Short-term power load prediction method based on graph structure learning and load period mode

Abstract

The invention belongs to the field of short-term power load prediction, and discloses a short-term power load prediction method based on graph structure learning and a load period mode, which comprises the following steps: acquiring historical load data of multi-scale time fine grit as input data; processing the input data by adopting a similar graph learning method to obtain a similar graph, and processing the input data by adopting a dynamic non-stationary graph learning method to obtain a dynamic non-stationary graph; extracting spatial features from the similarity graph and the dynamic non-stationary graph, and extracting time features from the input data; decoupling the spatial features and the time features based on a decoupling mechanism to obtain spatial-temporal features; and after fusing the input data and the spatial-temporal characteristics, predicting and outputting future short-term power load data by using a multi-layer perceptron. The method effectively improves the accuracy of short-term power load prediction.

Description

Technical Field

[0001] The present invention belongs to the field of short-term power load forecasting, and in particular relates to a short-term power load forecasting method based on graph structure learning and load cycle patterns. Background Art

[0002] Accurate short-term power load forecasting (STPLF) is a prerequisite for the effective implementation of demand-side response and therefore plays a vital role in demand-side response. However, unlike aggregate load forecasting, achieving accurate short-term power load forecasting is very difficult due to the complexity and uncertainty of load data in STPLF.

[0003] In recent years, many solutions for short-term power load forecasting have been proposed. According to the different types of data structures used, these methods can be divided into three categories: load data set (LDS)-based methods, that is, directly inputting the load data set into the model for load forecasting; graph structure (GS)-based methods, that is, converting the load data set into a graph structure and then inputting it into the model for load forecasting; and methods that integrate load data set and graph structure (LDS-GS), that is, inputting the graph structure and load data set into the model together for load forecasting.

[0004] Currently, various methods based on load datasets have been effectively applied to short-term power load forecasting, such as long short-term memory networks (LSTMs), gated recurrent units (GRUs), and transformers. At the same time, breakthroughs have been made in graph-based methods, with experiments demonstrating that effective spatial features can help improve the accuracy of short-term power load forecasting. In recent years, methods based on load datasets and graph structures have attracted considerable research attention due to their ability to simultaneously consider both spatial and temporal features between nodes (such as households and buildings). Despite these achievements, significant challenges remain in improving the reliability and accuracy of short-term power load forecasting.

[0005] First, the quality of the constructed graph structure is poor. Existing methods based on load datasets and graph structures use a variety of algorithms to construct the graph structure, including dynamic time warping (DTW), Pearson correlation coefficient (PCC), and k-means algorithm (KMA). However, these algorithms are overly sensitive to outliers and suffer from feature degradation.

[0006] Second, existing methods primarily employ a single-scale, fine-grained input strategy, often ignoring the inherent cyclical patterns of load data. This limits the model's ability to extract spatial and temporal features from the load data. For example, when predicting the load at 5:00 PM on May 31, 2025, existing methods typically only use limited historical load data prior to that time point. However, the load pattern at 5:00 PM on May 31, 2025, may also be similar to the load patterns around 5:00 PM on May 30, 2025, and around 5:00 PM on May 24, 2025. These recurring load patterns harbor important spatial and temporal characteristics.

[0007] Third, for most methods based on load data and graph structures, spatial and temporal features are typically extracted by the spatial and temporal modules of the forecasting model. However, due to the complexity of the load forecasting task, spatial and temporal features are often coupled, exhibiting complex spatiotemporal correlations. For example, during peak hours, a large number of nodes in a specific area consume electricity simultaneously, generating load data that contains spatial and temporal characteristics between nodes. Summary of the Invention

[0008] The purpose of the present invention is to provide a short-term power load forecasting method based on graph structure learning and load cycle mode to improve the accuracy of short-term power load forecasting.

[0009] To achieve the above object, the technical solution adopted by the present invention is:

[0010] A short-term power load forecasting method based on graph structure learning and load cycle patterns includes:

[0011] Obtain multi-scale time-granular historical load data as input data;

[0012] The similarity graph learning method is used to process the input data to obtain a similarity graph, and the dynamic non-stationary graph learning method is used to process the input data to obtain a dynamic non-stationary graph;

[0013] Extract spatial features from similarity graphs and dynamic non-stationary graphs, and extract temporal features from input data;

[0014] Based on the decoupling mechanism, spatial features and temporal features are decoupled to obtain spatiotemporal features;

[0015] After fusing the input data and spatiotemporal features, the multi-layer perceptron is used to predict and output future short-term power load data.

