Multi-task short-term power load prediction method based on attention time sequence decomposition

By combining attention time series decomposition and multi-task graph convolution network in short-term power load prediction, the problems of deep feature extraction and generalization capabilities are solved, and more accurate and stable power load prediction is achieved.

CN119090312AActive Publication Date: 2024-12-06ZHEJIANG UNIV OF FINANCE & ECONOMICS +1

Patent Information

Application Number
CN202411286400.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-12-06
Estimated Expiration
2044-09-13

AI Technical Summary

Technical Problem

The prior art is difficult to fully capture deep-level temporal and spatial characteristics in short-term power load prediction, and the generalization ability of graph convolutional networks decreases when the number of nodes increases, resulting in a decrease in overfitting risk and prediction performance.

Method used

Using multi-task short-term power load prediction method based on attention time series decomposition, combining STL decomposition and graph convolution network through space-time embedding module, attention time series decomposition module, gated time convolution module and multi-task graph convolution module , ,using attention mechanism weighted components and multitasking learning to extract temporal and spatial features, reducing the number of nodes to enhance generalization capabilities.

Benefits of technology

Effectively capture time characteristics and improve model generalization capabilities, reduce the risk of overfitting, and improve the accuracy and robustness of short-term power load prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090312A_ABST
    Figure CN119090312A_ABST
Patent Text Reader

Abstract

The invention discloses a multitask short-term power load prediction method based on attention time series decomposition, which comprises the following steps: firstly, converting historical power loads from different regions into corresponding space-time diagram representation, and then decomposing the space-time diagram representation into a trend component, a seasonal component and a residual component; the method comprises the following steps: decomposing a plurality of components of the power load, performing weighting processing on the components obtained by decomposition, splicing the weighted components into weighted power load data, performing time feature capture on the weighted power load data through gating time convolution, and finally performing linear transformation through multi-task graph convolution and non-Euclidean spatial feature extraction to obtain a prediction result. According to the method, the attention mechanism is used for performing weighting processing on the components decomposed from the power load data through the STL, so that the model is helped to concentrate on more important components, and time features are captured more effectively. The power load data is divided into different clusters for multi-task learning, and the generalization ability of the model is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application belongs to the technical field of power load forecasting, and in particular, relates to a multi-task short-term power load forecasting method based on attention time series decomposition. Background Art

[0002] Accurate short-term power load forecasting (STLF) is crucial to the safe and efficient operation of modern power systems. On the one hand, accurate short-term power load forecasting helps grid operators prevent equipment damage or grid failure caused by sudden increases in power load. On the other hand, it also helps power plants to scientifically arrange power generation plans to improve power generation efficiency and reduce operating costs.

[0003] Short-term power load forecasting methods can be divided into three categories: statistical methods, machine learning methods, and deep learning methods. Statistical methods, such as autoregressive model (AR), autoregressive moving average model (ARMA), and autoregressive integrated moving average model (ARIMA), have the advantages of simplicity, strong interpretability, and easy implementation, but it is difficult to capture nonlinear features. Machine learning methods, such as support vector regression (SVR), multi-layer perceptron (MLP), and feedforward neural network (FNN), are superior in processing nonlinear relationships, but it is difficult to process complex functions and feature engineering. Deep learning methods, such as gated recurrent unit (GRU), convolutional neural network (CNN), long short-term memory network (LSTM), and temporal convolutional network (TCN), have powerful feature extraction capabilities and have been widely used in the field of power load forecasting in the past decade. However, although deep learning-based methods have achieved certain results in the field of short-term load forecasting, they still face great challenges in capturing the deep temporal and spatial features of power load data.

[0004] STL (Seasonal and Trend decomposition using Loess) is an algorithm for time series decomposition, which decomposes time series data into trend component, seasonal component and residual component. Each component provides corresponding features for deep learning methods, which helps to extract complex time information in power load data. However, current research ignores the weight differences of different components decomposed by STL when extracting time features, and thus cannot fully extract deep time features.

[0005] Graph convolutional network (GCN) is a neural network that processes non-Euclidean spatial relationships and is often used to extract spatial features of power load data. Assuming that each network node represents a set of power load data, the graph convolutional network can aggregate the information of neighboring nodes and update the state of the current node after nonlinear transformation. However, as the number of nodes continues to increase, the number of edges and the size of the adjacency matrix will also increase, causing the spatial pattern of power load data to become too complex, increasing the risk of overfitting and reducing the generalization ability of the graph convolutional network. Summary of the invention

[0006] The purpose of this application is to provide a multi-task short-term power load forecasting method based on attention time series decomposition to overcome the technical problems pointed out in the background technology.

