A power distribution network multi-node short-term power load prediction method

Abstract

The present application relates to the technical field of power load prediction, in particular to a power distribution network multi-node short-term power load prediction method. In order to solve the technical problems of the existing power distribution network multi-node short-term power load prediction method, such as mixed multi-frequency characteristics, weak long sequence modeling capability, insufficient irregular node association mining, and lack of adaptability of space-time feature fusion, a new power distribution network multi-node short-term power load prediction method is provided, which comprises the following steps: constructing a historical load prediction sample sequence, performing discrete wavelet decomposition, building a graph connection mask and a space weight matrix, calculating a local space correlation index, extracting spatial correlation features and time dependence features, obtaining fused space-time features through an adaptive gating fusion module, and obtaining complete load prediction results of each load node at a future time through a frequency prediction module. The method improves the accuracy and adaptability of complex power distribution network multi-node short-term load prediction.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, specifically a method for short-term power load forecasting of multiple nodes in a distribution network. Background Technology

[0002] With the acceleration of urbanization and the advancement of energy structure transformation, the load of multiple nodes in the distribution network is simultaneously affected by complex factors such as weather conditions, holidays, electricity consumption behavior, and electrical connections between nodes, exhibiting significant spatial correlation and multi-timescale fluctuation characteristics. Short-term power load forecasting is the core foundation for the safe and economical operation of the distribution network, and its accuracy directly affects the efficiency of renewable energy absorption, reserve capacity configuration, and power supply reliability. However, existing methods for short-term power load forecasting of multiple nodes in the distribution network still have the following pain points that urgently need to be addressed:

[0003] 1) Existing multi-node short-term power load forecasting methods for distribution networks typically model the original load sequence that simultaneously contains long-term trends, periodic changes, and short-term fluctuations in a unified manner, which makes it difficult to fully characterize the variation patterns of different frequency components. At the same time, traditional time series models such as RNN (Recurrent Neural Network) and LSTM (Long Short-Term Memory Network) suffer from problems such as low efficiency of serial computation and gradual decay of early historical information under long input windows, which limit the extraction of long-term dependent features.

[0004] 2) Distribution network nodes form an irregular network. The relationship between nodes is not only affected by the physical topology, but may also be affected by the similarity of load change patterns. Existing short-term power load forecasting methods for multiple nodes in distribution networks mostly rely on fixed topology or local neighborhood information. They do not make sufficient use of the relationships between nodes that are not directly connected but have strong correlations in load changes, and it is difficult to fully extract multi-hop and long-distance spatial correlation features.

[0005] 3) Different load nodes and different frequency components of the same load node have significantly different degrees of dependence on spatial correlation information and their own historical patterns. Moreover, this degree of dependence will change continuously with the operating status and external environment. Existing multi-node short-term power load forecasting methods for distribution networks mostly use fixed ratio fusion or fixed threshold for feature selection, which makes it difficult to dynamically adjust the contribution of time and spatial features according to the current state of the node, and easily leads to insufficient utilization of features. Summary of the Invention

[0006] To address the technical problems of existing methods for predicting short-term power load across multiple nodes in distribution networks, such as mixed multi-frequency features, weak long-sequence modeling capabilities, insufficient mining of irregular node associations, and lack of adaptability in spatiotemporal feature fusion, this invention provides a new method for predicting short-term power load across multiple nodes in distribution networks.

[0007] This invention is achieved using the following technical solution:

[0008] A method for short-term power load forecasting at multiple nodes in a distribution network includes the following steps:

[0009] Step 1: Obtain historical load time series data of multiple load nodes, electrical topology connection relationships between multiple load nodes, and auxiliary features (auxiliary features refer to temperature, humidity, hour, weekday and holiday identifiers, etc.). Perform data preprocessing on the historical load time series data and auxiliary features, and construct a historical load prediction sample sequence using a sliding window method.

[0010] Step 2: Perform discrete wavelet decomposition on the historical load time series data of each load node in the historical load prediction sample sequence constructed in Step 1 to obtain multiple frequency sub-signals composed of low-frequency trend sub-signals and multiple high-frequency detail sub-signals;

[0011] Step 3: Construct a graph connection mask and spatial weight matrix based on the electrical topology connections between multiple load nodes and the temporal similarity between them;

[0012] Step 4: Based on the spatial weight matrix established in Step 3, calculate the local spatial correlation index between each load node within the current sliding window, and perform standardization processing to obtain the spatial correlation guidance quantity.

