]

[0025] Among them, H (l+1) represents the output feature of the (l + 1)-th graph convolution, T k represents the k-th order Chebyshev polynomial, and K represents the total order of the Chebyshev polynomial; represents the normalized graph Laplacian, I represents the identity matrix, and Θ k represents the weight matrix of the k-th order Chebyshev polynomial in the l-th graph convolution. The weight matrices of K Chebyshev polynomials in the l-th graph convolution form the weight matrix Θ (l) ;

[0026] The structure of the temporal convolution is as follows:

[0027]

[0028] Among them, Z (l+1) represents the output feature of the (l + 1)-th temporal convolution, represents the output feature of the one-dimensional convolution in the l-th temporal convolution, denotes the convolution kernel of the one-dimensional convolution in the (l + 1)-th layer of the temporal convolution, denotes the bias of the one-dimensional convolution in the (l + 1)-th layer of the temporal convolution; denotes the convolution kernel of the gated linear unit in the (l + 1)-th layer of the temporal convolution, denotes the bias of the gated linear unit in the (l + 1)-th layer of the temporal convolution, and * represents the convolution operation.

[0029] Furthermore, the adaptive module includes M adaptive layers connected in sequence. The adaptive layer includes node feature transformation, adaptive attention weight calculation, and attention weight normalization;

[0030] The node feature transformation is specifically:

[0031] F (m+1) = σ(B (m+1) F (m) )

[0032] where F (m+1) denotes the node feature matrix output by the (m + 1)-th layer of the adaptive layer, and the node feature matrix output by the first layer of the adaptive layer where denotes the feature output by the spatio-temporal convolution module, B (1) denotes the weight matrix of the first layer of the adaptive layer, and σ(.) represents the activation function; B (m+1) denotes the weight matrix of the (m + 1)-th layer of the adaptive layer, m = 1, 2, 3, …, M; M represents the total number of adaptive layers;

[0033] The adaptive attention weight calculation is specifically:

[0034]

[0035] where, denotes the adaptive attention weight between node i and node j in the m-th layer of the adaptive layer; LeakyReLU represents the leaky rectified linear activation function, || represents vector concatenation, and respectively denote the features of node i and node j in the m-th layer of the adaptive layer; B (m) denotes the weight matrix of the m-th layer of the adaptive layer, and a (m) is the set of normalized adaptive attention weights in the m-th layer of the adaptive layer;

[0036] The attention weight normalization is specifically:

[0037]

[0038] where, denotes the value after normalization of the adaptive attention weight between node $i$ and node $j$ in the $m$-th adaptive layer. $V(i)$ represents the set of adjacent nodes of node $i$, and $|V(i)|$ represents the total number of nodes in the set $V(i)$. denotes the adaptive attention weight between node $i$ and node $k'$ in the set $V(i)$ in the $m$-th adaptive layer.

[0039] Furthermore, the scalable module includes $N$ layers of scalable layers connected in sequence. The scalable layer includes multi-head graph convolution and hierarchical aggregation. The multi-head graph convolution is as follows:

[0040]

[0041] where represents the feature matrix of node $i$ output by the $(n + 1)$-th layer of multi-head graph convolution. $Q$ represents the number of attention heads, and $B$ (n+1,q) represents the weight matrix of the $q$-th attention head in the $(n + 1)$-th layer of multi-head graph convolution. represents the feature matrix of node $i$ output by the $n$-th layer of multi-head graph convolution. The normalized adaptive attention $a$ in the $M$-th adaptive layer M is decomposed into $Q$ attention heads, and these $Q$ attention heads are used as the $Q$ attention heads in the first layer of multi-head graph convolution. The feature matrix of node $i$ output by the first layer of multi-head graph convolution has the expression represents the attention weight between node $i$ and node $j$ calculated by the $q$-th attention head in the first layer of multi-head graph convolution. $F$ (M) represents the node feature matrix output by the $M$-th adaptive layer, where $n = 1, 2, \ldots, N$.

[0042] An electronic device / system for a gas power station power prediction method includes a processor and a memory. The memory stores the execution instructions of the processor, and the processor is configured to execute the execution instructions to implement the gas power station power prediction method.

[0043] A computer-readable storage medium is used to store a program, and executing the program implements the gas power station power prediction method.