[0007] In order to achieve the above purpose, the technical solution of this application is as follows: A multi-task short-term power load forecasting method based on attention time series decomposition, comprising: Constructing and training a prediction network model, wherein the prediction network model includes a spatiotemporal embedding module, an attention time series decomposition module, a gated temporal convolution module, and a multi-task graph convolution module; In the spatiotemporal embedding module, the historical power loads from different regions are converted into corresponding spatiotemporal graph representations through spatiotemporal embedding; In the attention time series decomposition module, the spatiotemporal graph representation is decomposed into trend component, seasonal component and residual component, and each component obtained by decomposition is weighted, and each weighted component is spliced ​​into weighted power load data; In the gated time convolution module, the time characteristics of the weighted power load data are captured to obtain a feature expression containing the time characteristics; The feature expression containing temporal features is input into the multi-task graph convolution module, and the non-Euclidean spatial features are extracted based on the graph convolution network with a learnable adjacency matrix to obtain the feature expression containing spatial features. The feature expression containing spatial features is then linearly transformed to obtain the prediction result.

[0008] Furthermore, the time-space diagram representation is decomposed into a trend component, a seasonal component and a residual component, and each component obtained by the decomposition is weighted, and each weighted component is spliced ​​into weighted power load data, including: The spatiotemporal graph representation is decomposed into trend component, seasonal component and residual component by using STL time series decomposition method; Global average pooling is used to compress the feature map of each component into a scalar; Use the gating mechanism to process the scalar corresponding to each component to obtain the weight of each component; Multiply each component by the weight and concatenate them to obtain the weighted power load data.

[0009] Furthermore, the gating mechanism is used to process the scalar corresponding to each component to obtain the weight of each component, including: in, is a scalar generated by the trend component, is a scalar generated by the seasonal component, is a scalar generated by the residual component, is the weight of the trend component, is the weight of the seasonal component, is the weight of the residual component, and is the learnable weight matrix, is the ReLU function, is the Sigmoid function.

[0010] Furthermore, the feature expression containing the spatial feature is linearly transformed to obtain a prediction result, and the following operations are performed: in, is the prediction result, represents the learnable weight matrix, is the bias, Represents a feature expression containing spatial features.

[0011] Furthermore, the multi-task short-term power load forecasting method based on attention time series decomposition also includes: When training the prediction network model, the spatiotemporal graph representation after spatiotemporal embedding is divided into different clusters, and then the prediction results are divided into cluster prediction results corresponding to each cluster according to the number of clusters. The average loss corresponding to each cluster is calculated, and the parameters of the prediction network model are optimized to obtain a trained prediction network model.

[0012] Furthermore, the average loss calculation formula is as follows: in, is the jth cluster to be divided, is the number of clusters, is the cluster prediction result corresponding to the jth cluster, is the average loss.

[0013] The multi-task short-term power load forecasting method based on attention time series decomposition proposed in this application has the following technical effects: On the one hand, an ASTL method combining attention mechanism with STL is proposed, that is, the attention mechanism is used to weight the components decomposed from the power load data by STL to help the model focus on more important components, thereby capturing temporal features more effectively.

[0014] On the other hand, a multi-task graph convolutional network (MG) combining multi-task learning with graph convolutional networks is proposed. That is, the power load data is divided into different clusters for multi-task learning, which helps to learn simple spatial patterns with fewer nodes, thereby enhancing the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of the multi-task short-term power load forecasting method based on attention time series decomposition in this application.

[0016] Figure 2 This is a schematic diagram of the prediction network model structure for this application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] An embodiment of the present application, such as Figure 1 As shown, a multi-task short-term power load forecasting method based on attention time series decomposition is provided, including: Step S1, construct and train a prediction network model, wherein the prediction network model includes a spatiotemporal embedding module, an attention time series decomposition module, a gated temporal convolution module, and a multi-task graph convolution module.

[0019] This embodiment constructs and trains a prediction network model (abbreviated as: MG-ASTL). The framework of the model is as follows: Figure 2As shown. The MG-ASTL model includes a spatiotemporal embedding module, an attention time series decomposition module (ASTL), a gated temporal convolution module (gated TCN), and a multi-task graph convolution module (MG). First, the spatiotemporal embedding module converts the historical power loads in different regions into corresponding spatiotemporal graphs for representation. Next, the ASTL module uses STL to decompose the power load data into three components: trend component, seasonal component, and residual component, and weights them through the attention mechanism. Then, the gated TCN module further extracts the temporal features. Finally, the MG module divides the power load data into different clusters for multi-task learning, and extracts spatial features through graph convolution GCN.

[0020] Step S2: In the spatiotemporal embedding module, historical power loads from different regions are converted into corresponding spatiotemporal graph representations through spatiotemporal embedding.

[0021] Spatiotemporal embedding is the basis of the spatiotemporal prediction model based on the graph neural network GNN, which is a relatively mature technology in this field. In order to embed the power load data of different regions into the graph, this embodiment assumes that each node represents a region, and the historical power loads from different regions can be converted into corresponding spatiotemporal graph representations by splicing. ,in Indicates time The historical power load characteristic matrix of different regions at that time is is the feature number, is the current time, is the number of regions, is the historical time step.

[0022] Step S3: In the attention time series decomposition module, the space-time graph representation is decomposed into a trend component, a seasonal component and a residual component, and each component obtained by the decomposition is weighted, and each weighted component is spliced ​​into weighted power load data.

