Power load prediction method based on space-time diagram convolutional network in extreme weather

Abstract

The invention provides a power load prediction method based on a space-time diagram convolutional network in extreme weather. Comprising the following steps: collecting historical load data and regional meteorological element data of a plurality of load nodes in a power system, and screening key meteorological characteristics which have obvious influence on loads through a mode of combining model interpretation and regression analysis to construct a meteorological characteristic vector; multivariable empirical mode decomposition and singular value decomposition are adopted to carry out multi-scale reconstruction on load data, and smooth and effective load feature tensors are extracted. On the basis, a graph network structure is constructed in combination with a node physical connection relationship, and load and meteorological characteristics are fused in a time dimension to form node time sequence characteristics. And predicting the load by using the space-time diagram convolutional network model. According to the method, the space-time dependency relationship of the load data can be effectively modeled, the prediction accuracy of the load change under the extreme weather condition is enhanced, and the method has good robustness and generalization ability.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load forecasting, and specifically relates to a power load forecasting method based on a spatiotemporal graph convolutional network under extreme weather conditions. Background Art

[0002] When extreme weather events occur, such as high temperatures or severe cold, power load tends to surge, creating new challenges for short-term load forecasting. Daily activities and meteorological factors contribute significantly to short-term load volatility and uncertainty, significantly impacting the accuracy of load forecasting. Power load is influenced by factors such as season, climate, holiday schedules, region, and economic policies, exhibiting complex variations. However, existing short-term load forecasting methods have limited ability to predict the nonlinear and dynamically changing load characteristics under extreme weather conditions. Accurate and stable short-term load forecasting is crucial for the daily operation of power systems and forms the foundation for building a new power system. It can provide a basis for decision-making in addressing power supply and demand imbalances. Currently, load forecasting methods primarily include traditional mathematical statistics-based methods and data-driven machine learning-based methods. Traditional statistical methods include time series analysis using autoregressive integrated moving averages, multivariate linear regression analysis, and exponential smoothing. While statistical methods have simple model structures and fast computational speed, they perform poorly for forecasting load series with large fluctuations or less pronounced periodicity.

[0003] Deep learning algorithms such as convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and bidirectional long short-term memory networks (BiLSTMs) have been widely used in load forecasting, demonstrating promising performance. However, these methods generally only consider the characteristics of time series data, ignoring the spatial correlations between nodes in the power system. This makes it difficult to fully capture the complex spatiotemporal dependencies of load changes under extreme weather conditions. Furthermore, power load data under extreme weather conditions is highly volatile and non-stationary, making it difficult to achieve satisfactory forecasting accuracy by directly applying deep learning models.

[0004] In the prior art, Chinese patent CN118428544A discloses a method and related device for periodic multi-node power load forecasting, including: obtaining historical load and influencing factor data of the load node to be predicted and adjacent nodes, and constructing a multi-node power load original data set; then preprocessing to obtain a multi-node load sample data set, screening influencing factors from the multi-node load sample data set, and constructing a final multi-node load sample data set; then dividing the data according to working days and holidays to construct a periodic multi-node load sample data set; then fusing a graph convolutional network and a bidirectional gated recurrent unit to build an initial periodic multi-node power load forecasting model; finally, optimizing the hyperparameters of the initial periodic multi-node power load forecasting model based on the data set, thereby determining the final periodic multi-node power load forecasting model.

[0005] However, this method has the following limitations: 1. The screening of influencing factors is based only on correlation coefficient measurement, ignoring the possible nonlinear and causal relationship between meteorological factors and load changes, and cannot ensure that the selected features still have sufficient explanatory power under extreme conditions; 2. Although a graph convolutional network is used to capture the spatial structure, this spatial model is connected in series with a bidirectional gated recurrent unit, and spatial and temporal features are processed separately, making it difficult to achieve joint spatiotemporal feature fusion under multi-node and full topology structures, especially when the physical connectivity between load nodes changes dynamically. The adaptability is insufficient.

[0006] Therefore, the existing technology still lacks a short-term power load forecasting method that can effectively identify nonlinear and causally related key meteorological characteristics under extreme weather conditions, realize deep fusion of multi-node full-topology spatiotemporal characteristics on the real topology structure of the power grid, and be sufficiently robust to the strong volatility and multi-scale dynamic changes of the load sequence. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a power load forecasting method based on spatiotemporal graph convolutional network under extreme weather conditions.

[0008] The purpose of the present invention can be achieved by the following technical solutions: On one hand, the present invention provides a method for power load forecasting under extreme weather conditions based on a spatiotemporal graph convolutional network, including the following features: Collect historical load data and regional meteorological data of multiple load nodes in the power system; Based on regional meteorological element data, the characteristic contribution value of each meteorological feature is calculated by combining model interpretation with regression analysis, and several key meteorological features are selected to construct a meteorological characteristic vector; Perform multivariate empirical mode decomposition on the historical load data of multiple load nodes in the power system, extract multi-order intrinsic mode function components, perform singular decomposition on each order modal component, construct a low-rank approximate embedding matrix, and obtain the load characteristic tensor after multi-scale smooth reconstruction; Based on the physical connection relationship of each load node in the power system, a graph network structure and adjacency matrix are constructed. The load characteristic tensor and meteorological characteristic vector of the corresponding node are spliced ​​at each time slice to form a temporal characteristic representation of the graph network node. Based on the constructed graph network structure and adjacency matrix, a power load forecasting model based on spatiotemporal graph convolutional network is constructed, and power load forecasting is realized based on the constructed power load forecasting model.