[0016] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution. They are merely further supplements or optimizations. Under the premise that there are no technical or logical contradictions, each optional method can be combined separately for the above-mentioned overall solution, or multiple optional methods can be combined.

[0017] Preferably, the obtaining of multi-scale time-granular historical load data includes:

[0018] Get recent load data: Recent load data is ,in is the current time point, is the length of recent load data, is the number of nodes, For time point The load data, For time point Load data;

[0019] Get daily cycle load data: Daily cycle load data is ,in is the length of the time window, is the daily historical load data, and , is the sampling rate of load data, For time point The load data, For time point The load data, For time point Before The load data of the sampling points, For time point after Load data of sampling points;

[0020] Get weekly cycle load data: Weekly cycle load data is , is the weekly historical load data, and , For time point The load data, For time point The load data, For time point Before The load data of the sampling points, For time point after Load data of sampling points;

[0021] The recent load data, daily cycle load data and weekly cycle load data are combined to obtain multi-scale time-granular historical load data.

[0022] Preferably, the step of processing the input data using a similarity graph learning method to obtain a similarity graph includes:

[0023] Use multi-scale dilated convolution to extract multi-scale features from input data;

[0024] Splicing multi-scale features and mapping them into feature matrix ;

[0025] Using similarity algorithms to process feature matrices , and obtain the similarity graph.

[0026] Preferably, the step of processing input data using a dynamic non-stationary graph learning method to obtain a dynamic non-stationary graph comprises:

[0027] Use gating mechanism to control recent load data and the feature matrix Perform adaptive feature fusion to obtain the adaptive fusion feature matrix , and adaptively fused feature matrix based on Gumbel Softmax sampling strategy Construct dynamic non-stationary graphs.

[0028] Preferably, the extracting of spatial features from the similarity graph and the dynamic non-stationary graph includes: taking the similarity graph and the dynamic non-stationary graph as inputs of a graph attention convolution, and extracting and outputting the spatial features by the graph attention convolution.

[0029] Preferably, the extracting time features from the input data includes: using a stacked LSTM network to extract time features from the input data.

[0030] Preferably, the decoupling mechanism is used to decouple spatial features from temporal features to obtain spatiotemporal features, including:

[0031]

[0032]

[0033] in, represents the weight matrix, is the activation function, is the spatial feature matrix, is the time feature matrix, is the learnable weight matrix corresponding to the spatial feature matrix, is the learnable weight matrix corresponding to the time feature matrix, is a learnable bias term, is the spatiotemporal feature matrix, is the Hadamard product.

[0034] Preferably, after fusing the input data and spatiotemporal features, the method of using a multi-layer perceptron to predict and output future short-term power load data includes:

[0035]

[0036]

[0037]

[0038]

[0039] in, For input data, is the fully connected layer, is the transformation feature matrix, is the spatiotemporal feature matrix, represents the concatenation operation, is the fusion feature matrix, is a multi-layer perceptron, is the output of the multi-layer perceptron, For time point Short-term power load data forecast value.

[0040] Preferably, the method further includes a training phase, wherein a graph regularization function and a mean absolute error are calculated in the training phase, and a weighted sum of the graph regularization function and the mean absolute error is used as a total loss function;

[0041] The graph regularization function is calculated based on the input data and the similarity graph, and the formula is as follows:

[0042]

[0043] in, is the graph regularization function, is the number of nodes, The first The load data of each node, The first The load data of each node, represents the square of the Euclidean norm, represents the square of the Frobenius norm, is a hyperparameter, is the adjacency matrix of the similarity graph, Represents the relationship between the i-th node and the j-th node in the adjacency matrix of the similarity graph;

[0044] The mean absolute error is calculated based on the predicted data and the true data, and the formula is as follows:

[0045]

[0046] in, is the mean absolute error, For time point The short-term power load data forecast value of the i-th node, For time point The true value of the short-term power load data of the i-th node.