[0013] Step 5: For the multiple frequency sub-signals obtained from Step 2, the Graph Transformer (Graph Self-Attention Transformer Network) is used to extract the spatial correlation features of each load node in parallel, and the Mamba (Selective State Space Network) is used to extract the temporal dependence features of each load node in parallel.

[0014] Step 6: Input the spatial correlation guidance quantity, spatial association features, and temporal dependency features into the adaptive gating fusion module to generate fusion weights, and then perform weighted fusion of spatial association features and temporal dependency features based on the generated fusion weights to obtain fused spatiotemporal features;

[0015] Step 7: Input the fused spatiotemporal features obtained in Step 6 into the frequency prediction module to predict the future time prediction results of each load node on each frequency sub-signal, and reconstruct and sum the future time prediction results to obtain the complete load prediction results of each load node at the future time.

[0016] Furthermore, in step two, the expression for discrete wavelet decomposition is:

[0017] ;

[0018] in, Indicates the first The load node at the ... Historical load values ​​at any given time Represents the th wavelet after discrete wavelet decomposition. Low-frequency trend sub-signals of the layer Represents the th wavelet after discrete wavelet decomposition. Layer high-frequency detail sub-signals, This indicates the number of discrete wavelet decomposition layers.

[0019] Furthermore, in step three, when there is a physical direct path between two load nodes, or when there is no physical direct path but the temporal similarity between the two load nodes reaches a set threshold (as is known to those skilled in the art, this set threshold is obtained by calculating the Pearson correlation coefficient), a corresponding graph connectivity mask is set. (Direct information interaction is allowed); otherwise, the corresponding graph connection mask is set. (Prevent direct information exchange).

[0020] Furthermore, in step three, when there is a direct physical path between two load nodes, the corresponding spatial weight is assigned as 1. When there is no direct physical path between two load nodes but the temporal similarity between the two load nodes reaches a set threshold, the corresponding spatial weight is calculated and assigned based on the Pearson coefficient. When there is no direct physical path between two load nodes and the temporal similarity between the two load nodes does not reach the set threshold, the corresponding spatial weight is assigned as 0. Finally, the assigned spatial weight matrix is ​​normalized to obtain the normalized spatial weight matrix. ,in, Indicates the first The load node and the first Normalized spatial weights among load nodes.

[0021] Furthermore, in step four, the formula for calculating the local spatial correlation index is:

[0022] ;

[0023] in, Indicates the first The load node and the first Local spatial correlation index of each frequency sub-signal Indicates the first The load node at the ... The statistical feature vector corresponding to each frequency sub-signal This indicates that all load nodes are at the . The average eigenvalues ​​corresponding to each frequency sub-signal This indicates that all load nodes are at the 1st The degree of dispersion of the corresponding eigenvalues ​​on each frequency sub-signal This represents a very small positive number that prevents the denominator from being zero;

[0024] Furthermore, in step five, the Graph Transformer is used to extract the spatial correlation features of each load node in parallel. The expression is:

[0025] ;

[0026] in, , and These represent the query vector, key vector, and value vector obtained after linear mapping of the load node, respectively. Indicates the feature dimension; Represents the trainable spatial structure bias coefficient; This represents the normalized exponential function operation;

[0027] Furthermore, in step five, the specific steps for extracting the time dependency features of each load node in parallel using Mamba are as follows: For the first... The first frequency sub-signal concatenates the historical load time-series data of each load node at that frequency with auxiliary features. After linear mapping, it is transformed into a load node feature vector of a unified dimension. This vector is then input into Mamba to extract the long-term trend, daily cycle, weekly cycle, short-term fluctuations, and multi-scale time-series dependencies of the current load node, thus obtaining the first frequency sub-signal. The load node at the ... Time-dependent features on each frequency sub-signal .

[0028] Furthermore, in step six, the weights are merged. The expression:

[0029] ;

[0030] in, This indicates a feature concatenation operation. This represents the Sigmoid activation function. and These represent the trainable weight parameters and bias parameters of the adaptive gating fusion module, respectively.