[0009] Furthermore, the load nodes are each substation node, the historical load data are the active power values ​​of the power grid measured corresponding to each substation node within a preset sampling period, and the regional meteorological element data are multiple meteorological characteristics measured in the same sampling period in the area where the substation is located, and the meteorological characteristics include atmospheric temperature, dew point temperature, atmospheric pressure, relative humidity, wind speed, precipitation, and sunshine duration.

[0010] Furthermore, the characteristic contribution value of each meteorological feature is calculated based on the regional meteorological element data by combining model interpretation with regression analysis, specifically including: Based on the regional meteorological element data and the load data at the corresponding time, a multivariate linear regression model is constructed to obtain the regression coefficients of each meteorological feature. ; Based on the regional meteorological element data and the existing load forecasting model, the model interpretation algorithm is used to calculate the model interpretation contribution value of each meteorological feature ; According to the regression coefficient of each meteorological feature and the model explanation contribution value, the characteristic contribution value of each meteorological feature to the load data is calculated: in, For the i The characteristic contribution value of each meteorological feature, 、 Respectively i The regression coefficients of each meteorological characteristic and the explanatory contribution of the model; For all meteorological characteristics Sort the values ​​from largest to smallest, select the first k corresponding meteorological features, constructing meteorological feature vectors .

[0011] Furthermore, the multivariate linear regression model is constructed based on the regional meteorological element data and the load data at the corresponding time to obtain the regression coefficient of each meteorological feature. , specifically including: At each sampling moment t Load increment As the dependent variable, the meteorological characteristic vector at the same time As the independent variable, a multiple linear regression model is established: in, 、 are the moments in the historical load data t and time t-1 The load value, m is the total number of meteorological features, For the i The regression coefficient of a meteorological characteristic represents the linear influence of the meteorological characteristic on the load increment; Indicates the i Meteorological characteristics at time t Sample value, is the error term, is a constant term; The regression coefficients of each meteorological characteristic are obtained by the least squares method. .

[0012] Furthermore, based on the regional meteorological element data and the existing load forecasting model, the model interpretation algorithm is used to calculate the model interpretation contribution value of each meteorological feature. , specifically including: For the i Meteorological features, define the meteorological feature set as , for any feature that does not contain i Subset of , divide the input vector into subsets containing S The meteorological characteristic vector and contains subsets S and features to be measured i The meteorological characteristic vector , calculate the first i Model explanation contribution of each meteorological feature : in, Representation subset S The factorial of the number of meteorological features contained in , Indicates the remaining The factorial of the number of meteorological features after removing the feature to be measured, Indicates the total number of meteorological features m The factorial of 、 They represent the input meteorological feature vectors as and The output of the existing load forecasting model.

[0013] Furthermore, the method of performing multivariate empirical mode decomposition on historical load data of multiple load nodes in the power system, extracting multi-order intrinsic mode function components, and performing singular decomposition on each order modal component specifically includes: In the power system N The historical load series of the load nodes are decomposed by multivariate empirical mode decomposition to obtain K Intrinsic mode function sequence ,in, Represents a load node j At the moment t No. Strip modal components; For each modal component of each load node , with sliding window length L Construct the Hankel embedding matrix: in, Represents a load node j No. Strip modal components Constructed after time delay embedding matrix; Represents an operation that converts a one-dimensional time series into a two-dimensional matrix with a time structure, so that each anti-diagonal element in the matrix is ​​the same. Indicates the total number of sampling time steps in the historical load data, is the length of the sliding window; Hankel embedding matrix for each load node Perform singular value decomposition to obtain the singular value diagonal matrix : in, is the left singular vector matrix, is the singular value matrix, is the right singular vector matrix and T is its transpose.

[0014] Furthermore, the construction of the low-rank approximate embedding matrix to obtain the multi-scale smoothed reconstructed load feature tensor specifically includes: Preserve the singular value diagonal matrix after singular decomposition Beforer Largest singular value and the corresponding left and right singular vectors 、 , construct a low-rank approximate embedding matrix: in, Represents a load node j No. The low-rank approximate embedding matrix on the strip modal components, is a preset constant, Load nodes j No. The first modal component p singular values, left singular vectors, and right singular vectors; And through the inverse Hankel operation Reconstructed into load nodes j In the k Load reconstruction signal on the strip modal component: , obtain the load feature tensor after multi-scale smooth reconstruction: in, Represents a load node j At the moment t The load feature tensor on , Represents a load node j In the k The moment of the modal component t The load reconstruction signal.

