A regional short-term load forecasting method for tracing load flow causality

Abstract

A regional short-term load forecasting method for tracking the causality of load flow, comprising the following steps: step 1: normalizing and seasonally decomposing historical load data to extract high-frequency residual components after removing the interference of trend items and seasonal items; step 2: applying a sliding window mechanism to the residual components, applying a fast causal inference algorithm in each time window, performing conditional independence test on the residual sequence between regions, identifying causal dependence edges, and constructing a dynamic causal adjacency matrix that evolves over time; step 3: inputting the historical load data and the dynamic causal adjacency matrix into the spatial feature extraction module and the temporal feature extraction module of the space-time graph neural network framework, respectively; step 4: fusing the spatial features and the temporal features; step 5: outputting multi-step load prediction values through a fully connected layer and performing denormalization. The present application describes the dynamic conduction relationship between regional loads by constructing a causal structure that changes over time, which can more accurately reflect the load flow mechanism compared to traditional static modeling methods based on correlation, thereby improving the prediction accuracy and model interpretability.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a short-term load forecasting method for analyzing the spatiotemporal characteristics of power load in multiple regions. Specifically, it is a regional short-term load forecasting technology that relies solely on historical load data, integrates seasonal trend decomposition, sliding window causal inference, and a spatiotemporal graph neural network framework. This invention is applicable to power grid dispatch optimization, demand response management, renewable energy consumption, and improvement of smart grid operation stability. Background Technology

[0002] As the penetration rate of renewable energy continues to increase, the operation of the power grid faces the challenge of balancing supply and demand due to the superposition of intermittent power sources and uncertainties on the demand side. Regional short-term load forecasting, as a core foundational technology for power dispatch, economic dispatch, and renewable energy consumption, directly impacts power generation planning, peak shaving and frequency regulation resource optimization, and power market trading decisions with its forecasting accuracy.

[0003] Traditional short-term load forecasting methods often assume that regional loads are independent or that inter-regional correlations are fixed, neglecting the causal transmission mechanism of load flows across regions. This leads to the loss of spatial structure information and reduced forecast accuracy. Mainstream spatiotemporal forecasting research often uses geographical location, electrical distance, or load data correlation to construct static adjacency matrices. Some studies introduce graph attention networks or learnable adjacency matrices to dynamically adjust spatial weights. However, these methods are essentially still based on statistical correlation modeling and cannot accurately characterize the causal mechanisms and directionality of load flows between regions. Correlation analysis can confuse spurious correlations with real causal effects. When the grid topology or operating conditions change, the spatial structure constructed based on correlations will suffer severe mismatch.

[0004] While existing research recognizes the existence of load transmission effects between regions, few studies have removed the inherent trend and periodic components in load sequences before conducting causal analysis. Since actual load sequences simultaneously contain multiple superimposed components such as intraday cycles, intraweekly seasonality, and long-term trends, directly performing causal tests on the original load sequences can lead to spurious common trends, severely interfering with causal inferences and failing to reflect the true dynamic causal relationships between regions. Furthermore, the causal relationships of load flows between regions are not static but dynamically change with operating conditions, seasonal variations, and electricity consumption behavior. Most existing methods employ static adjacency matrices calculated based on full historical data, failing to capture this time-varying characteristic. Even those few studies that introduce dynamic graph learning rely on gradient optimization of model parameters rather than direct measurement of physical causal relationships, resulting in insufficient interpretability. In addition, existing spatiotemporal load forecasting methods are insufficient in deep interactive modeling of temporal and spatial features. The coupling information between these two types of features is easily weakened during simple splicing or weighted summation, limiting the model's ability to model complex load evolution patterns.