[0031] Furthermore, in step six, spatiotemporal features are integrated. The expression is:

[0032] ;

[0033] in, This indicates element-wise multiplication.

[0034] Furthermore, in step one, the length of the sliding window is... The predicted length is The dimension of the historical load prediction sample sequence is ,in, This represents the number of features for each load node in the distribution network at each sampling time. This represents the total number of load nodes.

[0035] Furthermore, in step seven, the frequency prediction module adopts a prediction structure with multi-step output capability. That is, for the low-frequency trend sub-signal and the high-frequency detail sub-signal, independent prediction output mapping branches are established to decouple the output of each load node in the future. Prediction results of low-frequency trend sub-signals at each sampling time. and high-frequency detail sub-signal prediction results ,in Indicates the sampling period and The prediction results of each frequency sub-signal corresponding to each load node are reconstructed and summed to obtain the prediction results of each load node in the future. Complete load forecast results at each sampling time point Its expression is:

[0036] ;

[0037] in, Indicates the current moment.

[0038] Principle Description: This invention first employs discrete wavelet decomposition to perform multi-frequency decomposition on historical load prediction sample sequences to separate long-term trends and short-term fluctuations at different scales. Subsequently, it constructs load node association relationships by combining the electrical topology relationships between multiple load nodes in the distribution network and the temporal similarity among them. It then uses a local spatial correlation index to measure the spatial correlation between each load node and its associated load nodes at different frequency components, serving as guiding information for subsequent feature fusion. Based on this, it uses Graph Transformer and Mamba to extract the spatial association features and temporal dependency features of each load node, respectively, and dynamically adjusts the fusion ratio of the two types of features through an adaptive gating module. Finally, it uses a frequency prediction module to predict the complete load prediction results for each load node at future times.

[0039] The beneficial effects of this invention are as follows: The short-term power load forecasting method for multiple nodes in the distribution network described in this invention improves the accuracy and adaptability of short-term load forecasting for multiple nodes in complex distribution networks through multi-frequency decomposition, spatial correlation quantification, parallel extraction of spatiotemporal features and adaptive gating fusion. It can be applied to load forecasting in scenarios such as substations, feeders, distribution areas and parks, and provides data support for power grid operation and dispatch. Attached Figure Description

[0040] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of the short-term power load forecasting method for multiple nodes in a power distribution network according to the present invention. Detailed Implementation

[0043] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0045] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0046] A method for short-term power load forecasting at multiple nodes in a distribution network includes the following steps:

[0047] Step 1: Obtain historical load time-series data for multiple load nodes, the electrical topology connections between these nodes, and auxiliary characteristics (such as temperature, humidity, hour, day of the week, and holiday indicators). Preprocess the historical load time-series data and auxiliary characteristics, and construct a historical load prediction sample sequence using a sliding window method. The sliding window length is [value missing]. The predicted length is The dimension of the historical load prediction sample sequence is ,in, This represents the number of features for each load node in the distribution network at each sampling time. This represents the total number of load nodes.

[0048] In practice, the data preprocessing for historical load time series data and auxiliary features includes filling in missing values, correcting outliers, and aligning them according to a uniform sampling interval.

[0049] Step 2: Perform discrete wavelet decomposition on the historical load time series data of each load node in the historical load prediction sample sequence constructed in Step 1 to obtain multiple frequency sub-signals composed of a low-frequency trend sub-signal and multiple high-frequency detail sub-signals. The expression for this is:

[0050] ;

[0051] in, Indicates the first The load node at the ... Historical load values ​​at any given time Represents the th wavelet after discrete wavelet decomposition. Low-frequency trend sub-signals of the layer Represents the th wavelet after discrete wavelet decomposition. Layer high-frequency detail sub-signals, This indicates the number of discrete wavelet decomposition layers.