[0015] Furthermore, the power load forecasting model based on spatiotemporal graph convolutional network includes: The spatial graph convolution module consists of multiple stacked graph convolution layers, each of which receives a time slice. t The node feature matrix and the normalized adjacency matrix are input, and the spatial features are output through graph convolution operation; The temporal convolution module includes multiple layers of causal one-dimensional convolution. Each layer of causal one-dimensional convolution performs convolution on the node features of the same load node and outputs time series features. The spatiotemporal fusion module includes a feature concatenation layer and a fusion mapping layer, which is used to fuse the spatial features output by the spatial graph convolution module with the temporal features output by the temporal convolution module and map them to the prediction dimension to generate spatiotemporal features; The output prediction submodule is used to map the fused spatiotemporal features of each load node to the load forecast value of the next period through one or more fully connected layers.

[0016] Furthermore, the construction of a graph network structure and an adjacency matrix based on the physical connection relationship of each load node in the power system specifically includes: Based on the meteorological characteristic vectors and load characteristic tensors of each load node and their physical connection relationships, a graph network structure at multiple historical moments is constructed, specifically including: All load nodes in the power system constitute a node set V , and form an edge set based on the physical connectivity relationship of the transmission and distribution lines between each load node E , where the node characteristics of the graph network structure corresponding to each moment are: in, is the node of the graph network structure corresponding to time t The eigenvector of Load node j At the moment t The load feature tensor on , Load node j At the moment t The meteorological characteristic vector on Represents vector concatenation operation; The edge weights of the graph network structure corresponding to each moment are determined by the physical properties of the transmission and distribution lines between the load nodes, and a normalized adjacency matrix with self-loops is constructed. : in, Representation node and The edge weights of is the identity matrix, N represents the number of load nodes, is the degree matrix, whose diagonal elements ; With node set V, normalized adjacency matrix And node characteristics , defining the graph network structure at each moment , obtain the graph network structure set .

[0017] Furthermore, based on the constructed graph network structure and adjacency matrix, a power load forecasting model based on a spatiotemporal graph convolutional network is constructed, specifically including: For every historical moment t , the graph network structure Normalized adjacency matrix in and node feature matrix Input the spatial graph convolution module of the power load forecasting model, perform graph convolution operations, aggregate neighbor features, and obtain the time t Spatial characteristics of each node: Collect from the moment The spatial features of each node to t are input into the temporal convolution module, which performs multi-layer causal one-dimensional convolution or dilated convolution to capture short-term mutations and long-term trends between nodes and output the temporal features of each node: The spatial features and temporal features are spliced ​​in the feature dimension and mapped to a unified dimension through the spatiotemporal fusion module to generate node-level spatiotemporal features; Input the spatiotemporal features into the output prediction submodule to generate the load forecast value of each node at the next moment; The actual load value of each node at the next moment is used as the label to calculate the mean square error loss, and the parameters of the power load forecasting model are iteratively updated using back propagation and optimizer based on the loss to achieve the training of the power load forecasting model.

[0018] Compared with the prior art, the present invention has the following advantages: (1) In the prior art, the screening of meteorological influencing factors relies solely on the Pearson correlation coefficient metric, that is, the influencing factors are selected by comparing the linear correlation between each meteorological variable and the power load. This method can only reflect the linear relationship between variables, and cannot characterize the nonlinear contribution of meteorological factors to load changes, nor can it identify the true causal drive. It is easy to mistakenly select factors that are accidentally correlated or affected by other potential variables as key features. In the feature screening stage, the present invention first constructs a multivariate linear regression model of regional meteorological elements and load increments, and obtains the regression coefficient of each meteorological factor by the least squares method. , used to quantify the linear causal strength of the factor on load changes; then the marginal contribution value of each meteorological factor is calculated based on the existing load forecasting model using the model interpretation algorithm , to fully capture its impact on the model output in a nonlinear environment; finally, the two are multiplied to obtain the comprehensive feature contribution , and sort by this indicator to select the most driving k In this way, the present invention not only eliminates those variables that have only accidental linear correlation with load changes but lack true causal effects, but also retains those meteorological factors that have significant marginal effects in complex nonlinear scenarios, thereby significantly improving the ability of the selected features to explain power load fluctuations under extreme weather conditions and enhancing the robustness and prediction accuracy of the load forecasting model.