[0005] In summary, there is an urgent need for a load forecasting method that can dynamically identify the causal relationships of load flows between regions from historical load data, based on the elimination of trend and seasonal interference, and explicitly integrate the causal structure into spatiotemporal modeling. Summary of the Invention

[0006] The purpose of this invention is to solve the following technical problems: Existing regional short-term load forecasting methods rely on static adjacency matrices or statistical correlation models to model regional associations, which cannot accurately capture the dynamic causal transmission paths of load flow between regions; when performing causal inference, trend and seasonal terms in the load sequence are not removed, leading to false common trends interfering with causal test conclusions; the spatiotemporal feature interaction and fusion mechanism is weak, and the coupling information of temporal evolution and spatial propagation is easily weakened. These defects result in low forecasting accuracy and poor robustness of regional power grids in complex power grid environments with high proportions of renewable energy integration and increased load fluctuations, making it difficult to meet the actual needs of precise power grid dispatch and stable operation.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for regional short-term load forecasting that tracks the causality of load flows includes the following steps: Step 1: Obtain the historical load data sequence of the target area, perform normalization processing, decompose the load sequence of each area into trend term, seasonal term and residual term, and extract the residual term as high-frequency residual component; Step 2: Apply a sliding window mechanism to the high-frequency residual components, define the causal analysis window and step size, and divide the residual sequence into a series of time window subsets; within each time window, perform conditional independence tests on the residual sequences between regions to identify causal dependency edges between regions; construct the adjacency matrix corresponding to the window based on the identified causal edges, thereby generating a dynamic causal adjacency matrix sequence that evolves over time. Step 3: Input the historical load data and the dynamic causal adjacency matrix corresponding to the current moment generated in Step 2 into the classic spatiotemporal graph neural network framework to extract the spatial causal features between regions and the dynamic evolution features in the time dimension. Step 4: Fuse spatial and temporal features to obtain spatiotemporal fusion features; Step 5: Input the spatiotemporal fusion features into the fully connected layer, output the regional load prediction values ​​for multiple future time steps, and perform inverse normalization to obtain the final prediction result.

[0008] In step 1, the input data is processed. Min-Max normalization is performed, where T is the time step and N is the number of regions; the load sequence for each region i after normalization is then processed. ,by For periodicity, the STL algorithm is applied for decomposition, separating the original sequence into trend terms. Seasonal items and residuals The decomposition relationship is shown in equation (1): (1); Take the residual terms of each region Forming the residual matrix This serves as input for subsequent causal analysis.

[0009] In step 2, the causal analysis window size is defined. With window size M, From the prediction window The value is determined by the window multiplier M, and the calculation method is shown in equation (2): (2); by For step size, from Start sliding to Generate time window sequence ,in , , For each time window, extract the corresponding subset from the residual matrix R. ,right The FCI algorithm is applied to N region sequences to generate the dynamic causal adjacency matrix of the window. And store the matrix of all windows in a window range Dictionary with keys In the middle; when At that time, a static adjacency matrix generated in the same way using the complete residual sequence is adopted. ;when At that time, based on the window to which the current time step belongs, from Select the corresponding matrix from the list. If the corresponding window matrix has not yet been generated, select the window matrix that is closest in time.

[0010] In step 2, the classic Fast Causal Inference (FCI) algorithm is applied to perform conditional independence tests on the residual sequences between regions. Specifically, the FCI algorithm is used to test the conditional independence of the residual sequences within the window. and The independence of the execution conditions is tested using the partial correlation coefficient under the condition set Z. Based on this, the test statistic P-value is calculated as shown in equation (3): (3); Where n is the number of samples within the window. The number of elements in the condition set Z; at the significance level Below, by comparing the Z-value with the critical value of the standard normal distribution, the directed causal dependencies between regions are determined, and the following types of causal edges are output: express right One-way causal influence; Indicates a two-way causal relationship; This represents a causal relationship with uncertain direction. Based on the edge information above, an adjacency matrix is ​​constructed. Initially a zero-based matrix, after filling according to edge type, the matrix is ​​symmetricized and self-joined, ultimately generating a 0-1 adjacency matrix. This indicates that there is a causal coupling between regions i and j. This indicates that there is no direct causal relationship.