[0052] Step 3: Construct a graph connection mask and spatial weight matrix based on the electrical topology connections between multiple load nodes and the temporal similarity between them;

[0053] In practical implementation, for graph connectivity masks, when there is a physical direct path between two load nodes, or when there is no physical direct path but the temporal similarity between the two load nodes reaches a set threshold (as is known to those skilled in the art, this set threshold is calculated using the Pearson correlation coefficient), a corresponding graph connectivity mask is set. (Direct information interaction is allowed); otherwise, the corresponding graph connection mask is set. (Preventing direct information interaction);

[0054] For the spatial weight matrix, when there is a direct physical path between two load nodes, the corresponding spatial weight is assigned as 1. When there is no direct physical path between two load nodes but their temporal similarity reaches a set threshold, the corresponding spatial weight is calculated and assigned based on the Pearson coefficient. When there is no direct physical path between two load nodes and their temporal similarity does not reach the set threshold, the corresponding spatial weight is assigned as 0. Finally, the weighted spatial weight matrix is ​​normalized to obtain the normalized spatial weight matrix. ,in, Indicates the first The load node and the first Normalized spatial weights among load nodes;

[0055] Step 4: Based on the spatial weight matrix established in Step 3, calculate the local spatial correlation index between each load node within the current sliding window, and perform standardization processing to obtain the spatial correlation guidance quantity.

[0056] The formula for calculating the local spatial correlation index is as follows:

[0057] ;

[0058] in, Indicates the first The load node and the first Local spatial correlation index of each frequency sub-signal Indicates the first The load node at the ... The statistical feature vector corresponding to each frequency sub-signal This indicates that all load nodes are at the 1st The average eigenvalues ​​corresponding to each frequency sub-signal This indicates that all load nodes are at the 1st The degree of dispersion of the corresponding eigenvalues ​​on each frequency sub-signal This represents a very small positive number that prevents the denominator from being zero;

[0059] Step 5: For the multiple frequency sub-signals obtained from Step 2, the Graph Transformer (Graph Self-Attention Transformer Network) is used to extract the spatial correlation features of each load node in parallel, and the Mamba (Selective State Space Network) is used to extract the temporal dependence features of each load node in parallel.

[0060] In practice, the Graph Transformer is used to extract the spatial correlation features of each load node in parallel. The expression is:

[0061] ;

[0062] in, , and These represent the query vector, key vector, and value vector obtained after linear mapping of the load node, respectively. Indicates the feature dimension; Represents the trainable spatial structure bias coefficient; This represents the normalized exponential function operation;

[0063] The specific steps for extracting the time dependency features of each load node in parallel using Mamba are as follows: For the first node... The first frequency sub-signal concatenates the historical load time-series data of each load node at that frequency with auxiliary features. After linear mapping, it is transformed into a load node feature vector of a unified dimension. This vector is then input into Mamba to extract the long-term trend, daily cycle, weekly cycle, short-term fluctuations, and multi-scale time-series dependencies of the current load node, thus obtaining the first frequency sub-signal. The load node at the ... Time-dependent features on each frequency sub-signal ;

[0064] Step 6: Input the spatial correlation guidance quantity, spatial association features, and temporal dependency features into the adaptive gating fusion module to generate fusion weights, and then perform weighted fusion of spatial association features and temporal dependency features based on the generated fusion weights to obtain fused spatiotemporal features;

[0065] Among them, the fusion weight The expression:

[0066] ;

[0067] in, This indicates a feature concatenation operation. This represents the Sigmoid activation function. and These represent the trainable weight parameters and bias parameters of the adaptive gating fusion module, respectively.

[0068] Integrating spatiotemporal features The expression is:

[0069] ;

[0070] in, This indicates element-wise multiplication;

[0071] Step 7: Input the fused spatiotemporal features obtained in Step 6 into the frequency prediction module to predict the future time prediction results of each load node on each frequency sub-signal, and reconstruct and sum the future time prediction results to obtain the complete load prediction results of each load node at the future time.

[0072] In practical implementation, the frequency prediction module adopts a prediction structure with multi-step output capability. That is, for the low-frequency trend sub-signal and the high-frequency detail sub-signal, independent prediction output mapping branches are established to decouple the output of each load node in the future. Prediction results of low-frequency trend sub-signals at each sampling time. and high-frequency detail sub-signal prediction results ,in Indicates the sampling period and The prediction results of each frequency sub-signal corresponding to each load node are reconstructed and summed to obtain the prediction results of each load node in the future. Complete load forecast results at each sampling time point Its expression is:

[0073] ;

[0074] in, Indicates the current moment.