[0019] (2) The existing technology usually takes the historical load series of each load node as the model input directly, ignoring the strong noise, high-frequency jitter and non-stationary characteristics of the load time series under extreme weather conditions. When there are sudden peaks or periodic and non-periodic components interwoven in the sequence, directly feeding it into the deep learning network will not only increase the difficulty of model fitting, but also easily lead to overfitting or insufficient response to severe fluctuations, thereby reducing the prediction accuracy and robustness. To this end, the present invention first applies multivariate empirical mode decomposition to the historical load data of multiple load nodes at the same time, decomposing the original signal into several intrinsic mode functions IMF and residual terms, thereby separating the trends, cycles and noise components of different frequency bands in the load series. Each IMF reflects the dominant fluctuation mode on its corresponding time scale, eliminating the modal aliasing problem between different modes. Subsequently, each IMF is time-delayed and embedded and the Hankel matrix is ​​constructed. The singular value decomposition is used to extract its predecessor. r The main singular values ​​and corresponding singular vectors are used to complete the low-rank approximate reconstruction. This step not only removes high-frequency noise, but also retains the most representative dynamic features of each mode, making the reconstructed signal smooth and rich in multi-scale information. Finally, the reconstructed values ​​of each node on each mode are calculated by time. t The data is spliced ​​into a load feature tensor, which reflects the load variation patterns of nodes at different time scales. Through the above processing, the present invention converts the non-stationary and heavily noise-contaminated original load sequence into a smooth feature tensor containing multi-scale physically meaningful modal components, significantly reducing the non-stationarity and noise interference of the data. It provides more stable and discriminative input for the spatial and temporal convolution modules of the spatiotemporal graph convolutional network, significantly enhancing the model's ability to capture sudden peaks and multi-frequency fluctuations in extreme weather, thereby improving the accuracy and robustness of short-term power load forecasting.

[0020] (3) In the existing technology, spatial convolution is simply connected in series with time series models (such as LSTM or GRU), resulting in the independence of spatial features and temporal features, insufficient interaction, and difficulty in synchronously capturing the instantaneous spatial diffusion effect and historical dynamic evolution trend of power load under extreme weather conditions. To address this defect, the present invention designs a multi-layer stacked graph convolution layer under the same framework to deeply integrate the node features of each time slice with the network topology, truly reflecting the electrical coupling between different substations through transmission and distribution lines, and automatically aggregating neighborhood information to obtain a more accurate and generalizable spatial representation. Subsequently, the multi-layer causal one-dimensional convolution or dilated convolution module processes these spatial representations in parallel along the time dimension, which can not only capture short-term sudden load peaks, but also perceive long-term fluctuation trends at multiple scales, making up for the limitations of RNN-type algorithms in parallel computing and long-range dependencies. Subsequently, the spatiotemporal fusion module splices the above two outputs in the feature dimension and maps them to a unified dimension, enabling the model to simultaneously utilize the information of spatial topological association and temporal dynamic evolution to generate true node-level spatiotemporal features. Ultimately, these features rich in spatiotemporal interaction information are mapped to the load forecast for the next moment through a fully connected submodule. Backpropagation is then used to continuously adjust the parameters of the graph convolution kernel, temporal convolution kernel, and fusion mapping layer during end-to-end training. This structural design effectively addresses the problem of insufficient spatial and temporal coupling in existing technologies. It can more sensitively capture cross-node load contagion effects and mutation evolution patterns under extreme weather conditions, significantly improving the accuracy and robustness of short-term power load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a method for predicting power load according to an embodiment of the present invention; Figure 2 Schematic diagram of a power load forecasting model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0023] Example 1: This embodiment provides a method for predicting power load under extreme weather conditions based on spatiotemporal graph convolutional networks. Figure 1 As shown, the following features are included: Step S1: Collect historical load data and regional meteorological element data of multiple load nodes in the power system; wherein, the load nodes are each substation node, the historical load data are the active power values ​​of the power grid measured at each substation node within a preset sampling period, and the regional meteorological element data are multiple meteorological characteristics measured at the area where the substation is located within the same sampling period, and the meteorological characteristics include atmospheric temperature, dew point temperature, atmospheric pressure, relative humidity, wind speed, precipitation, and sunshine duration.

[0024] The collected historical load data and regional meteorological element data are preprocessed, including: using linear interpolation or sliding average methods to fill in data with missing values, and using box plot method or Z-score method to detect and eliminate data with outliers; normalizing the features of each dimension according to maximum and minimum value normalization or Z-score standardization to eliminate dimensional effects and improve model training stability; at the same time, aligning the load data and meteorological data according to a unified sampling period and timestamp to ensure the consistency of input features on the time axis.

[0025] Step S2: Based on the regional meteorological element data, the characteristic contribution value of each meteorological feature is calculated by combining model interpretation with regression analysis, and several key meteorological features are selected to construct a meteorological characteristic vector, specifically including: Based on the regional meteorological element data and the load data at the corresponding time, a multivariate linear regression model is constructed to obtain the regression coefficients of each meteorological characteristic. , specifically including: At each sampling moment t Load increment As the dependent variable, the meteorological characteristic vector at the same time As the independent variable, a multiple linear regression model is established: in, 、 are the moments in the historical load data t and time t-1 The load value, m is the total number of meteorological features, For the i The regression coefficient of a meteorological characteristic represents the linear influence of the meteorological characteristic on the load increment; Indicates the i Meteorological characteristics at time t Sample value, is the error term, is a constant term; The regression coefficients of each meteorological characteristic are obtained by solving the least squares method .

