Multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing

Abstract

The application discloses a multi-factor short-term power load prediction system based on adaptive weight sharing of a graph structure, and comprises the following steps: multi-dimensional power load data preprocessing, implicit graph structure and dynamic space correlation modeling, adaptive weight sharing graph convolution design, multi-factor enhancement module fusion, and end-to-end space-time graph network training.The application solves the problem of poor data quality after new energy access through multi-dimensional data preprocessing; through the cooperation of dynamic implicit graph and weight sharing graph convolution, the problem of the inability of a static graph to capture dynamic implicit correlation and the loss of space-time characteristics is solved; through weight sharing and multi-factor fusion, the problems of model parameter redundancy and insufficient feature fusion are solved, high-precision and efficient short-term power load prediction is realized, and the needs of users are met.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing. Background Technology

[0002] Short-term power load forecasting serves as a core technical support for power system dispatch optimization, energy supply and demand balance, and demand-side response. Existing short-term power load forecasting mostly uses graph networks for forecasting. With the rapid development of smart grids and distributed energy, the power system is shifting from traditional total load forecasting to highly volatile net load forecasting. The randomness of new energy sources causes the load curve to exhibit sawtooth fluctuations.

[0003] Traditional models still rely on static rules and fixed parameters, which leads to the following problems: First, when a large number of new energy sources are connected to the graph network, it will exacerbate network load fluctuations, and problems such as noise, outliers, missing values, and spatiotemporal deviations in the original data will become more prominent. Second, traditional explicit graphs cannot capture dynamic / implicit spatial relationships, and the separation of spatiotemporal features in modeling is prone to information loss. Third, traditional graph networks have redundant parameters and insufficient fusion of multiple factors and spatiotemporal features, which have limitations.

[0004] The reasons for this problem are as follows: Problem 1: The randomness and intermittency of new energy sources such as photovoltaics and wind power lead to increased net load fluctuations. Data collection is easily affected by external interference, resulting in a decrease in the quality of raw data and an increase in noise, outliers, and missing values. Problem 2: Traditional graph structures are difficult to dynamically adapt to changes in power system load relationships, especially in capturing implicit relationships between functionally similar but geographically distant buildings. The spatiotemporal feature separation modeling method leads to the loss of feature information during transmission, resulting in a decrease in the model's overall predictive ability. Problem 3: When processing large-scale data, the number of parameters in traditional graph network models increases dramatically, leading to increased model complexity and longer training and inference times. At the same time, the models are insufficient in integrating multiple factors and spatiotemporal features, and cannot fully utilize this external information to improve prediction accuracy, thus having limitations. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing, comprising the following steps: S1. Multi-dimensional power load data preprocessing: collect multi-source data, then send the data to the data cleaning module, then use the 3σ criterion and isolated forest algorithm to detect outliers, then use KNN time series interpolation to complete the data, and finally divide the training set, validation set and test set in a 7:2:1 ratio; S2. Implicit graph structure and dynamic spatial association modeling: The preprocessed composite data is sent to the node embedding module to generate a trainable embedding vector for each node. Then, the association degree between nodes is calculated by cosine similarity to dynamically construct an adaptive graph structure. Next, L1 regularization is introduced to achieve graph structure sparsification and output the dynamic adjacency matrix to the downstream network. S3. Adaptive weight-sharing graph convolution design, dynamic adjacency matrix and node features, initial grouping according to node functional features, assign basic convolution kernels to each group of nodes, and then achieve personalized weight adjustment through fine-tuning matrix to complete spatial feature extraction and output to the temporal modeling unit; S4. Multi-factor enhancement module fusion: external features are normalized and encoded, and the weights of each feature are dynamically allocated through the attention weighting module. Then, the feature is gated and fused with the spatiotemporal features. Finally, the fused multi-factor enhancement node representation is sent to the prediction output layer. S5. End-to-end spatiotemporal graph network training with multi-factor enhanced features: First, the feature is input into the graph attention-gated recurrent network for spatiotemporal joint feature extraction. Then, it enters the prediction output layer to generate the load prediction value for the future time step. Next, the Huber loss function is used to evaluate the prediction error. Finally, it enters the parameter optimization module to update the network weights through gradient pruning and backpropagation.