[0044] Beneficial effects: The present invention combines the advantages of graph convolutional neural network and temporal convolutional network, and can effectively capture the complex spatio-temporal dependence relationships in the gas power station power data. At the same time, an adaptive module is introduced to dynamically adjust the model parameters to adapt to the changing characteristics of gas power generation. The introduction of a scalable module enables the model to be extended as the data volume and the number of nodes in the gas power station increase, improving the gas power station power prediction accuracy and enhancing the adaptability of the model. Brief Description of the Drawings

[0045] Figure 1 This is the flowchart of the method of the present invention.

[0046] Figure 2 This is the block diagram of the method of the present invention.

[0047] Figure 3 This is the comparison chart of the effects of three prediction methods. Detailed implementation mode

[0048] The accompanying drawings that form a part of the present invention are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0049] As Figure 1 , as shown in FIG. 2, the gas power station power prediction method provided by the present invention includes the following steps:

[0050] 1) Collect historical data of the gas power station, and perform data cleaning and normalization processing. The historical data includes power output data, weather data, and gas supply data;

[0051] For this embodiment, first collect the historical power output data of the gas power station, and record the power output value per hour or per minute; collect weather data related to power generation, including temperature, humidity, wind speed, rainfall, and air pressure, etc.; obtain the data of gas supply, including gas supply volume, pressure, and temperature, etc. Secondly, perform data cleaning on the collected data, and use the method of forward filling or backward filling to process the missing power data; use the Z-score method to detect and process outliers. Convert the time stamp to the datetime type to facilitate time series processing. Finally, use the min-max normalization method to perform normalization processing on the data, and perform standardization processing on the output data.

[0052] (1) The collection of historical data of the gas power station can be expressed as:

[0053] D = {P, W, G}

[0054] Where D represents the historical data set; P represents the power output data; W represents the weather data; G represents the gas supply data.

[0055] (2) The data cleaning and normalization processing is expressed as:

[0056] D clean = f clean (D)

[0057] D norm = f norm (D clean )

[0058] Where Dclean Represents a data cleaning operation; f clean Represents a data cleaning function; D represents the historical data set, D norm Represents a data normalization operation; f norm Represents a data normalization function.

[0059] (3) The data normalization process described above also includes:

[0060]

[0061] Among them, x′ represents the normalized data; x represents the original data; min(x) represents the minimum value of the data; max(x) represents the maximum value of the data.

[0062] 2) Construct the different data sources of the gas power station into a spatio-temporal graph, and construct an adjacency matrix to represent the spatial relationship between nodes.

[0063] (1) The construction of the different data sources of the gas power station into a spatio-temporal graph described above can be expressed as:

[0064] H = (V, E, A)

[0065] Among them, H represents the spatio-temporal graph of the gas power station; V represents the set of nodes, and a gas power station represents a node; E represents the set of edges, and the edges represent the spatio-temporal relationship between time steps; A represents the spatio-temporal correlation between gas power stations.

[0066] (2) The construction of an adjacency matrix to represent the spatial relationship between nodes can be expressed as:

[0067] Use the Gaussian kernel function to construct the element A in the adjacency matrix ij :

[0068]

[0069] Among them, A ij represents the element in the i-th row and j-th column of the adjacency matrix A, and i, j represent nodes i, j respectively; p i and p j represent the geographical locations of nodes i and j respectively, d(i, j) is the correlation coefficient between the historical powers of nodes i and j, and α and κ are parameters that control the sparsity of the adjacency matrix.

[0070] 3) Conduct the design of an adaptive scalable spatio-temporal graph convolutional network model, including an input layer, a spatio-temporal convolution module, an adaptive module, a scalable module, and a fully connected layer;

[0071] In the input layer, the preprocessed data is input into the model, including historical power data, weather data, and gas supply data; in the spatio-temporal convolution module, temporal convolution and graph convolution are respectively performed to capture the correlations in time and space; an adaptive module is introduced to dynamically adjust the model parameters to adapt to the changing characteristics of gas power generation; a scalable module is designed to enable the model to scale as the data volume and the number of nodes increase, maintaining efficient computation.

[0072] (1) The input layer described above can be expressed as:

[0073] The preprocessed data is input into the model, including historical power data, weather data, and gas supply data.

[0074] (2) The spatio-temporal convolution module includes L spatio-temporal convolution blocks connected in sequence:

[0075] The spatio-temporal convolution block uses graph convolution operations to extract the spatial features of gas power generation.