[0047] The present invention provides a short-term power load forecasting method based on graph structure learning and load cycle patterns. First, a new graph structure learning method is proposed, which uses a similarity algorithm and a Gumbel Softmax sampling strategy to address the problem of poor graph structure quality. Second, a new multi-scale time fine-grained input strategy is proposed to capture the inherent periodic patterns in load data. Third, a new decoupling mechanism is proposed to enhance the ability to separate spatial and temporal features. Multiple experimental results show that the present method outperforms existing mainstream forecasting methods in terms of robustness and real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flowchart of a short-term power load forecasting method based on graph structure learning and load cycle model of the present invention;

[0049] Figure 2 A data flow diagram of a short-term power load forecasting method based on graph structure learning and load cycle patterns according to the present invention;

[0050] Figure 3 Flowchart of the decoupling mechanism of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0053] The task of short-term power load forecasting can be viewed as a multi-source forecasting problem, where each node represents a data source. is defined as a tensor. The goal of short-term power load forecasting is to Learn a function f that is used to predict the next time point The load data can be described as formula (1).

[0054] (1)

[0055] To address the aforementioned multi-source forecasting problem, this example proposes a novel multi-scale graph tidal model (MSGT) consisting of an input strategy, a GSL module, a spatiotemporal module, a decoupling mechanism, and an output module to improve the accuracy of short-term power load forecasting. First, a novel graph structure learning method is proposed, utilizing a similarity algorithm and a Gumbel Softmax sampling strategy to address the issue of poor graph structure quality. Second, a novel multi-scale temporal fine-grained input strategy is proposed to help MSGT capture the inherent cyclical patterns in load data. Third, a novel decoupling mechanism is proposed to enhance MSGT's ability to separate spatial and temporal features. Multiple experimental results demonstrate that MSGT outperforms existing mainstream forecasting models in terms of robustness and real-time performance.

[0056] like Figure 1 and Figure 2 As shown, this embodiment proposes a short-term power load forecasting method based on graph structure learning and load cycle pattern, including the following steps:

[0057] Step 1: Input strategy: Obtain multi-scale time-granular historical load data as input data.

[0058] The load data at time point t is recorded as , where N is the number of nodes (such as homes, buildings, etc.) in the power system. Load data exhibits a periodic pattern with daily and weekly cycles, and recent load data also contains future load trends. Therefore, the multi-scale time fine-grained input strategy includes three different time fine-grains: 1) recent load data ;2) Daily cycle load data ;3) Weekly load data ,This strategy makes it easier for MSGT to capture the inherent periodic pattern of load data. To solve the periodic time offset problem, the time window Ensure that the input data includes daily and weekly historical load data. Detailed information is as follows.

[0059] (1) Recent load data: Recent load data is ,in is the current time point, is the length of recent load data, is the number of nodes, For time point The load data, For time point load data.

[0060] (2) Daily cycle load data: Daily cycle load data is ,in Time window length, is the daily historical load data, and , is the sampling rate of load data, that is, load data is sampled S times a day, For time point The load data, For time point The load data, For time point Before The load data of the sampling points, For time point after The load data of each sampling point.

[0061] (3) Weekly load data: Weekly load data is , is the weekly historical load data, and , For time point The load data, For time point The load data, For time point Before The load data of the sampling points, For time point after The load data of each sampling point.

[0062] In summary, the combination of recent load data, daily cycle load data and weekly cycle load data gives the multi-scale time-granular historical load data: , where L is equal to .

[0063] Step 2, GSL module: Use similarity graph learning method to process input data to obtain similarity graph, and use dynamic non-stationary graph learning method to process input data to obtain dynamic non-stationary graph.

[0064] In this embodiment, two directed graphs and are defined as similarity graph and dynamic non-stationary graph respectively. V represents the node set, and represents the edge set, and Represents the adjacency matrix. is a mathematical expression of graph connectivity, if but ,if but ,in and Represent nodes i and j respectively, Representing the similarity graph and Similarly, The relationships between nodes in dynamic non-stationary graphs are represented in a similar way.

[0065] Spatial features of different nodes are crucial for accurate short-term power load forecasting. However, existing models for establishing spatial features are overly sensitive to outliers and suffer from feature degradation. Therefore, this paper proposes a new GSL module (graph structure learning) to alleviate this problem.