[0007] Preferably, the implicit graph structure and dynamic space association modeling in step S2 specifically includes: The node embedding module assigns a d-dimensional trainable embedding vector to each building load node, calculates the similarity between node embeddings through an attention mechanism, generates an implicit adjacency matrix, and initializes the nodes. Combined with L1 regularization constraints, it ensures the sparsity and interpretability of the graph structure.

[0008] Preferably, the implicit adjacency matrix is ​​generated using a dynamic update mechanism based on attention weighting.

[0009] Preferably, in step S3, the adaptive weight-sharing graph convolution design specifically includes: A weight-sharing pool is established to assign a unified convolution kernel to groups of nodes with similar functions. Node-level fine-tuning is achieved through adaptive adjustment of the matrix. Parameters are utilized during spatial feature aggregation to reduce the number of model parameters.

[0010] Preferably, the weight sharing mechanism adopts a grouping strategy, dynamically grouping nodes based on their functional characteristics and load pattern similarity, with each group of nodes sharing the basic convolutional kernel. And by adjusting the matrix Enables fine-tuning of node-level weights.

[0011] Preferably, in step S4, the multi-factor enhancement module fusion specifically includes a meteorological feature encoding module, a date attribute encoding module, and a building feature encoding module: The meteorological feature coding module is used to process temperature, humidity, and light intensity data; The date attribute encoding module is used to process holiday and seasonal information; The building feature encoding module is used to extract functional type features and achieves dynamic fusion of multi-source features through an attention weighting mechanism. The resulting multi-factor enhanced node representation is then sent to the prediction output layer for load prediction.

[0012] Preferably, the multi-factor fusion adopts a gating mechanism, specifically implemented as follows: Multi-source data is mapped to a unified dimension by feature normalization, attention weights are used to dynamically adjust the contribution of each feature, and a fully connected layer is used for feature compression. The resulting multi-factor enhanced node representation is transmitted to the prediction output layer to complete the final load prediction.

[0013] Preferably, the dynamic graph structure in step S2 and the weight sharing mechanism in step S3 work together, specifically including: The dynamic adjacency matrix generated by the implicit graph learner is directly used as the input structure for the weight-shared graph convolution. The weight-shared pool automatically adjusts the grouping strategy according to the dynamic graph structure, achieving coordinated optimization of spatial association capture and parameter efficiency.

[0014] Preferably, the end-to-end training in step S5 employs a phased optimization strategy, specifically including: First, the parameters of the implicit graph structure learner are fixed, and the weight-shared graph convolutional unit is trained. Then, the multi-factor fusion module is jointly optimized. Finally, end-to-end fine-tuning is performed, and gradient pruning and early stopping strategies are used to ensure training stability.

[0015] Preferably, in step S5, the end-to-end spatiotemporal graph network training also includes a visualization analysis module. The visualization analysis module is used to visualize the learned implicit graph structure, analyze the dynamic changes of spatial correlations in different time periods, verify the effectiveness of the weight sharing mechanism, and provide interpretability analysis of the prediction results.

[0016] The technical effects and advantages of this invention are as follows: 1. This invention addresses the problem of poor data quality after new energy access through multi-dimensional data preprocessing; it solves the problem of static graphs failing to capture dynamic implicit associations and spatiotemporal feature loss through the collaborative convolution of dynamic implicit graphs and weight-sharing graphs; and it solves the problems of redundant model parameters and insufficient feature fusion through weight sharing and multi-factor fusion, thereby achieving high-precision and efficient prediction of short-term power load and meeting user needs.

[0017] 2. In terms of data processing and spatial correlation, this invention improves data quality through dual anomaly detection and temporal interpolation. The dynamic implicit graph combined with regularization can accurately capture implicit correlations and adapt to load changes, making up for the shortcomings of traditional static graphs and enhancing the ability to extract spatial features.