[0076] To ensure the stability of the convolution operation, it is necessary to normalize the adjacency matrix A. The normalized adjacency matrix is

[0077]

[0078] where J represents the degree matrix, J ii = ∑ j A ij .

[0079] Next, calculate the graph Laplacian operator. The normalized graph Laplacian operator is:

[0080]

[0081] where represents the normalized graph Laplacian operator; I represents the identity matrix.

[0082] Adopt Chebyshev polynomial approximation to achieve localized convolution and extract the features of each node and its neighborhood in the gas power station:

[0083]

[0084] where H (l+1) represents the output feature of the (l + 1)-th layer of graph convolution, T k represents the k-th order Chebyshev polynomial, K represents the total order of the Chebyshev polynomial; Θ k represents the weight matrix of the k-th order Chebyshev polynomial in the l-th layer of graph convolution, and H (l) represents the output feature of the l-th layer of graph convolution.

[0085] Then, one-dimensional convolution operation is used to extract temporal features, and gated linear units are used to capture the temporal dynamic behavior of the gas power station.

[0086] The one-dimensional convolution operation can be expressed as:

[0087]

[0088] Combining the above graph convolution operation and temporal convolution operation, a spatio-temporal convolution block is formed to capture the spatio-temporal features of the gas power station, and the feature extraction ability is improved by stacking multiple spatio-temporal convolution blocks. Each spatio-temporal convolution block includes a graph convolution and a temporal convolution. The graph convolution is used to extract spatial features, and the temporal convolution is used to extract temporal features. To enhance the expressive power of the model, multiple spatio-temporal convolution blocks are stacked. The output of each convolution block is used as the input of the next convolution block. And activation function (ReLU) and regularization method (Dropout) are used to prevent overfitting.

[0089] Among them, the spatio-temporal convolution block capturing the spatio-temporal features of the gas power station is expressed as:

[0090]

[0091] Where, represents the feature matrix output by the spatio-temporal convolution block of the (l + 1)-th layer; Θ (l) represents the weight matrix of the graph convolution of the l-th layer, and the weight matrices Θ k of K Chebyshev polynomials in the l-th layer graph convolution constitute the weight matrix Θ (l) of the l-th layer graph convolution; Z (l) represents the feature matrix output by the temporal convolution of the l-th layer, and H (l) represents the output feature of the graph convolution of the l-th layer; represents the bias of the one-dimensional convolution in the temporal convolution of the l-th layer, GraphConv(.) represents the graph convolution operation, and TemporalConv(.) represents the temporal domain convolution operation, l = 0, 1, 2,..., L - 1.

[0092] (3) The adaptive module can be expressed as: Dynamically adjust the model parameters through the adaptive graph attention mechanism to adapt to the changing characteristics of gas power generation. The adaptive module includes node feature transformation, adaptive attention weight calculation, and attention weight normalization processes.

[0093] Among them, the node feature transformation can be expressed as:

[0094]

[0095] F (m+1) = σ(B(m+1) F (m) )

[0096] Among them, F (1) , F (m) , F (m+1) respectively represent the node feature matrices output by the 1st, m-th, and (m + 1)-th adaptive layers; B (1) , B (m+1) represent the weight matrices of the 1st and (m + 1)-th adaptive layers.

[0097] Adaptive attention weight calculation:

[0098]

[0099] Among them, represents the adaptive attention weight between node i and node j in the m-th adaptive layer; LeakyReLU represents the leaky rectified linear activation function; a (m) is the set of normalized adaptive attention weights in the m-th adaptive layer, and the features of node i and node j are concatenated together; and respectively represent the features of node i and node j in the m-th adaptive layer, and B (m) represents the weight matrix of the m-th adaptive layer.

[0100] Attention weight normalization:

[0101]

[0102] Among them, represents the value after normalization of the adaptive attention weight between node i and node j in the m-th adaptive layer, V(i) represents the set of adjacent nodes of node i, and |V(i)| represents the total number of nodes in the set V(i), represents the adaptive attention weight between node i and node k' in the set V(i) in the m-th adaptive layer.

[0103] Through this adaptive attention mechanism, the model can dynamically adjust the importance of each node in the graph, so as to adapt to the changing characteristics of gas power generation.