[0075] To verify the effectiveness of the multi-node short-term power load forecasting method for distribution networks described in this invention, experiments were first conducted using the SMART-DS (Synthetic Models for Advanced, Realistic Testing: Distribution Systems and Scenarios) dataset developed by the U.S. Department of Energy. The SMART-DS dataset is a large-scale, realistic synthetic distribution network dataset designed for distribution system research. It contains complete distribution network topology and multi-timescale load, meteorological, and other scenario data, and supports sub-modeling based on substations and feeders, making it suitable for multi-node load forecasting and spatial correlation analysis in distribution networks.

[0076] A typical distribution feeder in the Greensboro (GSO) area of ​​the SMART-DS dataset was selected as the experimental object. 32 load nodes with continuous load records and complete topological connections were selected from this distribution feeder to construct a multi-node load prediction dataset, and a load node diagram was established based on the connection relationship of the distribution line.

[0077] Historical load time-series data were time-aligned using a uniform 15-minute sampling interval and divided into training, validation, and test sets in chronological order, with proportions of 70%, 10%, and 20%, respectively. A sliding window was used to construct historical load prediction sample sequences, using the historical load of the past 48 hours to predict the load of the next 6 hours, with the input window length... Prediction window length .

[0078] To reduce the impact of differences in the magnitude of different load nodes on model training, the mean and standard deviation of the training set are used to standardize the historical load sequence data, and the results are restored to the original scale after prediction is completed for performance evaluation.

[0079] The experiment used MAE (mean absolute error), RMSE (root mean square error), and MAPE (mean absolute percentage error) as evaluation indicators.

[0080] LSTM (Long Short-Term Memory), PatchTST (Patch Temporal Transformer), iTransformer (Inverse Transformer), STGCN (Spatiotemporal Graph Convolutional Network), and Graph WaveNet (Graph Dilated Causal Convolutional Temporal Network) were selected as comparison methods. All methods used the same data partitioning, input length, and prediction length to ensure the comparability of experimental results.

[0081] The prediction results of different methods on the SMART-DS dataset are shown in the table below.

[0082]

[0083] As shown in Table 1, the method of this invention achieves superior prediction results in all three evaluation indicators: MAE, RMSE, and MAPE. Compared with Graph WaveNet, which has better overall performance, the method of this invention reduces MAE from 0.344kW to 0.321kW, RMSE from 0.548kW to 0.514kW, representing reductions of approximately 6.7% and 6.2% respectively, and MAPE from 5.06% to 4.89%. The results demonstrate that the method of this invention can effectively improve the accuracy of short-term load forecasting for multiple nodes in distribution networks.

[0084] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. A method for short-term power load forecasting at multiple nodes in a distribution network, characterized in that, Includes the following steps: Step 1: Obtain historical load time series data of multiple load nodes, electrical topology connection relationships and auxiliary features between multiple load nodes, perform data preprocessing on historical load time series data and auxiliary features, and construct historical load prediction sample sequences using a sliding window method; Step 2: Perform discrete wavelet decomposition on the historical load time series data of each load node in the historical load prediction sample sequence constructed in Step 1 to obtain multiple frequency sub-signals composed of low-frequency trend sub-signals and multiple high-frequency detail sub-signals; Step 3: Construct a graph connection mask and spatial weight matrix based on the electrical topology connections between multiple load nodes and the temporal similarity between them; Step 4: Based on the spatial weight matrix established in Step 3, calculate the local spatial correlation index between each load node within the current sliding window, and perform standardization processing to obtain the spatial correlation guidance quantity. Step 5: For the multiple frequency sub-signals obtained from Step 2, use Graph Transformer to extract the spatial correlation features of each load node in parallel, and use Mamba to extract the temporal dependence features of each load node in parallel. Step 6: Input the spatial correlation guidance quantity, spatial association features, and temporal dependency features into the adaptive gating fusion module to generate fusion weights, and then perform weighted fusion of spatial association features and temporal dependency features based on the generated fusion weights to obtain fused spatiotemporal features; Step 7: Input the fused spatiotemporal features obtained in Step 6 into the frequency prediction module to predict the future time prediction results of each load node on each frequency sub-signal, and reconstruct and sum the future time prediction results to obtain the complete load prediction results of each load node at the future time.

2. The method for short-term power load forecasting of multiple nodes in a distribution network according to claim 1, characterized in that, In step two, the expression for discrete wavelet decomposition is: ； in, Indicates the first The load node at the ... Historical load values ​​at any given time Represents the th wavelet after discrete wavelet decomposition. Low-frequency trend sub-signals of the layer Represents the th wavelet after discrete wavelet decomposition. Layer high-frequency detail sub-signals, This indicates the number of discrete wavelet decomposition layers.