[0026] Based on regional meteorological element data and existing load forecasting models, the model interpretation algorithm is used to calculate the model interpretation contribution value of each meteorological feature , specifically including: For the i Meteorological features, define the meteorological feature set as , for any feature that does not contain i Subset of , divide the input vector into subsets containing S The meteorological characteristic vector and contains subsets S and features to be measured i The meteorological characteristic vector , calculate the first i Model explanation contribution of each meteorological feature : in, Representation subset S The factorial of the number of meteorological features contained in , Indicates the remaining The factorial of the number of meteorological features after removing the feature to be measured, Indicates the total number of meteorological features m The factorial of 、 They represent the input meteorological feature vectors as and The output of the existing load forecasting model.

[0027] According to the regression coefficient of each meteorological feature and the model explanation contribution value, the characteristic contribution value of each meteorological feature to the load data is calculated: in, For the i The characteristic contribution value of each meteorological feature, 、 Respectively i The regression coefficients of each meteorological characteristic and the model explanation contribution; For all meteorological characteristics Sort the values ​​from largest to smallest, select the first k corresponding meteorological features, constructing meteorological feature vectors .

[0028] Through the above step S2, the present invention introduces an organic combination of linear regression analysis and model interpretation method in the process of meteorological feature screening, comprehensively considering the linear causal effect of each meteorological factor on load change and the marginal contribution in the nonlinear model, thereby constructing a more representative feature contribution evaluation index. , and based on this, the key meteorological features with the greatest driving effect are screened out. This technical feature effectively compensates for the nonlinear feature effects and false correlation problems that are ignored by the traditional reliance on correlation measurement, eliminates noise variables and redundant features, and improves the scientific nature and interpretability of feature selection. In terms of technical effects, this method significantly enhances the model's perception and response capabilities to meteorological driving factors in the face of extreme weather events, improves the stability, robustness and overall prediction accuracy of power load forecasting, and provides a reliable data foundation for the efficient training and accurate reasoning of subsequent models.

[0029] Step S3: Perform multivariate empirical mode decomposition on the historical load data of multiple load nodes in the power system, extract multi-order intrinsic mode function components, perform singular decomposition on each order modal component, construct a low-rank approximate embedding matrix, and obtain the load feature tensor after multi-scale smooth reconstruction, specifically including: In the power system N The historical load series of the load nodes are decomposed by multivariate empirical mode decomposition to obtain K Intrinsic mode function sequence ,in, Represents a load node j At the moment t No. Strip modal components; For each modal component of each load node , with sliding window length L Construct the Hankel embedding matrix: in, Represents a load node j No. Strip modal components Constructed after time delay embedding matrix; Represents an operation that converts a one-dimensional time series into a two-dimensional matrix with a time structure, so that each anti-diagonal element in the matrix is ​​the same. Indicates the total number of sampling time steps in the historical load data, is the length of the sliding window; Hankel embedding matrix for each load node Perform singular value decomposition to obtain the singular value diagonal matrix : in, is the left singular vector matrix, is the singular value matrix, is the right singular vector matrix and T is its transpose.

[0030] Preserve the singular value diagonal matrix after singular decomposition Before r Largest singular value and the corresponding left and right singular vectors 、 , construct a low-rank approximate embedding matrix: in, Represents a load node j No. The low-rank approximate embedding matrix on the strip modal components, is a preset constant, Load nodes j No. The first modal component p singular values, left singular vectors, and right singular vectors; And through the inverse Hankel operation Reconstructed into load nodes j In the k Load reconstruction signal on the strip modal component: , obtain the load feature tensor after multi-scale smooth reconstruction: in, Represents a load node j At the moment t The load feature tensor on , Represents a load node j In the k The moment of the modal component t The load reconstruction signal.

[0031] In this embodiment, step S3 combines multivariate empirical mode decomposition (MEMD) with singular value decomposition (SVD) to perform multiscale time series structural modeling and noise reduction on historical load data from multiple load nodes in the power system. This approach aims to extract smoother and more physically meaningful load characteristic tensors, providing higher-quality input data for subsequent prediction models. Multivariate empirical mode decomposition (MEMD) is introduced to decompose the original load time series into several intrinsic mode function (IMF) components with different time-scale characteristics. This process helps capture the varying characteristics of power loads at different frequency levels (such as high-frequency disturbances and low-frequency trends), enabling multiscale modeling of load signals and addressing the difficulty in effectively extracting multi-time-scale information from the original load data. For each modal component, the dominant low-rank structural features are obtained through Hankel matrix construction and singular value decomposition. The Hankel embedding operation expands the one-dimensional time series into a two-dimensional matrix with a time-delay structure, helping to explicitly preserve its temporal evolution. SVD, on the other hand, extracts the primary direction of variation, removes noise and local fluctuations, and achieves smooth reconstruction of the modal components. Through low-rank approximation, only the most representative dynamic mode in each mode is retained, which improves the robustness and predictability of the data. The smooth reconstruction results of all modal components are spliced ​​in the time dimension to form a complete multi-scale load feature tensor. This tensor not only retains the main trends and fluctuation patterns of the original load signal, but also effectively filters out noise and abnormal disturbances, and has higher learnability and stronger characterization capabilities. This step realizes the denoising, dimensionality reduction and multi-scale enhanced expression of historical load data, and solves the problems of load data in existing technologies such as violent fluctuations, difficult to handle abnormal disturbances, and aliasing of changes at different time scales. It provides a solid data foundation for the model to accurately characterize the load evolution law and improve the prediction performance under extreme weather conditions, significantly improving the accuracy and stability of the overall prediction.