[0104] The expandable module described in (4) can be expressed as: The expandable module includes N expandable layers, and the expandable layers enable the model to expand as the amount of gas power generation data and the number of nodes increase through multi-head graph convolution and hierarchical graph convolution, maintaining efficient computing.

[0105] Through the multi-head attention mechanism, the model can capture different graph structure information, thereby enhancing the expressive power and scalability of the model. Among them, the multi-head graph convolution can be expressed as:

[0106]

[0107] Among them, respectively represent the feature matrices of node i in the multi-head graph convolution of the 1st, nth, and (n + 1)th layers; Q represents the number of attention heads; the normalized adaptive attention a in the Mth adaptive layer (M) is decomposed into Q attention heads, and these Q attention heads are used as the Q attention heads in the multi-head graph convolution of the 1st layer, so that each attention head focuses on different aspects or features of the input data. represents the attention weight between node i and node j calculated by the qth attention head in the multi-head graph convolution of the 1st layer, F (M) represents the node feature matrix output by the Mth adaptive layer, n = 1, 2,..., N, B (n+1，q) represents the weight matrix of the qth attention head in the multi-head graph convolution of the (n + 1)th layer.

[0108] Through hierarchical aggregation, the model can capture high-order information in the graph structure layer by layer, thus achieving the scalability of the model.

[0109] Hierarchical aggregation can be expressed as:

[0110]

[0111] Finally, a fully connected layer is used to output the predicted power.

[0112] 4) Train the model in step 3;

[0113] Evaluate the prediction accuracy by defining the mean squared error (MSE) and root mean squared error (RMSE):

[0114]

[0115] Among them, n1 represents the number of samples, represents the predicted value of the i1th sample; y i1 represents the actual value of the i1th sample.

[0116] Use historical data for model training, input training data, perform forward propagation through the spatio-temporal convolution layer, adaptive module, scalable module, and output layer, and calculate the predicted value. Calculate the loss between the predicted value and the actual value according to the loss function and calculate the gradient of the loss with respect to the model parameters. Use the gradient descent optimization algorithm to update the model parameters:

[0117]

[0118] Among them, θ represents the model parameters, η represents the learning rate, represents the gradient of the loss function.

[0119] Repeat forward propagation, calculate the loss, perform backpropagation, and update the parameters until the model converges or reaches the set maximum number of iterations. Use the validation set data to evaluate the performance of the model and prevent overfitting.

[0120] 5) Input the new gas power generation data into the model in step 4 and output the predicted power value of the gas power station.

[0121] That is, input the new weather data and gas supply data into the model, perform gas power generation power prediction through the trained ASGCN model, and finally output the predicted power value of the gas power station.

[0122] For the above test system, this embodiment respectively uses four estimation methods, namely regression prediction of the traditional least squares (LR) algorithm, regression prediction of the long short-term memory recurrent neural network (LSTM) method, regression prediction of the spatio-temporal graph convolutional neural network (ST-GCN) method, and the adaptive scalable spatio-temporal graph convolutional network (ASGCN) method proposed in the present invention, for simulation and comparative analysis. Table 1 shows the regression prediction effects and algorithm calculation efficiencies of the four methods after 20 simulations.

[0123] Table 1

[0124] Method MSE RMSE Average Inference Time (s) Floating Point Operations per Second LR 10.24 3.20 1.32 / LSTM 6.79 2.61 3.28 2.45 ST-GCN 5.32 2.31 2.51 1.89 ASGCN 3.26 1.81 1.08 0.92

[0125] It can be seen from the data in the table that the traditional least squares estimation method has a large RMSE in the estimation result because it lacks the prediction ability for the dynamic changes of the meteorological and gas power station topological systems, indicating that the estimation result deviates greatly from the true state. The prediction accuracy of the long short-term memory recurrent neural network method has been greatly improved compared with the static estimation. At the same time, the ST-GCN algorithm can better consider the characteristics of each node and its neighborhood of the gas power station, and both the estimation result accuracy and the algorithm calculation efficiency are superior to those of LSTM. The method proposed in the present invention considers the factors of the number of gas power stations and the dynamic growth of nodes on the basis of ST-GCN, greatly improving the accuracy of the estimation result, and because of the improved convolutional kernel structure, it does not overly affect the algorithm calculation efficiency.