[0018] 3. In terms of model optimization and feature fusion, this invention simplifies parameters through a weight-sharing mechanism, fully utilizes external features through multi-factor encoding and gating fusion, and improves prediction accuracy and training stability by combining end-to-end training and visualization design, thus meeting user needs. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of data preprocessing in this invention; Figure 3 This is a schematic diagram of the implicit graph structure of the present invention; Figure 4 This is a schematic diagram of adaptive weight-sharing graph convolution according to the present invention; Figure 5 This is a schematic diagram of the multi-factor enhancement fusion of the present invention; Figure 6 This is a flowchart of the end-to-end training process of the present invention; Figure 7 This is a flowchart of the KNN repair and smoothing process of the present invention; Figure 8 This is a diagram of the end-to-end spatiotemporal network architecture of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides, for example Figure 1 The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing, as shown, includes the following steps: S1. Multi-dimensional power load data preprocessing: collect multi-source data, then send the data to the data cleaning module, then use the 3σ criterion and isolated forest algorithm to detect outliers, then use KNN time series interpolation to complete the data, and finally divide the training set, validation set and test set in a 7:2:1 ratio; S2. Implicit graph structure and dynamic spatial association modeling: The preprocessed composite data is sent to the node embedding module to generate a trainable embedding vector for each node. Then, the association degree between nodes is calculated by cosine similarity to dynamically construct an adaptive graph structure. Next, L1 regularization is introduced to achieve graph structure sparsification and output the dynamic adjacency matrix to the downstream network. S3. Adaptive weight-sharing graph convolution design, dynamic adjacency matrix and node features, initial grouping according to node functional features, assign basic convolution kernels to each group of nodes, and then achieve personalized weight adjustment through fine-tuning matrix to complete spatial feature extraction and output to the temporal modeling unit; S4. Multi-factor enhancement module fusion: external features are normalized and encoded, and the weights of each feature are dynamically allocated through the attention weighting module. Then, the feature is gated and fused with the spatiotemporal features. Finally, the fused multi-factor enhancement node representation is sent to the prediction output layer. S5. End-to-end spatiotemporal graph network training with multi-factor enhanced features: First, the feature is input into the graph attention-gated recurrent network for spatiotemporal joint feature extraction. Then, it enters the prediction output layer to generate the load prediction value for the future time step. Next, the Huber loss function is used to evaluate the prediction error. Finally, it enters the parameter optimization module to update the network weights through gradient pruning and backpropagation.

[0022] Example 1: In this example, the core principle of this system is to achieve high-precision and efficient short-term power load forecasting through multi-stage collaborative linkage. The specific functions, problems solved, and implementation methods of each step are as follows: S1. Multi-dimensional power load data preprocessing: This addresses the issues of noise, outliers, and missing values ​​in the raw data after new energy access. The method involves collecting multi-source power load related data, cleaning the data to remove redundant information, using the 3σ criterion and the isolated forest algorithm to detect various outliers, using KNN time-series interpolation to complete missing data, and finally dividing the training, validation, and test sets in a 7:2:1 ratio to provide high-quality input for subsequent models. S2. Implicit Graph Structure and Dynamic Spatial Association Modeling: This addresses the problem that traditional explicit graphs cannot capture dynamic / implicit spatial associations and that spatiotemporal feature separation modeling is prone to information loss. The implementation method is to generate trainable node embedding vectors through a node embedding module, calculate node association degree by combining an attention mechanism to generate an implicit adjacency matrix, introduce L1 regularization to ensure the sparsity and interpretability of the graph structure, and adopt an attention-weighted dynamic update mechanism to adapt to changes in load associations, outputting a dynamic adjacency matrix to the downstream network. S3. Adaptive weight-sharing graph convolution design: This design addresses the issues of redundant parameters and low training and inference efficiency in traditional graph networks. It is implemented by dynamically grouping nodes based on their functions and load patterns, assigning a unified basic convolution kernel to similar node groups, and adjusting the node-level weights through a fine-tuning matrix. This simplifies parameters while ensuring the accuracy of spatial feature extraction, and works in synergy with the dynamic graph structure to optimize spatial association capture and parameter efficiency. S4. Multi-factor enhancement module fusion: This addresses the problem of insufficient fusion of multi-factor and spatiotemporal features and the inability to fully utilize external information. The implementation method is to process various external features through three major coding modules: meteorology, date attribute, and building features. After normalization and attention weighting, feature weights are allocated, and a gating mechanism is used to deeply fuse with spatiotemporal features to generate multi-factor enhancement node representations for prediction. S5. End-to-end Spatiotemporal Graph Network Training: This addresses the issues of unstable model training, insufficient prediction accuracy, and poor interpretability. The approach involves extracting spatiotemporal joint features through a graph attention-gated recurrent network, generating load prediction values ​​through the prediction output layer, evaluating errors using the Huber loss function, and ensuring training stability by combining gradient pruning and phased optimization strategies. A visualization analysis module is also added to improve the interpretability of the prediction results.