[0032] Step S4: Based on the physical connection relationship of each load node in the power system, a graph network structure and an adjacency matrix are constructed. The load characteristic tensor and meteorological characteristic vector of the corresponding node are spliced ​​at each time slice to form a temporal feature representation of the graph network node, which specifically includes: Based on the meteorological characteristic vectors and load characteristic tensors of each load node and their physical connection relationships, a graph network structure at multiple historical moments is constructed, specifically including: All load nodes in the power system constitute a node set V , and form an edge set based on the physical connectivity relationship of the transmission and distribution lines between each load node E , where the node characteristics of the graph network structure corresponding to each moment are: in, is the node of the graph network structure corresponding to time t The eigenvector of Load node j At the moment t The load feature tensor on , Load node j At the moment t The meteorological characteristic vector on Represents vector concatenation operation; The edge weights of the graph network structure corresponding to each moment are determined by the physical properties of the transmission and distribution lines between the load nodes, and a normalized adjacency matrix with self-loops is constructed. : in, Representation node and The edge weights of is the identity matrix, N represents the number of load nodes, is the degree matrix, whose diagonal elements ; With node set V, normalized adjacency matrix And node characteristics , define the graph network structure at each moment , obtain the graph network structure set .

[0033] Step S5: Based on the constructed graph network structure and adjacency matrix, a power load forecasting model based on the spatiotemporal graph convolutional network is constructed, and power load forecasting is implemented based on the constructed power load forecasting model; Among them, the power load forecasting model based on spatiotemporal graph convolutional network is as follows: Figure 2 As shown, including: The spatial graph convolution module consists of multiple stacked graph convolution layers (Graph Conv), each of which receives a time slice. t The node feature matrix and the normalized adjacency matrix are input, and the spatial features are output through graph convolution operation; The temporal convolution module includes multiple layers of causal one-dimensional convolution (Causal Conv). Each layer of causal one-dimensional convolution performs convolution on the node features of the same load node and outputs time series features. The spatiotemporal fusion module (Spatiotemporal Fusion), including the feature splicing layer and the fusion mapping layer, is used to fuse the spatial features output by the spatial graph convolution module with the temporal features output by the temporal convolution module and map them to the prediction dimension to generate spatiotemporal features; The output prediction submodule is used to map the fused spatiotemporal features of each load node to the load forecast value of the next period through one or more fully connected layers.

[0034] Based on the constructed graph network structure and adjacency matrix, a power load forecasting model based on spatiotemporal graph convolutional network is constructed, which specifically includes: For every historical moment t , the graph network structure Normalized adjacency matrix in and node feature matrix Input the spatial graph convolution module of the power load forecasting model, perform graph convolution operations, aggregate neighbor features, and obtain the time t Spatial characteristics of each node: Collect from the moment The spatial features of each node to t are input into the temporal convolution module, which performs multi-layer causal one-dimensional convolution or dilated convolution to capture short-term mutations and long-term trends between nodes and output the temporal features of each node: The spatial features and temporal features are spliced ​​in the feature dimension and mapped to a unified dimension through the spatiotemporal fusion module to generate node-level spatiotemporal features; Input the spatiotemporal features into the output prediction submodule to generate the load forecast value of each node at the next moment; The actual load value of each node at the next moment is used as the label to calculate the mean square error loss, and the parameters of the power load forecasting model are iteratively updated using back propagation and optimizer based on the loss to achieve the training of the power load forecasting model.

[0035] Example 2: This embodiment provides a power load forecasting system based on a spatiotemporal graph convolutional network under extreme weather conditions, including: A data acquisition module is used to collect historical load data of multiple load nodes (i.e., substation nodes) in the power system and multi-dimensional meteorological element data of the corresponding area. The meteorological elements include but are not limited to temperature, humidity, air pressure, wind speed, precipitation, etc.; The data preprocessing module is used to clean, align, and normalize the collected data. It also selects the meteorological features most closely related to power load through a combination of correlation analysis and model sensitivity analysis to improve the effectiveness and generalization ability of the model input. The feature extraction module is used to perform multi-scale decomposition of the load time series based on multivariate empirical mode decomposition (MEMD), extract multi-order modal components, and perform singular value decomposition (SVD) and low-rank approximation reconstruction on each order mode to obtain a smoothed and denoised load feature tensor; A graph construction module is used to construct a graph-structured adjacency matrix between grid nodes based on geographic distance and load correlation, and obtain a graph representation that can be used in a graph convolutional network through normalization processing; The load forecasting model module includes multiple parallel graph convolution (Graph Conv) and causal convolution (CausalConv) units, which are used to extract spatial correlation features and temporal dependency features between nodes. The spatiotemporal features of each branch are integrated through a fusion mechanism (Spatiotemporal Fusion). The fused features are input into a fully connected network (Fully Connected), which outputs the predicted load value. The result output module is used to output the load forecast value of each node at the future moment, and compare it with the actual load value, and further calculate the error index for model evaluation and optimization.