[0011] In step 3, the spatial feature extraction module of the spatiotemporal graph neural network framework takes historical load data and a dynamic causal adjacency matrix as input, and aggregates the features of neighboring nodes to extract spatial features containing causal topological information between regions. The temporal feature extraction module takes historical load sequences as input and extracts short-term and long-term dynamic evolution features in the time dimension. The specific implementation of the spatial feature extraction module and the temporal feature extraction module is not limited, and appropriate spatial and temporal modeling methods can be selected according to the actual application scenario.

[0012] In step 4, the time series features output by the time feature extraction module are used. As the query vector Q, the spatial features output by the spatial feature extraction module As the key vector K and value vector V, a trainable weight matrix is ​​used. , , After linear projection, the cross-attention fusion features are calculated according to equation (4). : (4); And introduce residual connections, Compared with the original time series characteristics Adding them together yields the final spatiotemporal fusion feature. As shown in equation (5): (5); In step 5, the features of the last time step of the spatiotemporal fusion feature sequence are extracted. After linear projection through a fully connected layer, the predicted values ​​for the next P time steps are output. As shown in equation (6): (6); The predicted value Y is inversely normalized to obtain the final load forecast results for each region.

[0013] Compared with the prior art, the present invention has the following technical effects: (1) Before performing inter-regional causal inference, this invention first uses STL decomposition to remove the trend and seasonal terms from the load series, and only performs FCI causality test on the residual terms after removing the above systematic components. This effectively avoids the false causal effect caused by common trends, and makes the identified inter-regional causal relationship more realistically reflect the dynamic transmission mechanism of load flow, with clear physical interpretability.

[0014] (2) The residual sequence is divided into time-series window subsets by a sliding window mechanism, and a causal adjacency matrix is ​​independently constructed in each window to form a dynamic causal adjacency matrix sequence that evolves over time. This dynamic causal modeling method can accurately capture the time-varying characteristics of the causal relationship of load flow between regions with changes in operating conditions and seasons, and overcomes the inherent defect that static adjacency matrices cannot respond to topological evolution.

[0015] (3) This invention uses the dynamic causal adjacency matrix as the spatial structure input of the spatiotemporal graph neural network, so that the spatial feature extraction process is based on an interpretable causal graph structure, rather than statistical correlation or learnable implicit weights, which helps to improve the generalization robustness of the model when the operating conditions change.

[0016] (4) The specific implementation of the spatial feature extraction module and the temporal feature extraction module of the spatiotemporal graph neural network framework in this invention is not limited. Different graph neural network methods and temporal modeling methods can be selected according to computing power resources and accuracy requirements. The scheme has good scalability and engineering adaptability. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of a method for regional short-term load forecasting that tracks the causality of load flow according to the present invention; Figure 2 This is a flowchart illustrating the logic of generating and selecting the dynamic causal adjacency matrix in this invention. Figure 3 This is a schematic diagram of the time-series evolution of the dynamic causal adjacency matrix in the experimental system according to the present invention; Figure 4 and Figure 5 This is a heatmap of Granger causal correlation between the original and residual sequences in the loading region. Detailed Implementation

[0018] like Figure 1 As shown, the overall process of a regional short-term load forecasting method for tracking the causality of load flows includes the following steps: Step 1: Perform data preprocessing and seasonal trend decomposition; Step 2: Construct a dynamic causal adjacency matrix based on sliding window FCI causal inference; Step 3: Extract spatial causal features and temporal dynamic features respectively using a spatiotemporal graph neural network framework; Step 4: Fusion of spatiotemporal features; Step 5: Output load forecast.

[0019] In step 1, multi-regional load time series are obtained from the historical load database of the target area to form an input data matrix. Where T is the historical time step and N is the number of regions. First, Min-Max normalization is applied to X, mapping the load values ​​of each region to... Intervals are used to eliminate the impact of dimensional differences on subsequent modeling.