[0023] This invention provides, for example Figures 2 to 6 The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing, as shown, includes the following steps: S1. Multi-dimensional power load data preprocessing: collect multi-source data, then send the data to the data cleaning module, then use the 3σ criterion and isolated forest algorithm to detect outliers, then use KNN time series interpolation to complete the data, and finally divide the training set, validation set and test set in a 7:2:1 ratio; S2. Implicit graph structure and dynamic spatial association modeling: The preprocessed composite data is sent to the node embedding module to generate a trainable embedding vector for each node. Then, the association degree between nodes is calculated by cosine similarity to dynamically construct an adaptive graph structure. Next, L1 regularization is introduced to achieve graph structure sparsification and output the dynamic adjacency matrix to the downstream network. S3. Adaptive weight-sharing graph convolution design, dynamic adjacency matrix and node features, initial grouping according to node functional features, assign basic convolution kernels to each group of nodes, and then achieve personalized weight adjustment through fine-tuning matrix to complete spatial feature extraction and output to the temporal modeling unit; S4. Multi-factor enhancement module fusion: external features are normalized and encoded, and the weights of each feature are dynamically allocated through the attention weighting module. Then, the feature is gated and fused with the spatiotemporal features. Finally, the fused multi-factor enhancement node representation is sent to the prediction output layer. S5. End-to-end spatiotemporal graph network training with multi-factor enhanced features: First, the feature is input into the graph attention-gated recurrent network for spatiotemporal joint feature extraction. Then, it enters the prediction output layer to generate the load prediction value for the future time step. Next, the Huber loss function is used to evaluate the prediction error. Finally, it enters the parameter optimization module to update the network weights through gradient pruning and backpropagation.

[0024] The implicit graph structure and dynamic space association modeling in step S2 specifically includes: The node embedding module assigns a d-dimensional trainable embedding vector to each building load node, calculates the similarity between node embeddings through an attention mechanism, generates an implicit adjacency matrix, and initializes the nodes. Combined with L1 regularization constraints, it ensures the sparsity and interpretability of the graph structure.

[0025] The implicit adjacency matrix is ​​generated using a dynamic update mechanism based on attention weighting.

[0026] In step S3, the adaptive weight-sharing graph convolution design specifically includes: A weight-sharing pool is established to assign a unified convolution kernel to groups of nodes with similar functions. Node-level fine-tuning is achieved through adaptive adjustment of the matrix. Parameters are utilized during spatial feature aggregation to reduce the number of model parameters.

[0027] The weight-sharing mechanism employs a grouping strategy, dynamically grouping nodes based on their functional characteristics and load pattern similarity. Each group of nodes shares the basic convolutional kernel. And by adjusting the matrix Enables fine-tuning of node-level weights.

[0028] In step S4, the multi-factor enhancement module fusion specifically includes a meteorological feature coding module, a date attribute coding module, and a building feature coding module: The meteorological feature coding module is used to process temperature, humidity, and light intensity data; The date attribute encoding module is used to process holiday and seasonal information; The building feature encoding module is used to extract functional type features and achieves dynamic fusion of multi-source features through an attention weighting mechanism. The resulting multi-factor enhanced node representation is then sent to the prediction output layer for load prediction.

[0029] The multi-factor fusion adopts a gating mechanism, which is implemented as follows: Multi-source data is mapped to a unified dimension by feature normalization, attention weights are used to dynamically adjust the contribution of each feature, and a fully connected layer is used for feature compression. The resulting multi-factor enhanced node representation is transmitted to the prediction output layer to complete the final load prediction.