[0126] Figure 3 For the comparison chart of the effects of three different prediction methods, from Figure 3It can be seen that in terms of prediction accuracy, the curve of ASGCN (red) is relatively close to the actual value (black), and it can better track the change trend of the actual value in most time periods; in contrast, the fluctuation of LSTM (blue) is relatively large, especially in some time periods, it deviates greatly from the actual value; the predicted values of ST-GCN (purple) deviate more at some peaks and troughs. In terms of stability, the curve of ASGCN is relatively smooth with small fluctuations, showing good stability, which is particularly important for the gas power generation scenario; the curve of LSTM has a large fluctuation amplitude, and the prediction results are sometimes too drastic and less stable than ASGCN; ST-GCN also has large fluctuations in some time periods, and the predicted values are not stable enough. In terms of tracking trends, ASGCN has better tracked the overall trend of the actual value throughout the time period, especially in the middle and late stages; LSTM has not been able to track the actual trend well in some time periods, showing some large deviations; although ST-GCN performs well in some time periods, its overall trend tracking ability is not as good as ASGCN. In terms of outlier handling, ASGCN performs relatively robustly when dealing with outliers (such as the spikes and troughs in the figure) and does not show large deviations; when facing outliers, the curve of LSTM fluctuates greatly, showing insufficient ability to handle outliers; ST-GCN also has large fluctuations at some outliers and fails to handle outliers well.

[0127] In summary, the method proposed in the present invention shows higher accuracy, better stability, stronger trend tracking ability and better outlier handling ability when predicting gas power generation power. At the same time, the introduction of the adaptive and scalable module enables ASGCN to adapt to the changing characteristics of gas power generation, can be extended as the amount of gas power generation data and the number of nodes increase, can handle the nonlinearity and dynamic changes in the power data of gas power stations, and improves the robustness and adaptability of the model, having obvious advantages compared with LSTM and ST-GCN.