[0066] (1) Similarity Graph Learning (SGL): To address the problem that existing methods are too sensitive to outliers, this embodiment proposes a similarity graph learning method. First, multi-scale dilated convolution is used to extract the fine-grained input from the multi-scale temporal The multi-scale features extracted are used Represented by , where d is the feature dimension of the hidden layer. Multi-scale dilated convolution can capture rich contextual information, multi-scale features and long dependencies in the load data. Then, Mapped into feature matrix Finally, the feature matrix is ​​processed using the similarity algorithm To obtain the similarity graph. The above process can be described as formulas (2) to (7).

[0067] (2)

[0068] (3)

[0069] (4)

[0070] (5)

[0071] (6)

[0072] (7)

[0073] in, and Represent convolution operation and connection operation respectively; 、 and Represent convolution kernels with scales of 1, 2, and 5, and dilation rates of 1, 2, and 5, respectively; 、 、 、 and Both represent bias terms; 、 and Represents features of different scales, which contain more comprehensive key features of load data; represents a fully connected layer; Unnormalized adjacency matrix representing the similarity graph; express The relationship between the i-th node and the j-th node in the , is calculated based on the similarity algorithm. The larger the value, the stronger the similarity relationship, and the smaller the value, the weaker the similarity relationship. express The feature vector of the i-th node in; represents the transposed eigenvector of the jth node; represents the exponential function; and represents the activation function, the former ensures , which ensures that the values ​​are positive.

[0074] (2) Dynamic Non-Stationary Graph Learning Method (DNSGL): In order to solve the problem of feature degradation in existing methods, this embodiment proposes a dynamic non-stationary graph learning method. This method first uses a gating mechanism to learn the recent load data. and the feature matrix Adaptively perform feature fusion to obtain adaptive fusion features ; Then, a dynamic non-stationary graph is constructed through the Gumbel Softmax sampling strategy.

[0075] The gating mechanism is solved by feature fusion operation and The problem of inconsistent feature scales and weights between them can be described as formulas (8) to (12).

[0076] (8)

[0077] (9)

[0078] (10)

[0079] (11)

[0080] (12)

[0081] in, Depend on Transformed; and These are all learnable weight matrices in the gating mechanism; and Represents sigmoid and tanh activation functions respectively; is the Hadamard product; 、 and Represents features with different weights that maximize coverage and The feature space of is the adaptive fusion feature matrix.

[0082] Adopting Gumbel Softmax strategy to adaptively fuse features The process can be described as formula (13) to (15).

[0083] (13)

[0084] (14)

[0085] (15)

[0086] in, Represents the relationship between the i-th node and the j-th node in the adjacency matrix of a dynamic non-stationary graph. The larger the value, the stronger the fluctuation relationship, and the smaller the value, the weaker the fluctuation relationship. express The eigenvector of the i-th node in , Represents the output of the fully connected layer and The similarity of represents a Gumbel distribution with location parameter 0 and scale parameter 1; and are independent and identically distributed sample vectors, all randomly drawn from the Gumbel distribution; S represents the parameter that controls the smoothness of the sample, represents the logarithmic function, is the activation function.

[0087] Step 3, spatiotemporal module: extract spatial features from the similarity graph and dynamic non-stationary graph, and extract temporal features from the input data.

[0088] First, we use graph attention convolution. Graph attention convolution is a convolutional layer in the GNN (Graph Neural Network) model that embodies the core concept of the Graph Attention Network (GAT). Based on the self-attention mechanism, graph attention convolution can adaptively assign appropriate weights to each neighboring node to aggregate neighboring node features. The similarity graph and the dynamic non-stationary graph are used as input for the graph attention convolution, which extracts and outputs spatial features.

[0089] Secondly, in order to In this example, stacked LSTM is used to extract time features. By increasing the network depth, stacking LSTM not only strengthens the model's expressive power but also improves its ability to capture complex temporal features.

[0090] Step 4: Decoupling mechanism: Decouple spatial features and temporal features based on the decoupling mechanism to obtain spatiotemporal features.

[0091] In order to decouple spatial features from temporal features, this embodiment proposes a new decoupling mechanism, such as Figure 3 The mechanism can be described as formula (16)~(17).

[0092] (16)

[0093] (17)

[0094] in, represents the weight matrix, is the activation function, is the spatial feature matrix, is the time feature matrix, is the learnable weight matrix corresponding to the spatial feature matrix, is the learnable weight matrix corresponding to the time feature matrix, is a learnable bias term, is the spatiotemporal feature matrix, is the Hadamard product.