[0020] Subsequently, the time series of each region i in the normalized data was analyzed. ,by To determine the periodicity, the STL algorithm is applied for decomposition, representing the original sequence as a trend term. Seasonal items With residuals The superposition is shown in equation (1): (1); STL decomposition, through multiple iterations of Locally Weighted Regression (LOESS), separates the seasonal cyclical components and long-term electricity consumption trend components formed by intraday electricity consumption patterns from the original sequence, resulting in residual terms. This primarily preserves the short-term random fluctuations of regional loads, eliminating the systematic interference from intraday cycles and trends. The residual terms for each region are combined column-wise to obtain the residual matrix. This serves as the input for subsequent causal inference. If causal tests are directly performed on the original load series containing trend and seasonal components, the intraday electricity consumption patterns followed by loads in multiple regions will cause widespread spurious correlations, leading to serious distortion of causal inference conclusions. The STL decomposition pre-step fundamentally solves this problem.

[0021] In step 2, considering the prediction window To address the problem of insufficient data samples to support the FCI algorithm, the window factor M is first defined, and then the prediction window is determined accordingly. Calculate the causal analysis window size As shown in equation (2): (2); Set a larger window size M (e.g., ... or This ensures that each time window contains a sufficient sample size to meet the statistical reliability requirements of the FCI algorithm for conditional independence testing. For step size, from Start sliding to Time windows are generated sequentially. Extract a subset from R for each window. .

[0022] For each The FCI algorithm is applied to perform pairwise conditional independence tests on the residual sequences of N regions. The advantage of the FCI algorithm lies in its ability to correctly handle situations with potential confounding variables by calculating partial correlation coefficients. The test statistic is obtained by Fischer transformation, as shown in equation (3): (3); Where n is the number of samples within the window, Let Z be the size of the condition set. At the significance level... Below, by comparing the Z-value with the critical value of the standard normal distribution... ,judge and After excluding the influence of condition set Z, determine whether a significant dependency still exists, identify the causal direction, and output three types of directed causal edges: (One-way causality) (Two-way causality) and (A causal relationship with an uncertain direction).

[0023] (4); Based on the above edge information, a 0-1 adjacency matrix is ​​constructed as shown in equation (4). ,in This indicates that there is a causal coupling between regions i and j. This indicates no direct causal relationship, and the matrix is ​​symmetricized and self-connected to ensure the correctness of the graph neural network operations. The adjacency matrices of all windows are stored in... Dictionary with keys middle.

[0024] like Figure 2 As shown, the logic for generating and selecting the dynamic causal adjacency matrix is ​​as follows: At the current time step At that time, due to insufficient available historical windows, a static adjacency matrix pre-calculated based on the complete residual sequence was directly adopted. ;when At that time, according to the window interval to which the current time step belongs. From the dictionary The corresponding matrix is ​​retrieved from the database. If the window matrix has already been generated, it is used directly. If it has not yet been generated, the nearest generated window matrix is ​​selected by backtracking in time. This ensures that an effective adjacency matrix reflecting the causal structure between regions at any given time can be provided.

[0025] exist Figure 1 In step 3, historical load data and the current-moment dynamic causal adjacency matrix selected in step 2 are input into the spatiotemporal graph neural network framework. The framework consists of a spatial feature extraction module and a temporal feature extraction module; the specific implementation of these two modules is not limited. The spatial feature extraction module, based on the graph structure defined by the dynamic causal adjacency matrix, aggregates the load features of neighboring nodes through a graph neural network to extract spatial features containing causal topological information between regions. This allows spatial modeling to rely directly on physically meaningful causal relationships rather than data correlations. The time feature extraction module performs time-series modeling on historical load sequences, capturing multi-scale dynamic evolution features from short-term fluctuations to long-term trends over time. .

[0026] In one specific implementation, the spatial feature extraction module can be implemented using a two-layer graph convolutional network, and the temporal feature extraction module can be implemented using a structure containing one-dimensional convolution and a sparse self-attention encoder. However, the above are only examples, and other spatial and temporal modeling methods are also applicable to this scheme.

[0027] When constructing a two-layer graph convolutional network (GCN), the specific steps are as follows: Step 3-1) Perform linear projection on the input features, including: Load data within the current time window L is the window length, which serves as the initial node feature for the GCN. This feature is then processed through a linear layer and an activation function. Converted to model hidden layer dimension Matching features The process is represented as: (5); in, The trainable weight matrix is ​​the input projection layer. This refers to the process of applying a linear transformation to... The process of feature projection onto input data. For output features The shape of the tensor.