3. The method for short-term power load forecasting of multiple nodes in a distribution network according to claim 2, characterized in that, In step three, when there is a physical direct path between two load nodes, or when there is no physical direct path but the temporal similarity between the two load nodes reaches a set threshold, the corresponding graph connectivity mask is set. Otherwise, set the corresponding graph connection mask. .

4. The method for short-term power load forecasting of multiple nodes in a distribution network according to claim 3, characterized in that, In step three, when there is a direct physical path between two load nodes, the corresponding spatial weight is assigned as 1. When there is no direct physical path between two load nodes but their temporal similarity reaches a set threshold, the corresponding spatial weight is calculated and assigned based on the Pearson coefficient. When there is no direct physical path between two load nodes and their temporal similarity does not reach the set threshold, the corresponding spatial weight is assigned as 0. Finally, the assigned spatial weight matrix is ​​normalized to obtain the normalized spatial weight matrix. ,in, Indicates the first The load node and the first Normalized spatial weights among load nodes.

5. The method for short-term power load forecasting of multiple nodes in a distribution network according to claim 4, characterized in that, In step four, the formula for calculating the local spatial correlation index is: ； in, Indicates the first The load node and the first Local spatial correlation index of each frequency sub-signal Indicates the first The load node at the ... The statistical feature vector corresponding to each frequency sub-signal This indicates that all load nodes are at the 1st The average eigenvalues ​​corresponding to each frequency sub-signal This indicates that all load nodes are at the 1st The degree of dispersion of the corresponding eigenvalues ​​on each frequency sub-signal This represents a very small positive number that prevents the denominator from being zero.

6. The method for short-term power load forecasting of multiple nodes in a distribution network according to claim 5, characterized in that, In step five, the Graph Transformer is used to extract the spatial correlation features of each load node in parallel. The expression is: ； in, , and These represent the query vector, key vector, and value vector obtained after linear mapping of the load node, respectively. Indicates the feature dimension; Represents the trainable spatial structure bias coefficient; This indicates the operation of the normalized exponential function.

7. The method for short-term power load forecasting of multiple nodes in a distribution network according to claim 6, characterized in that, In step five, the specific steps for extracting the time dependency features of each load node in parallel using Mamba are as follows: For the first... The first frequency sub-signal concatenates the historical load time-series data of each load node at that frequency with auxiliary features. After linear mapping, it is transformed into a load node feature vector of a unified dimension. This vector is then input into Mamba to extract the long-term trend, daily cycle, weekly cycle, short-term fluctuations, and multi-scale time-series dependencies of the current load node, thus obtaining the first frequency sub-signal. The load node at the ... Time-dependent features on each frequency sub-signal .

8. The method for short-term power load forecasting of multiple nodes in a distribution network according to claim 7, characterized in that, In step six, the weights are merged. The expression: ； in, This indicates a feature concatenation operation. This represents the Sigmoid activation function. and These represent the trainable weight parameters and bias parameters of the adaptive gating fusion module, respectively. Integrating spatiotemporal features The expression is: ； in, This indicates element-wise multiplication.

9. A method for short-term power load forecasting of multiple nodes in a distribution network according to claim 8, characterized in that, In step one, the length of the sliding window is The predicted length is The dimension of the historical load prediction sample sequence is ,in, This represents the number of features for each load node in the distribution network at each sampling time. This represents the total number of load nodes.

10. A method for short-term power load forecasting of multiple nodes in a distribution network according to claim 9, characterized in that, In step seven, the frequency prediction module adopts a prediction structure with multi-step output capability. That is, for the low-frequency trend sub-signal and the high-frequency detail sub-signal, independent prediction output mapping branches are established to decouple the output of each load node in the future. Prediction results of low-frequency trend sub-signals at each sampling time. and high-frequency detail sub-signal prediction results ,in Indicates the sampling period and The prediction results of each frequency sub-signal corresponding to each load node are reconstructed and summed to obtain the prediction results of each load node in the future. Complete load forecast results at each sampling time point Its expression is: ； in, Indicates the current moment.

Images