[0095] At the same time, the information extracted from the similarity graph and the dynamic non-stationary graph also faces the problem of coupling between spatial and temporal features. Therefore, a decoupling mechanism is also used between them to alleviate the impact of this problem. That is, the information extracted from the similarity graph is regarded as spatial features, and the information extracted from the dynamic non-stationary graph is regarded as temporal features. Similarly, formulas (16) to (17) are used to finally decouple and obtain spatial features.

[0096] Step 5, output module: After fusing the input data and spatiotemporal features, the multi-layer perceptron is used to predict and output future short-term power load data.

[0097] After extracting important spatial and temporal features, the output module is used to effectively integrate the spatial and temporal features to improve prediction accuracy. The output module can be described as formulas (18) to (21).

[0098] (18)

[0099] (19)

[0100] (20)

[0101] (twenty one)

[0102] in, For input data, is the fully connected layer, is the transformation feature matrix, is the spatiotemporal feature matrix, represents the concatenation operation, is the fusion feature matrix, is a multi-layer perceptron (MLP), is the output of the multi-layer perceptron, For time point The specific operation of the multilayer perceptron can be described as formula (22).

[0103] (twenty two)

[0104] in, represents the ReLU activation function, Represents the first layer hidden layer, Represents the first The hidden layer processes the output features, Represents the first The features of the hidden layer output, is the total number of layers of the multilayer perceptron, , and the input of the first hidden layer is .

[0105] In the inference application phase, MSGT outputs the predicted value That is, the future short-term power load data. At the training node, after MSGT outputs the predicted value, the total loss function needs to be calculated. , so that the MSGT parameters can be iteratively updated through the gradient descent algorithm.

[0106] The reasonable connectivity and sparsity of similarity graphs play a vital role in improving the performance of GNN models. To this end, this embodiment uses a graph regularization function to control the connectivity and sparsity of similarity graphs, thereby optimizing the performance of MSGT. Defined as formula (23).

[0107] (twenty three)

[0108] in, is the graph regularization function, is the number of nodes, The first The load data of each node, The first The load data of each node, represents the square of the Euclidean norm, represents the square of the Frobenius norm, is a hyperparameter that controls the sparsity of the similarity graph. is the adjacency matrix of the similarity graph, Represents the relationship between the i-th node and the j-th node in the adjacency matrix of the similarity graph. The bigger, The smaller; Used to control how the graph learning process captures long spatial features between nodes.

[0109] Secondly, the prediction loss is calculated using the mean absolute error (MAE) , which is then combined with the graph regularization function Combined to calculate the total loss. This process can be described as formula (24) ~ (25).

[0110] (twenty four)

[0111] (25)

[0112] in, is the mean absolute error, For time point The short-term power load data forecast value of the i-th node, For time point The true value of the short-term power load data of the i-th node, is a hyperparameter used to adjust the weight of the graph regularization loss function.

[0113] In order to intuitively demonstrate the advantages of the multi-scale graph tidal model proposed in the present invention, the present invention conducts the following experiments:

[0114] (1) Datasets: This experiment uses two real load datasets, GEFCom2012 and GEFCom2017, to verify the effectiveness of the model. The GEFCom2012 dataset contains 39,576 load records from 20 nodes, spanning from January 1, 2004 to June 30, 2008. The GEFCom2017 dataset contains 61,344 load records from 500 nodes, spanning from January 1, 2005 to December 31, 2011. Each dataset is divided into three parts: 70% of the data is used for training, 10% for validation, and the remaining 20% ​​for testing.

[0115] (2) Baseline and benchmark:

[0116] This experiment compares MSGT with the following baseline models to verify the effectiveness of MSGT in short-term load forecasting.

[0117] A. HLSA (from Zang, HX, Xu, RQ, Cheng, LL, Ding, T., Liu,L., Wei, ZN, & Sun, GQ (2021). Residential load forecasting based on LSTM fusing self-attention mechanism with pooling. Energy, 229, 120682.): Load forecasting is performed through feature engineering methods, pooling strategies, and a combined deep learning framework.