[0128] The above is only the preferred solution of the present invention and is not intended to further limit the present invention. All equivalent changes made by using the content of the specification and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A method for predicting power of a gas power plant, characterized in that: The specific steps include: Step 1: Collect historical data of gas power plants as sample data; Step 2: Preprocess the sample data; Step 3: construct the preprocessed data into a spatiotemporal graph H; H = (V, E, A), V represents a node set, the node represents a gas power station, E represents an edge set, the edge is a connection edge between nodes, and A is an adjacency matrix; Step 4: Design an adaptive and scalable spatiotemporal graph convolutional network model; Step 5: Use the preprocessed data and adjacency matrix as training sets to train the adaptive scalable spatiotemporal graph convolutional network model; Step 6: Use the trained adaptive scalable spatiotemporal graph convolutional network model to predict the power of the gas power station; The adaptive scalable spatiotemporal graph convolutional network model is configured to perform the following operations: The spatiotemporal convolutional block is used to capture the spatiotemporal characteristics of the gas power station, and the feature extraction capability is improved by stacking L layers of spatiotemporal convolutional blocks. Then, the adaptive module is used to adjust the model parameters. The scalable module is used to expand according to the increase in the amount of gas power generation data and the number of nodes. Finally, the predicted power of the gas power station is output through the fully connected layer. Each spatiotemporal convolution block uses graph convolution operations to extract the features of each node and neighborhood of the gas power plant, uses the one-dimensional convolution operation in the time convolution to extract the time features, and uses the gated linear unit in the time convolution to capture the temporal dynamic behavior of the gas power plant. The structure of the spatiotemporal convolution block is as follows: in, represents the output features of the l+1th spatiotemporal convolutional block, TemporalConv(.) represents the temporal convolution operation, σ(.) represents the activation function, GraphConv(.) represents the graph convolution operation, and H (l) Represents the output features of the l-th layer of graph convolution, Z (l) represents the output features of the l-th layer of temporal convolution, represents the bias of the one-dimensional convolution in the l-th layer of temporal convolution, represents the normalized adjacency matrix, Where J represents the degree matrix, Θ (l) represents the weight matrix of the l-th layer of graph convolution; l = 0, 1, 2, ..., L-1, X is the sample data; The structure of graph convolution is as follows: Among them, H (l+1) represents the output features of the l+1th layer of graph convolution, T k represents the k-th order Chebyshev polynomial, K represents the total order of the Chebyshev polynomial; represents the normalized graph Laplacian operator, I represents the identity matrix, Θ k Represents the weight matrix of the k-th order Chebyshev polynomial in the l-th layer of graph convolution. The weight matrix of K Chebyshev polynomials in the l-th layer of graph convolution constitutes the weight matrix Θ of the l-th layer of graph convolution. (l) ; The structure of temporal convolution is as follows: Among them, Z (l+1) represents the output features of the l+1th layer of temporal convolution, represents the output feature of the one-dimensional convolution in the l-th layer of temporal convolution, represents the convolution kernel of the one-dimensional convolution in the l+1th layer of temporal convolution, Represents the bias of the one-dimensional convolution in the l+1th layer of temporal convolution; represents the convolution kernel of the gated linear unit in the l+1th layer of temporal convolution, represents the bias of the gated linear unit in the l+1th layer of temporal convolution, and * represents the convolution operation; The adaptive module includes M adaptive layers connected in sequence, wherein the adaptive layers include node feature transformation, adaptive attention weight calculation and attention weight normalization; The node feature transformation is specifically as follows: F (m+1 )=σ(B (m+1) F (m) ) Among them, F (m+1) Represents the node feature matrix output by the m+1th adaptive layer, and the node feature matrix output by the 1st adaptive layer in represents the features of the spatiotemporal convolutional block output, B (1) represents the weight matrix of the first adaptive layer, σ(.) represents the activation function; B (m+1) represents the weight matrix of the m+1th adaptive layer, m=1,2,3,…,M; M represents the total number of adaptive layers; The adaptive attention weight calculation is specifically as follows: in, represents the adaptive attention weight between node i and node j in the mth adaptive layer; LeakyReLU represents the linear rectification activation function with leakage, || represents the vector connection, and Respectively represent the features of node i and j in the mth adaptive layer; B (m) represents the weight matrix of the mth adaptive layer, a (m) is the set of normalized adaptive attention weights in the mth adaptive layer; The attention weight normalization is specifically: in, represents the normalized value of the adaptive attention weight between node i and node j in the mth adaptive layer, V(i) represents the set of adjacent nodes of node i, |V(i)| represents the total number of nodes in the set V(i), represents the adaptive attention weight between node i in the mth adaptive layer and node k′ in the set V(i); The scalable module includes N scalable layers connected in sequence, and the scalable layers include multi-head graph convolution and hierarchical aggregation. The multi-head graph convolution is as follows: in, represents the feature matrix of node i output by the n+1th layer of multi-head graph convolution, Q represents the number of attention heads, and B (n+1，q) represents the weight matrix of the qth attention head in the n+1th layer of multi-head graph convolution, Represents the feature matrix of node i output by the n-th layer multi-head graph convolution, and the normalized adaptive attention a in the M-th adaptive layer M Decompose into Q attention heads, use the Q attention heads as the Q attention heads in the first layer of multi-head graph convolution, and the feature matrix of node i output by the first layer of multi-head graph convolution is The expression is represents the attention weight between node i and node j calculated by the qth attention head in the first layer of multi-head graph convolution, F (M) Represents the node feature matrix output by the Mth adaptive layer, n=1,2,...,N.

2. A method for predicting power of a gas power plant according to claim 1, characterized in that: The sample data in step 1 includes historical power, weather data related to power generation, and gas supply data.

3. A method for predicting power of a gas power plant according to claim 1, characterized in that: The data preprocessing in step 2 specifically includes first cleaning the data and then normalizing it. The data cleaning specifically includes: processing the missing power data using the forward filling or backward filling method; then using the Z-score method to detect and process outliers; and converting the timestamp into a datetime type.

4. A method for predicting power of a gas power plant according to claim 1, characterized in that: In step 3, a Gaussian kernel function is used to construct an adjacency matrix, and the expression of the adjacency matrix A is: Where i, j represent nodes i, j in the node set respectively; p i and p j They represent the geographical locations of nodes i and j respectively, d(i,j) is the correlation coefficient between the historical powers of nodes i and j, and α and κ are parameters that control the sparsity of the adjacency matrix.

5. An electronic device for a gas power plant power prediction method, characterized in that: The method comprises a processor and a memory, wherein the memory stores execution instructions of the processor, and the processor is configured to execute the execution instructions to implement the power prediction method for a gas power station as claimed in any one of claims 1 to 4.

6. A computer-readable storage medium for storing a program, characterized in that: The program is executed to implement the power prediction method for a gas power plant as described in any one of claims 1-4.

Images