[0036] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0037] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for power load forecasting under extreme weather conditions based on spatiotemporal graph convolutional networks, characterized in that: Features include: Collect historical load data and regional meteorological data of multiple load nodes in the power system; Based on regional meteorological element data, the characteristic contribution value of each meteorological feature is calculated by combining model interpretation with regression analysis, and several key meteorological features are selected to construct a meteorological characteristic vector; Perform multivariate empirical mode decomposition on the historical load data of multiple load nodes in the power system, extract multi-order intrinsic mode function components, perform singular decomposition on each order modal component, construct a low-rank approximate embedding matrix, and obtain the load characteristic tensor after multi-scale smooth reconstruction; Based on the physical connection relationship of each load node in the power system, a graph network structure and adjacency matrix are constructed. The load characteristic tensor and meteorological characteristic vector of the corresponding node are spliced ​​at each time slice to form a temporal characteristic representation of the graph network node. Based on the constructed graph network structure and adjacency matrix, a power load forecasting model based on spatiotemporal graph convolutional network is constructed, and power load forecasting is realized based on the constructed power load forecasting model.

2. The method for power load forecasting based on spatiotemporal graph convolutional network under extreme weather conditions according to claim 1 is characterized in that: The load nodes are each substation node, the historical load data are the active power values ​​of the power grid measured corresponding to each substation node within a preset sampling period, and the regional meteorological element data are multiple meteorological characteristics measured in the same sampling period in the area where the substation is located. The meteorological characteristics include atmospheric temperature, dew point temperature, atmospheric pressure, relative humidity, wind speed, precipitation, and sunshine duration.

3. The method for power load forecasting based on spatiotemporal graph convolutional network under extreme weather conditions according to claim 1 is characterized in that: The characteristic contribution value of each meteorological feature is calculated based on regional meteorological element data by combining model interpretation with regression analysis, specifically including: Based on the regional meteorological element data and the load data at the corresponding time, a multivariate linear regression model is constructed to obtain the regression coefficients of each meteorological feature. ; Based on the regional meteorological element data and the existing load forecasting model, the model interpretation algorithm is used to calculate the model interpretation contribution value of each meteorological feature ; According to the regression coefficient of each meteorological feature and the model explanation contribution value, the characteristic contribution value of each meteorological feature to the load data is calculated: in, For the i The characteristic contribution value of each meteorological feature, 、 Respectively i The regression coefficients of each meteorological characteristic and the model explanation contribution; For all meteorological characteristics Sort the values ​​from largest to smallest, select the first k corresponding meteorological features, constructing meteorological feature vectors .

4. The method for power load forecasting based on spatiotemporal graph convolutional network under extreme weather conditions according to claim 3 is characterized in that: The multivariate linear regression model is constructed based on the regional meteorological element data and the load data at the corresponding time to obtain the regression coefficient of each meteorological feature. , specifically including: At each sampling moment t Load increment As the dependent variable, the meteorological characteristic vector at the same time As the independent variable, a multiple linear regression model is established: in, 、 are the moments in the historical load data t and time t-1 The load value, m is the total number of meteorological features, For the i The regression coefficient of a meteorological characteristic represents the linear influence of the meteorological characteristic on the load increment; Indicates the i Meteorological characteristics at time t Sample value, is the error term, is a constant term; The regression coefficients of each meteorological characteristic are obtained by the least squares method. .

5. The method for power load forecasting based on spatiotemporal graph convolutional network under extreme weather conditions according to claim 3 is characterized in that: Based on the regional meteorological element data and the existing load forecasting model, the model interpretation algorithm is used to calculate the model interpretation contribution value of each meteorological feature , specifically including: For the i Meteorological features, define the meteorological feature set as , for any feature that does not contain i Subset of , divide the input vector into subsets containing S The meteorological characteristic vector and contains subsets S and features to be measured i The meteorological characteristic vector , calculate the first i Model explanation contribution of each meteorological feature : in, Representation subset S The factorial of the number of meteorological features contained in , Indicates that from the remaining The factorial of the number of meteorological features after removing the feature to be measured, Indicates the total number of meteorological features m The factorial of 、 Respectively represent the input meteorological feature vectors as and The output of the existing load forecasting model.