[0118] B. GWN-GCN (from Lin, WX, Wu, D., & Boulet, B. (2021). Spatial-temporal residential short-term load forecasting via graph neural networks. IEEE Transactions on Smart Grid, 12(6), 5373-5384.): It integrates graph convolutional layers and temporal convolutional networks and can use graph structures learned from load data for load forecasting.

[0119] C. GCRNN (from Arastehfar, S., Matinkia, M., & Jabbarpour, MR (2022). Short-term residential load forecasting using graph convolutional recurrent neural networks. Engineering Applications of Artificial Intelligence, 116, 105358.): Combines graph convolutional recurrent neural networks with LSTM for load forecasting.

[0120] D. RLF-MGNN (from Wang, YF, Rui, LX, Ma, JH, & Jin Q. (2023). A short-term residential load forecasting scheme based on the multiple correlation-temporal graph neural networks. Applied Soft Computing, 146, 110629.): It uses a multiple correlation spatiotemporal graph convolutional network to extract latent features from load data, and then transfers the features to LSTM to enhance the spatiotemporal feature learning ability for load forecasting.

[0121] E. ATF (from Wu, D., & Lin, WX (2022). Efficient residential electric load forecasting via transfer learning and graph neural networks. IEEE Transactions on Smart Grid, 14(3), 2423-2431.): An attention transfer framework that can integrate GNN models trained in different domains, especially models trained in external domains with models trained in the load forecasting domain, to perform load forecasting.

[0122] The above baseline models use the hyperparameters recommended by the original literature. To evaluate the effectiveness of MSGT and the baseline models, this experiment uses three robust and indicative benchmarks: mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE).

[0123] (3) Experimental setup:

[0124] (3-1) Data preprocessing: In order to better extract the spatial and temporal characteristics of load data, the data were normalized using the Z-score.

[0125] (3-2) Experimental environment: This experiment uses PyTorch to implement MSGT and baseline models and runs on a computer equipped with an NVIDIA GeForce RTX 3090 GPU and an AMD Ryzen 7 3700X 8 CPU @ 3.6 GHz.

[0126] (3-3) Model hyperparameter setting: length of recent historical load data Set to 64, the time window length Set to 4, the feature dimension of the hidden layer is set to 32, and the coefficient in the graph regularization function The number of heads K of the graph attention convolution is set to 2, the dropout rate is set to 0.2, and the number of layers is set to 3; the dropout rate of the stacked LSTM is set to 0.1, and the number of layers is set to 3; the multilayer perceptron consists of 3 hidden layers; the total loss function The probability distribution is set to 0.3, the learning rate is set to 0.002, the batch size is set to 64, and the number of training rounds is set to 150. The loss function is MAE, and the optimizer is Adam. To ensure the fairness of the experiment, the parameters of all baseline models are configured according to the parameter description of the original paper.

[0127] (4) Performance comparison:

[0128] This section uses a real workload dataset to compare the performance of MSGT and the baseline model with different node numbers. Tables 1 and 2 show the average results of five repeated experiments for each model.

[0129] Table 1 Performance of each model on the GEFCom2012 dataset

[0130]

[0131] Table 2 Performance of each model on the GEFCom2017 dataset

[0132]

[0133] By analyzing the data in Table 1 and Table 2, the following conclusions are drawn:

[0134] A. All models perform better on GEFCom2017 than on GEFCom2012, indicating that GEFCom2012 has greater volatility and complexity than GEFCom2017.

[0135] B. As the number of nodes increases, all models experience performance fluctuations. However, MSGT has the least fluctuation and superior overall performance, indicating that MSGT has good robustness.

[0136] C. The overall performance of MSGT on the two datasets is better than other mainstream load forecasting models, which verifies the effectiveness of MSGT.

[0137] D. In the GEFCom2012 dataset, the performance of GCRNN is generally lower than that of other models, indicating that GCRNN has difficulties in effectively extracting and quantifying the spatial and temporal features of complex datasets, which leads to its poor performance.

[0138] E. In the GEFCom2017 dataset, the performance of GWN-GCN is generally lower than that of other models, indicating that GWN-GCN over-relies on complex structures and algorithms, resulting in poor performance on simple datasets.

[0139] (5) Ablation Experiments: This section compares MSGT and its various variants and evaluates their contributions to the overall performance. The variants compared are as follows:

[0140] Variant 1: MSGT variant without similarity graph learning.