[0028] Step 3-2) Perform propagation operations on the two-layer GCN, including: First, calculate the normalized form of the dynamic adjacency matrix A. ,in yes The degree matrix, yes N × N The identity matrix, features, and a normalized matrix are propagated together with trainable weights. The graph convolution operation aggregates node features with neighborhood structure information to extract spatial features. The propagation rule formula for the l-th layer is: (6); in, It is the node feature matrix of the l-th layer. It is the adjacency matrix with added self-loops, where I is the identity matrix. It is an activation function. Let be the weight matrix for the l-th layer. Spatial features are extracted using GCN, utilizing the normalized adjacency matrix. Correlation between aggregation regions; input feature matrix First, through the projection layer: (7); in, D represents the model dimension. Layer normalization and ReLU activation are used to enhance feature representation. Then, two layers of GCN are applied to normalize the adjacency matrix. Spatial aggregation is performed, and the graph convolution operation is as follows: (8); (9); in, The spatial features output by the first layer GCN This represents the spatial features output by the second-layer GCN.

[0029] Step 3-3) Extract local temporal features using one-dimensional convolution, including: The original payload sequence X is first processed through a one-dimensional convolutional layer. This operation utilizes small-scale convolutional kernels sliding along the time dimension to extract local trends and short-term fluctuation features in the sequence, transforming X into... : (10); Steps 3-4) utilize the Informer sparse self-attention mechanism for computation, including: Will Input the Informer encoder and generate a matrix of query Q, key K, and value V. The encoder applies the ProbSparse self-attention mechanism by calculating the sparsity index. Filter out a small number of important queries It calculates attention only for these queries, thereby reducing computational complexity. Sparsity metrics The calculation is as follows: (11); The final formula for calculating ProbSparse self-attention is: (12); The Informer encoder then applies self-attention distillation to reduce the dimensionality of the features, further improving efficiency, and finally outputs temporal embedding features containing long- and short-term dependency information. .

[0030] In step 4, a cross-attention mechanism is used to deeply fuse temporal and spatial features. Using temporal features... As the query vector Q, with spatial features As the key vector K and value vector V, a trainable weight matrix is ​​used. , , Perform linear projections separately and calculate the cross-attention output according to equation (13). : (13); in, Let be the dimension of the key vector. The cross-attention mechanism enables temporal features to adaptively query and integrate correlation information from spatial causal features, achieving deep interaction between spatiotemporal features. Residual connections are introduced to maintain the stability of temporal features, resulting in spatiotemporal fused features. As shown in equation (14): (14); In step 5, the features of the last time step of the spatiotemporal fusion feature sequence are extracted. After linear projection through the fully connected layer, the predicted load values ​​for each region at the next P time steps are output. As shown in equation (15): (15); The predicted value Y is subjected to Min-Max inverse normalization to restore the load prediction results for each region under the original dimensions.

[0031] Example: This invention was validated on the ISO-England multi-regional historical load dataset, which includes eight interconnected power supply areas. Historical load data was collected at a 1-hour time resolution, covering a full calendar year and including data segments with different seasons and electricity consumption characteristics. Root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) were used as evaluation metrics for predictive performance.

[0032] During the data preprocessing stage, STL decomposition was performed on the load time series of each region with a 24-hour period to extract the residual matrix R. A window factor was set. Predicting step size This determines the causal analysis window. The FCI algorithm is run window-by-window in 120 time steps to generate a dynamic causal adjacency matrix sequence. Figure 3 The time-series evolution results of the dynamic causal adjacency matrix shown indicate that the causal relationships between regions exhibit significant dynamic evolution characteristics with changes in seasons and operating conditions, verifying the necessity of introducing a sliding window for dynamic causal inference.