6. The method for power load forecasting based on spatiotemporal graph convolutional network under extreme weather conditions according to claim 1 is characterized in that: The method of performing multivariate empirical mode decomposition on historical load data of multiple load nodes in the power system, extracting multi-order intrinsic mode function components, and performing singular decomposition on each order mode component specifically includes: In the power system N The historical load series of the load nodes are decomposed by multivariate empirical mode decomposition to obtain K Intrinsic mode function sequence ,in, Represents a load node j At the moment t No. Strip modal components; For each modal component of each load node , with sliding window length L Construct the Hankel embedding matrix: in, Represents a load node j No. Strip modal components Constructed after time delay embedding matrix; Represents an operation that converts a one-dimensional time series into a two-dimensional matrix with a time structure, so that each anti-diagonal element in the matrix is ​​the same. Indicates the total number of sampling time steps in the historical load data, is the length of the sliding window; Hankel embedding matrix for each load node Perform singular value decomposition to obtain the singular value diagonal matrix : in, is the left singular vector matrix, is the singular value matrix, is the right singular vector matrix and T is its transpose.

7. The method for power load forecasting under extreme weather conditions based on spatiotemporal graph convolutional networks according to claim 1 or 6, characterized in that: The low-rank approximate embedding matrix is ​​constructed to obtain the load feature tensor after multi-scale smooth reconstruction, specifically including: Preserve the singular value diagonal matrix after singular decomposition Before r Largest singular value and the corresponding left and right singular vectors 、 , construct a low-rank approximate embedding matrix: in, Represents a load node j No. The low-rank approximate embedding matrix on the strip modal components, is a preset constant, Load nodes j No. The first modal component p singular values, left singular vectors, and right singular vectors; And through the inverse Hankel operation Reconstructed into load nodes j In the k Load reconstruction signal on the strip modal component: , obtain the load feature tensor after multi-scale smooth reconstruction: in, Represents a load node j At the moment t The load feature tensor on , Represents a load node j In the k The moment of the modal component t The load reconstruction signal.

8. The method for power load forecasting under extreme weather conditions based on spatiotemporal graph convolutional networks according to claim 1 is characterized in that: The power load forecasting model based on spatiotemporal graph convolutional network includes: The spatial graph convolution module consists of multiple stacked graph convolution layers, each of which receives a time slice. t The node feature matrix and the normalized adjacency matrix are input, and the spatial features are output through graph convolution operation; The temporal convolution module includes multiple layers of causal one-dimensional convolution. Each layer of causal one-dimensional convolution performs convolution on the node features of the same load node and outputs time series features. The spatiotemporal fusion module includes a feature concatenation layer and a fusion mapping layer, which is used to fuse the spatial features output by the spatial graph convolution module with the temporal features output by the temporal convolution module and map them to the prediction dimension to generate spatiotemporal features; The output prediction submodule is used to map the fused spatiotemporal features of each load node to the load forecast value of the next period through one or more fully connected layers.

9. The method for power load forecasting under extreme weather conditions based on spatiotemporal graph convolutional networks according to claim 1 is characterized in that: The construction of a graph network structure and an adjacency matrix based on the physical connection relationship of each load node in the power system specifically includes: Based on the meteorological characteristic vectors and load characteristic tensors of each load node and their physical connection relationships, a graph network structure at multiple historical moments is constructed, specifically including: All load nodes in the power system constitute a node set V , and form an edge set based on the physical connectivity relationship of the transmission and distribution lines between each load node E , where the node characteristics of the graph network structure corresponding to each moment are: in, is the node of the graph network structure corresponding to time t The eigenvector of Load node j At the moment t The load feature tensor on , Load node j At the moment t The meteorological characteristic vector on Represents vector concatenation operation; The edge weights of the graph network structure corresponding to each moment are determined by the physical properties of the transmission and distribution lines between the load nodes, and a normalized adjacency matrix with self-loops is constructed. : in, Representation node and The edge weights of is the identity matrix, N represents the number of load nodes, is the degree matrix, whose diagonal elements ; With node set V, normalized adjacency matrix And node characteristics , define the graph network structure at each moment , obtain the graph network structure set .

10. The method for power load forecasting under extreme weather conditions based on spatiotemporal graph convolutional networks according to claim 1, 8 or 9, characterized in that: The power load forecasting model based on the spatiotemporal graph convolutional network is constructed based on the constructed graph network structure and adjacency matrix, specifically including: For every historical moment t , the graph network structure Normalized adjacency matrix in and node feature matrix Input the spatial graph convolution module of the power load forecasting model, perform graph convolution operations, aggregate neighbor features, and obtain the time t Spatial characteristics of each node: Collect from the moment The spatial features of each node to t are input into the temporal convolution module, which performs multi-layer causal one-dimensional convolution or dilated convolution to capture short-term mutations and long-term trends between nodes and output the temporal features of each node: The spatial features and temporal features are spliced ​​in the feature dimension and mapped to a unified dimension through the spatiotemporal fusion module to generate node-level spatiotemporal features; Input the spatiotemporal features into the output prediction submodule to generate the load forecast value of each node at the next moment; The actual load value of each node at the next moment is used as the label to calculate the mean square error loss, and the parameters of the power load forecasting model are iteratively updated using back propagation and optimizer based on the loss to achieve the training of the power load forecasting model.

Images