[0141] Variant 2: MSGT variant with dynamic non-stationary graph learning removed.

[0142] Variant 3: MSGT variant without the decoupling mechanism.

[0143] Variant 4: MSGT variant with the output module removed.

[0144] Variant 5: MSGT variant without the graph regularization function.

[0145] Next, we evaluate MSGT and its variants on the aforementioned payload dataset. Tables 3 and 4 show the average results of five replicates for each variant.

[0146] Table 3 Performance of each variant on the GEFCom2012 dataset

[0147]

[0148] Table 4 Performance of each variant on the GEFCom2012 dataset

[0149]

[0150] By analyzing the data in Table 3 and Table 4, the following conclusions are drawn:

[0151] A. On both datasets, Variant 1 performs worse than MSGT, indicating that similarity graph learning can effectively alleviate the defect that the model is too sensitive to outliers.

[0152] B. On both datasets, the performance of Variant 2 is worse than MSGT, indicating that dynamic non-stationary graph learning can effectively alleviate the problem of model feature degradation.

[0153] C. On both datasets, the performance of variant 3 is worse than MSGT, indicating that the decoupling mechanism effectively alleviates the coupling of spatial and temporal features by assigning weights.

[0154] D. On both datasets, Variant 4 performs worse than MSGT, indicating that the output module can effectively integrate spatial and temporal features, thereby improving the prediction accuracy of the model.

[0155] E. On both datasets, the performance of Variant 5 is worse than MSGT, indicating that the graph regularization function can effectively control the connectivity and sparsity of the graph, thereby improving the performance of graph attention convolution.

[0156] (6) Non-stationary Conditions: This section compares the prediction performance of MSGT with the baseline model on a non-stationary dataset. The non-stationary dataset is a subset of the load datasets GEFCom2012 and GEFCom2017. Tables 5 and 6 show the average results of five repeated experiments for each model.

[0157] Table 5 Performance of each model on the GEFCom2012 non-stationary dataset

[0158]

[0159] Table 6 Performance of each model on the GEFCom2017 non-stationary dataset

[0160]

[0161] By analyzing the data in Table 5 and Table 6, the following conclusions are drawn:

[0162] A. The performance of the proposed MSGT under non-stationary conditions is generally better than that of the baseline model, indicating that MSGT exhibits strong robustness under non-stationary conditions.

[0163] B. As the number of nodes increases, the prediction accuracy of each model decreases. However, MSGT only shows slight fluctuations and its performance is relatively stable, indicating that MSGT can remain robust under non-stationary conditions.

[0164] C. Compared with the performance on the above two datasets, the performance difference of MSGT under non-stationary conditions is relatively small, indicating that MSGT can extract important spatial and temporal features even in non-stationary environments, indicating its scalability.

[0165] In another embodiment, a short-term power load forecasting device based on graph structure learning and load cycle patterns is provided, comprising:

[0166] Input strategy module, used to obtain multi-scale time-granular historical load data as input data;

[0167] The GSL module is used to process the input data using the similarity graph learning method to obtain a similarity graph, and to process the input data using the dynamic non-stationary graph learning method to obtain a dynamic non-stationary graph;

[0168] The spatiotemporal module is used to extract spatial features from similarity maps and dynamic non-stationary maps, and temporal features from input data;

[0169] A decoupling mechanism is used to decouple spatial features and temporal features based on the decoupling mechanism to obtain spatiotemporal features;

[0170] The output module is used to fuse the input data and spatiotemporal features, and then use the multi-layer perceptron to predict and output future short-term power load data.

[0171] For the specific limitations of the short-term power load forecasting device based on graph structure learning and load cycle model, please refer to the limitations of the short-term power load forecasting method based on graph structure learning and load cycle model above, which will not be repeated here.

[0172] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art would be able to make numerous modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A short-term power load forecasting method based on graph structure learning and load cycle pattern, characterized in that: The short-term power load forecasting method based on graph structure learning and load cycle pattern includes: Obtain multi-scale time-granular historical load data as input data; The similarity graph learning method is used to process the input data to obtain a similarity graph, and the dynamic non-stationary graph learning method is used to process the input data to obtain a dynamic non-stationary graph; Extract spatial features from similarity graphs and dynamic non-stationary graphs, and extract temporal features from input data; Based on the decoupling mechanism, spatial features and temporal features are decoupled to obtain spatiotemporal features; After fusing the input data and spatiotemporal features, the multi-layer perceptron is used to predict and output future short-term power load data.