[0030] The dynamic graph structure in step S2 works in conjunction with the weight-sharing mechanism in step S3, specifically including: The dynamic adjacency matrix generated by the implicit graph learner is directly used as the input structure for the weight-shared graph convolution. The weight-shared pool automatically adjusts the grouping strategy according to the dynamic graph structure, achieving coordinated optimization of spatial association capture and parameter efficiency.

[0031] The end-to-end training in step S5 employs a phased optimization strategy, specifically including: First, the parameters of the implicit graph structure learner are fixed, and the weight-shared graph convolutional unit is trained. Then, the multi-factor fusion module is jointly optimized. Finally, end-to-end fine-tuning is performed, and gradient pruning and early stopping strategies are used to ensure training stability.

[0032] In step S5, the end-to-end spatiotemporal graph network training also includes a visualization analysis module. This module is used to visualize the learned implicit graph structure, analyze the dynamic changes in spatial relationships at different time periods, verify the effectiveness of the weight sharing mechanism, and provide interpretability analysis of the prediction results.

[0033] Example 2: In this example, in step 1, the collected multi-source data includes real-time electricity consumption data of different types of buildings, local meteorological data, date-related information, and information about the buildings themselves. These collected data are aggregated, and useless, duplicate, or blank data are removed. Then, abnormal or irregular values ​​are identified. Two methods are used in combination: one to quickly identify obvious anomalies, and the other to identify hidden or less obvious anomalies, ensuring that all anomalies are identified. These two methods are the 3σ criterion and the Isolation Forest algorithm, respectively. Afterward, for missing or incomplete data, KNN time-series interpolation is used to fill in the gaps, ensuring data continuity, by referring to the normal data before and after it. Finally, the processed data is divided into three parts: one part for training the model, one part for adjusting model parameters, and one part for testing the model's prediction performance, namely the training set, validation set, and test set. In step S2, all the data processed in the first step is input into the node embedding module. An adjustable feature vector is assigned to each building's electricity consumption node, representing its electricity consumption characteristics. Cosine similarity is used to analyze the similarity between the feature vectors of each node, determining the degree of correlation between electricity consumption nodes in different buildings, and generating a matrix that reflects this correlation. This matrix is ​​essentially a relationship graph, reflecting which nodes have similar electricity consumption patterns and high correlation. Simultaneously, L1 regularization simplifies this relationship graph, removing weak and useless relationships, making the graph structure clearer and easier to understand. Furthermore, this graph is not static; it is adjusted in real-time according to data changes to ensure it accurately reflects the electricity consumption correlation between each node. Finally, the adjusted relationship graph is passed to the next processing stage. Receive the relationship graph and feature data of each node generated in the previous step. First, based on the building's purpose and electricity consumption patterns, nodes with similar electricity consumption characteristics are divided into several groups. A unified basic convolution kernel is assigned to each group of similar nodes. This basic convolution kernel can extract the common electricity consumption features of the nodes in the group. At the same time, some minor adjustments are made for the individual electricity consumption characteristics of each node to ensure that the commonalities of the nodes within the group are taken into account while adapting to the individuality of each node. Through this calculation method, the spatial correlation features between each node are extracted. Finally, the extracted features are passed to the next processing stage. It should be noted that the basic convolution kernel allocation in this step and the relationship graph in the previous step are mutually supportive. When the relationship graph changes, the grouping of nodes will also be adjusted accordingly, and the calculation template will also be adapted accordingly to ensure that the correlation between nodes can be accurately captured. In step S4, this step mainly integrates the impact of various external factors on the electricity load. It is divided into three parts, which process different types of external data respectively, and finally merge them together: (1) Meteorological feature processing: Adjust meteorological data such as temperature, humidity, and light intensity to a uniform range to eliminate the influence of different data units, and then extract the impact characteristics of meteorological data on electricity load; (2) Date attribute processing: Convert information such as whether it is a holiday or which season it belongs to into a form that the model can recognize, and then extract the impact characteristics of date information on electricity load; (3) Building feature processing: The building's purpose, size, electrical equipment, and other information are converted into a form that the model can recognize. After adjusting to a unified range, the influence of the building's own characteristics on the electricity load is extracted. The features extracted from the above three parts are integrated, and different weights are assigned according to the magnitude of the influence of different features on the electricity load, highlighting the features with the greatest impact. Then, these external features are combined with the spatial correlation features extracted in the previous step to form a comprehensive feature. This comprehensive feature can fully reflect the influence of various factors on the electricity load and is finally passed to the prediction stage. In step S5, the comprehensive features obtained in the previous step are input into the graph attention-gated recurrent network. This model can capture both the spatial relationships between nodes and the time-varying patterns of electricity load, achieving comprehensive extraction of spatial and temporal features. After feature extraction, the predicted output layer generates predicted electricity load values ​​for a future period. Then, the predicted values ​​are compared with the actual electricity load values ​​to determine the accuracy of the prediction. The Huber loss function is used to evaluate the prediction error, ensuring that the prediction results are closer to the actual situation. Simultaneously, the parameter optimization module adjusts various parameters of the model to avoid excessive prediction deviations and parameter runaway, ensuring stable model operation. The training process is divided into three stages: In the first stage, the relevant parameters of the relationship graph are fixed, and only the parameters of the calculation template are optimized; in the second stage, the parameters of the relationship graph are unfixed, and the relevant parameters of the calculation template and external feature fusion are optimized; in the third stage, all parameters of the entire model are fine-tuned to ensure the model's prediction effect reaches its best, while avoiding overtraining and inaccurate predictions, meeting user needs.