[0033] To further verify the causal dependence of load flows between regions and to theoretically support the construction of a dynamic causal adjacency matrix, this invention additionally conducted Granger causality verification experiments. Before performing the Granger causality test, the stationarity of the original sequences and high-frequency residual sequences of each load region was first tested (ADF test).

[0034] Table 1 Original Sequence ADF Verification Result Table

[0035] Table 2 Residual Sequence ADF Verification Result Table

[0036] (Note: significance index " "This indicates that at a significance level of α = 0.05, the null hypothesis is rejected, the series is stationary, and the Critical Value is the confidence threshold." The test results show that the ADF statistics of the original and residual sequences in all regions are below the 5% significance threshold, and the p-values ​​are all significantly less than 0.05, indicating that all sequences are stationary time series and meet the prerequisites for Granger causality test. The ADF test results are shown in Tables 1 and 2.

[0037] To ensure data stationarity, a two-way Granger causality test was used to analyze whether statistically significant causal effects exist between load regions. The lag order was set to 24 to capture short-term dependencies in hourly load sequences. For any two regions i and j, the hypotheses "region i is not a Granger cause of region j" and "region j is not a Granger cause of region i" were tested respectively. When the p-value was less than the significance level of 0.05, the null hypothesis was rejected, and a Granger causal relationship was considered to exist.

[0038] like Figure 4 , Figure 5 As shown, significant causal relationships exist between the original and residual sequences in each load region, and the intensity of influence between different regions varies significantly; that is, the darker the color, the more significant the causal relationship. Among these, the high-frequency residual sequences, compared to the original sequences, more clearly highlight the contrast in the strength of causal relationships between regions, avoiding interference from trend and periodic components. This further demonstrates the rationality and superiority of constructing a spatial causal adjacency matrix based on the high-frequency residual sequences after STL decomposition.

[0039] To comprehensively compare prediction performance, the comparison methods are set as follows: (1) a spatiotemporal graph neural network baseline based on the generation of a static adjacency matrix based on geographical proximity, denoted as M1; (2) a spatiotemporal graph neural network baseline based on the generation of a static adjacency matrix based on an electrical distance threshold, denoted as M2; (3) a comparison method based on data-driven adaptive generation of a dynamic adjacency matrix, denoted as M3; (4) a traditional time series prediction baseline, denoted as M4. The method adopted in this invention is denoted as M5.

[0040] Table 3 Comparison method prediction error results

[0041] As shown in Table 3, the prediction error comparison results demonstrate that this invention achieves the best RMSE and MAPE indices compared to other methods. The RMSE is significantly lower than the static adjacency matrix baseline method, verifying the effectiveness of the dynamic causal adjacency matrix in improving prediction accuracy. Compared to methods based on data-driven dynamic adjacency matrix generation, the causal inference-based method proposed in this invention can more accurately characterize spatial correlation, further improving prediction accuracy. Experimental results demonstrate the advantages of causal structure over statistical correlation in prediction performance, proving the effectiveness and practicality of this scheme.

Claims

1. A method for regional short-term load forecasting that tracks the causal relationship of load flows, characterized in that, Includes the following steps: Step 1: Obtain the historical load data sequence of the target area, perform normalization processing, decompose the load sequence of each area into trend term, seasonal term and residual term, and extract the residual term as high-frequency residual component; Step 2: Apply a sliding window mechanism to the high-frequency residual components, define the causal analysis window and step size, and divide the residual sequence into a series of time window subsets; Within each time window, conditional independence is tested on the residual sequences between regions to identify causal dependency edges between regions; based on the identified causal edges, the adjacency matrix corresponding to that window is constructed, thereby generating a dynamic causal adjacency matrix sequence that evolves over time. Step 3: Input the historical load data and the dynamic causal adjacency matrix corresponding to the current moment generated in Step 2 into the classic spatiotemporal graph neural network framework to extract the spatial causal features between regions and the dynamic evolution features in the time dimension, respectively. Step 4: Fuse spatial and temporal features to obtain spatiotemporal fusion features; Step 5: Input the spatiotemporal fusion features into the fully connected layer, output the regional load prediction values ​​for multiple future time steps, and perform inverse normalization to obtain the final prediction result.