2. The short-term power load forecasting method based on graph structure learning and load cycle pattern according to claim 1 is characterized in that: The acquisition of multi-scale time-granular historical load data includes: Get recent load data: Recent load data is ,in is the current time point, is the length of recent load data, is the number of nodes, For time point The load data, For time point Load data; Get daily cycle load data: Daily cycle load data is ,in is the length of the time window, is the daily historical load data, and , is the sampling rate of load data, For time point The load data, For time point The load data, For time point Before The load data of the sampling points, For time point after Load data of sampling points; Get weekly cycle load data: Weekly cycle load data is , is the weekly historical load data, and , For time point The load data, For time point The load data, For time point Before The load data of the sampling points, For time point after Load data of sampling points; The recent load data, daily cycle load data and weekly cycle load data are combined to obtain multi-scale time-granular historical load data.

3. The short-term power load forecasting method based on graph structure learning and load cycle pattern according to claim 2 is characterized in that: The method of using a similarity graph learning method to process input data to obtain a similarity graph includes: Use multi-scale dilated convolution to extract multi-scale features from input data; Splicing multi-scale features and mapping them into feature matrix ; Using similarity algorithms to process feature matrices , and obtain the similarity graph.

4. The short-term power load forecasting method based on graph structure learning and load cycle pattern according to claim 3 is characterized in that: The method of using a dynamic non-stationary graph learning method to process input data to obtain a dynamic non-stationary graph includes: Use gating mechanism to control recent load data and the feature matrix Perform adaptive feature fusion to obtain the adaptive fusion feature matrix , and adaptively fused feature matrix based on Gumbel Softmax sampling strategy Construct dynamic non-stationary graphs.

5. The short-term power load forecasting method based on graph structure learning and load cycle pattern according to claim 1 is characterized in that: The method of extracting spatial features from the similarity graph and the dynamic non-stationary graph includes: taking the similarity graph and the dynamic non-stationary graph as inputs of a graph attention convolution, and extracting and outputting spatial features by the graph attention convolution.

6. The short-term power load forecasting method based on graph structure learning and load cycle pattern according to claim 1 is characterized in that: The extracting time features from the input data includes: using a stacked LSTM network to extract the time features in the input data.

7. The short-term power load forecasting method based on graph structure learning and load cycle pattern according to claim 1 is characterized in that: The decoupling mechanism is used to decouple spatial features from temporal features to obtain spatiotemporal features, including: ； ； in, represents the weight matrix, is the activation function, is the spatial feature matrix, is the time feature matrix, is the learnable weight matrix corresponding to the spatial feature matrix, is the learnable weight matrix corresponding to the time feature matrix, is a learnable bias term, is the spatiotemporal feature matrix, is the Hadamard product.

8. The short-term power load forecasting method based on graph structure learning and load cycle pattern according to claim 1 is characterized in that: After fusing the input data and spatiotemporal features, the multi-layer perceptron is used to predict and output future short-term power load data, including: ； ； ； ； in, For input data, is the fully connected layer, is the transformation feature matrix, is the spatiotemporal feature matrix, represents the concatenation operation, is the fusion feature matrix, is a multi-layer perceptron, is the output of the multi-layer perceptron, For time point Short-term power load data forecast value.

9. The short-term power load forecasting method based on graph structure learning and load cycle pattern according to claim 1 is characterized in that: It also includes a training phase, in which the graph regularization function and the mean absolute error are calculated, and the weighted sum of the graph regularization function and the mean absolute error is used as the total loss function; The graph regularization function is calculated based on the input data and the similarity graph, and the formula is as follows: ； in, is the graph regularization function, is the number of nodes, The first The load data of each node, The first The load data of each node, represents the square of the Euclidean norm, represents the square of the Frobenius norm, is a hyperparameter, is the adjacency matrix of the similarity graph, Represents the relationship between the i-th node and the j-th node in the adjacency matrix of the similarity graph; The mean absolute error is calculated based on the predicted data and the true data, and the formula is as follows: ； in, is the mean absolute error, For time point The short-term power load data forecast value of the i-th node, For time point The true value of the short-term power load data of the i-th node.

Images