[0034] This invention provides, for example Figures 7 to 8 The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing, as shown, includes the following steps: S1. Multi-dimensional power load data preprocessing: collect multi-source data, then send the data to the data cleaning module, then use the 3σ criterion and isolated forest algorithm to detect outliers, then use KNN time series interpolation to complete the data, and finally divide the training set, validation set and test set in a 7:2:1 ratio; S2. Implicit graph structure and dynamic spatial association modeling: The preprocessed composite data is sent to the node embedding module to generate a trainable embedding vector for each node. Then, the association degree between nodes is calculated by cosine similarity to dynamically construct an adaptive graph structure. Next, L1 regularization is introduced to achieve graph structure sparsification and output the dynamic adjacency matrix to the downstream network. S3. Adaptive weight-sharing graph convolution design, dynamic adjacency matrix and node features, initial grouping according to node functional features, assign basic convolution kernels to each group of nodes, and then achieve personalized weight adjustment through fine-tuning matrix to complete spatial feature extraction and output to the temporal modeling unit; S4. Multi-factor enhancement module fusion: external features are normalized and encoded, and the weights of each feature are dynamically allocated through the attention weighting module. Then, the feature is gated and fused with the spatiotemporal features. Finally, the fused multi-factor enhancement node representation is sent to the prediction output layer. S5. End-to-end spatiotemporal graph network training with multi-factor enhanced features: First, the feature is input into the graph attention-gated recurrent network for spatiotemporal joint feature extraction. Then, it enters the prediction output layer to generate the load prediction value for the future time step. Next, the Huber loss function is used to evaluate the prediction error. Finally, it enters the parameter optimization module to update the network weights through gradient pruning and backpropagation.

[0035] Example 3: In this example, the implicit adjacency matrix is ​​generated using an attention-weighted dynamic update mechanism, specifically expressed as follows:

[0036] in A learnable attention vector; Embed vectors for nodes. This indicates a vector concatenation operation.