2. The method according to claim 1, characterized in that, In step 1, the input data is processed. Min-Max normalization is performed, where T is the time step and N is the number of regions; the load sequence for each region i after normalization is then processed. ,by For periodicity, the Seasonal Trend Decomposition (STL) algorithm is applied to decompose the original sequence into trend terms. Seasonal items and residuals , t This is the index for the current time step.

3. The method according to claim 2, characterized in that, Separate the original sequence into trend terms. Seasonal items and residuals As shown in equation (1): （1）； Take the residual terms of each region Forming the residual matrix This serves as input for subsequent causal analysis.

4. The method according to any one of claims 1 to 3, characterized in that, In step 2, the causal analysis window size is defined. With window size M, The prediction window is set The value is determined by the window multiplier M, and the calculation method is shown in equation (2): （2）； by For step size, from Start sliding to Generate time window sequence ,in , , For each time window, extract the corresponding subset from the residual matrix R. ,right The FCI algorithm is applied to N region sequences to generate the dynamic causal adjacency matrix of the window. And store the matrix of all windows in a window range Dictionary with keys middle.

5. The method according to claim 4, characterized in that, when At that time, a static adjacency matrix generated in the same way using the complete residual sequence is adopted. ;when At that time, based on the window to which the current time step belongs, from Select the corresponding matrix from the list. If the corresponding window matrix has not yet been generated, select the window matrix that is closest in time.

6. The method according to claim 4, characterized in that, In step 2, the classic Fast Causal Inference (FCI) algorithm is applied to perform conditional independence tests on the residual sequences between regions. Specifically, the FCI algorithm is used to test the conditional independence of the residual sequences within the window. and Perform a conditional independence test, using the partial correlation coefficient under the condition set Z. Based on this, the test statistic P-value is calculated as shown in equation (3): （3）； Where n is the number of samples within the window. The number of elements in the condition set Z at the significance level. Below, by comparing the Z-value with the critical value of the standard normal distribution, the directed causal dependencies between regions are determined, and the following types of causal edges are output: express right One-way causal influence; Indicates a two-way causal relationship; It indicates a causal relationship with an uncertain direction.

7. The method according to claim 6, characterized in that, Based on the edge information, construct the adjacency matrix. Initially a zero-based matrix, after filling according to edge type, the matrix is ​​symmetricized and self-joined to finally generate a 0-1 adjacency matrix; where This indicates that there is a causal coupling between regions i and j. This indicates that there is no direct causal relationship.

8. The method according to claim 1, characterized in that, In step 3, spatial causal features between regions are extracted through the spatial feature extraction module. The spatial feature extraction module of the spatiotemporal graph neural network framework takes historical load data and dynamic causal adjacency matrix as input, and aggregates the features of neighboring nodes to extract spatial features containing causal topological information between regions. The dynamic evolution features in the time dimension are extracted through the temporal feature extraction module. The temporal feature extraction module takes historical load sequence as input and extracts short-term and long-term dynamic evolution features in the time dimension.

9. The method according to claim 1, 2, 3, 5, 6, or 8, characterized in that, In step 4, the time series features output by the time feature extraction module are used. As the query vector Q, the spatial features output by the spatial feature extraction module As the key vector K and value vector V, a trainable weight matrix is ​​used. , , After linear projection, the cross-attention fusion features are calculated according to equation (4). : （4）； And introduce residual connections, Compared with the original time series characteristics Adding them together yields the final spatiotemporal fusion feature. As shown in equation (5): （5）； in, Key vector K The feature dimensions.

10. The method according to claim 9, characterized in that, In step 5, the features of the last time step of the spatiotemporal fusion feature sequence are extracted. After linear projection through a fully connected layer, the predicted values ​​for the next P time steps are output. As shown in equation (6): （6） The predicted value Y is inversely normalized to obtain the final load forecast results for each region; among them, This is the trainable weight matrix for the fully connected layer. This is the corresponding bias term.

Images