[0037] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing, characterized in that, Includes the following steps: S1. Multi-dimensional power load data preprocessing: collect multi-source data, then send the data to the data cleaning module, then use the 3σ criterion and isolated forest algorithm to detect outliers, then use KNN time series interpolation to complete the data, and finally divide the training set, validation set and test set in a 7:2:1 ratio; S2. Implicit graph structure and dynamic spatial association modeling: The preprocessed composite data is sent to the node embedding module to generate a trainable embedding vector for each node. Then, the association degree between nodes is calculated by cosine similarity to dynamically construct an adaptive graph structure. Next, L1 regularization is introduced to achieve graph structure sparsification and output the dynamic adjacency matrix to the downstream network. S3. Adaptive weight-sharing graph convolution design, dynamic adjacency matrix and node features, initial grouping according to node functional features, assign basic convolution kernels to each group of nodes, and then achieve personalized weight adjustment through fine-tuning matrix to complete spatial feature extraction and output to the temporal modeling unit; S4. Multi-factor enhancement module fusion: external features are normalized and encoded, and the weights of each feature are dynamically allocated through the attention weighting module. Then, the feature is gated and fused with the spatiotemporal features. Finally, the fused multi-factor enhancement node representation is sent to the prediction output layer. S5. End-to-end spatiotemporal graph network training with multi-factor enhanced features: First, the feature is input into the graph attention-gated recurrent network for spatiotemporal joint feature extraction. Then, it enters the prediction output layer to generate the load prediction value for the future time step. Next, the Huber loss function is used to evaluate the prediction error. Finally, it enters the parameter optimization module to update the network weights through gradient pruning and backpropagation.

2. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 1, characterized in that, The implicit graph structure and dynamic space association modeling in step S2 specifically includes: The node embedding module assigns a d-dimensional trainable embedding vector to each building load node, calculates the similarity between node embeddings through an attention mechanism, generates an implicit adjacency matrix, and initializes the nodes. Combined with L1 regularization constraints, it ensures the sparsity and interpretability of the graph structure.

3. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 2, characterized in that, The implicit adjacency matrix is ​​generated using a dynamic update mechanism based on attention weighting.

4. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 1, characterized in that, In step S3, the adaptive weight-sharing graph convolution design specifically includes: A weight-sharing pool is established to assign a unified convolution kernel to groups of nodes with similar functions. Node-level fine-tuning is achieved through adaptive adjustment of the matrix. Parameters are utilized during spatial feature aggregation to reduce the number of model parameters.

5. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 4, characterized in that, The weight-sharing mechanism employs a grouping strategy, dynamically grouping nodes based on their functional characteristics and load pattern similarity, with each group sharing the basic convolutional kernel. And by adjusting the matrix Enables fine-tuning of node-level weights.

6. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 1, characterized in that, In step S4, the multi-factor enhancement module fusion specifically includes a meteorological feature coding module, a date attribute coding module, and a building feature coding module: The meteorological feature coding module is used to process temperature, humidity, and light intensity data; The date attribute encoding module is used to process holiday and seasonal information; The building feature encoding module is used to extract functional type features and achieves dynamic fusion of multi-source features through an attention weighting mechanism. The resulting multi-factor enhanced node representation is then sent to the prediction output layer for load prediction.

7. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 6, characterized in that, The multi-factor fusion adopts a gating mechanism, specifically implemented as follows: Multi-source data is mapped to a unified dimension by feature normalization, attention weights are used to dynamically adjust the contribution of each feature, and a fully connected layer is used for feature compression. The resulting multi-factor enhanced node representation is transmitted to the prediction output layer to complete the final load prediction.

8. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 1, characterized in that, The dynamic graph structure in step S2 works in conjunction with the weight-sharing mechanism in step S3, specifically including: The dynamic adjacency matrix generated by the implicit graph learner is directly used as the input structure for the weight-shared graph convolution. The weight-shared pool automatically adjusts the grouping strategy according to the dynamic graph structure, achieving coordinated optimization of spatial association capture and parameter efficiency.

9. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 1, characterized in that, The end-to-end training in step S5 employs a phased optimization strategy, specifically including: First, the parameters of the implicit graph structure learner are fixed, and the weight-shared graph convolutional unit is trained. Then, the multi-factor fusion module is jointly optimized. Finally, end-to-end fine-tuning is performed, and gradient pruning and early stopping strategies are used to ensure training stability.

10. The multi-factor short-term power load forecasting system based on graph structure adaptive weight sharing according to claim 1, characterized in that, In step S5, the end-to-end spatiotemporal graph network training also includes a visualization analysis module. This module is used to visualize the learned implicit graph structure, analyze the dynamic changes in spatial relationships at different time periods, verify the effectiveness of the weight sharing mechanism, and provide interpretability analysis of the prediction results.

Images