[0023] STL (Seasonal-Trend decomposition procedure based on Loess) is a data preprocessing method that can decompose power load data into three components: trend component, seasonal component and residual component. Generally speaking, different components have different importance for temporal feature extraction. Therefore, the ASTL module of this embodiment performs weighted processing on the components decomposed by STL through the attention mechanism to improve the ability of STL to capture temporal features.

[0024] First, the space-time graph identifier is decomposed into three components using the STL time series decomposition method. The process is shown in formula (1):

[0025] (1) in, A spatiotemporal graph representation of historical power loads from different regions, is the trend component, is the seasonal component, is the residual component, It is an STL operation.

[0026] Next, global average pooling is used to compress the feature map of each component into a scalar, which represents the global distribution of features among the components. The process is shown in formula (2-4):

[0027] (2) (3) (4) in, is a scalar generated by the trend component, is a scalar generated by the seasonal component, is a scalar generated by the residual component, and is the index, is the number of regions, It is the input window.

[0028] Then, the gating mechanism is used to process the scalar corresponding to each component to obtain the weight of each component. This process is shown in formula (5-7).

[0029] (5) (6) (7) in, is the weight of the trend component, is the weight of the seasonal component, is the weight of the residual component, and is the learnable weight matrix, is the ReLU function, is the Sigmoid function.

[0030] Finally, each component is multiplied by a weight, so that the MG-ASTL model can focus on more important components to achieve more effective temporal feature extraction. This process is shown in Equation (8-11).

[0031] (8) (9) (10) (11) in, is the weighted trend component, is the weighted seasonal component, is the weighted residual component, represents the weighted power load data, It is a splicing operation.

[0032] This embodiment achieves attention-based time series decomposition by weighting each decomposed component. Different components have different importance for time feature extraction. The ASTL module of this embodiment weights the components decomposed by STL through the attention mechanism to improve the ability of STL to capture time features.

[0033] Step S4: In the gated time convolution module, the time characteristics of the weighted power load data are captured to obtain a feature expression including the time characteristics.

[0034] The gated TCN module uses TCN to further capture temporal features and uses a gating mechanism to control information flow to avoid gradient vanishing or gradient exploding. The process is shown in equations (12-17). Equation (12) calculates the input gate; Equation (13) calculates the forget gate; Equation (14) calculates the output gate; Equation (15) calculates the candidate cell state; Equation (16) uses the forget gate and input gate to control The cell state at each moment and The information flow of candidate cell states at time The cell state at the moment; Formula (17) uses the output gate to control The information flow of the cell state at each moment is obtained By iteratively calculating in the time dimension, the gated TCN module finally obtains a feature expression containing time features .

[0035] (12) (13) (14) (15) (16) (17) in, express The input gate of the moment, express The door of forgetfulness of time, express The output gate at time, express The candidate cell states at time t, express The cell state at a moment in time, express The feature expression of the moment containing the time feature, express Weighted power load data at the moment; , , and is the deviation, is the time step, is the Sigmoid function, is the Tanh function, It is a TCN operation. It is Hadamard.

[0036] Step S5: Input the feature expression containing the temporal feature into the multi-task graph convolution module, extract the non-Euclidean spatial features based on the graph convolution network with a learnable adjacency matrix, obtain the feature expression containing the spatial feature, and then perform a linear transformation on the feature expression containing the spatial feature to obtain the prediction result.

[0037] In this embodiment, the feature expression containing time features obtained by the gated time convolution module is input to the MG module. In the MG module, a graph convolutional network based on a learnable adjacency matrix is ​​used to extract non-Euclidean space features. The process is shown in Equations (18-19).

[0038] (18) (19) in, represents the feature expression containing time features, represents the feature expression containing spatial features, represents the learnable adjacency matrix, represents the diagonal node degree matrix, represents the identity matrix, represents the learnable weight matrix, represents the node embedding matrix, is the ReLU function, is the SoftMax function.

[0039] Then, the corresponding prediction results are generated through linear transformation. The process is shown in formula (20).

[0040] (20) in, is the prediction result, represents the learnable weight matrix, is bias.

[0041] In another embodiment of the present application, since the graph convolutional network can extract non-Euclidean spatial features in power load data. However, as the number of nodes increases, the generalization ability of the graph convolutional network decreases. Therefore, this embodiment of a multi-task short-term power load forecasting method based on attention time series decomposition also includes:

[0042] When training the prediction network model, the spatiotemporal graph representation after spatiotemporal embedding is divided into different clusters, and then the prediction results are divided into cluster prediction results corresponding to each cluster according to the number of clusters. The average loss corresponding to each cluster is calculated, and the parameters of the prediction network model are optimized to obtain a trained prediction network model.

[0043] Specifically, this embodiment uses a multi-task learning approach to train the network model, thereby learning simple spatial patterns with fewer nodes to enhance the generalization ability of the graph convolutional network. For example, OPTICS (Ordering points to identify the clustering structure) is used to divide the spatiotemporal graph identifiers corresponding to the power load data of different regions into different clusters, and the prediction results are divided according to the number of clusters. The power load data and cluster prediction results of each cluster generate a task for multi-task learning. Then the loss value of each task is calculated by the mean square error loss function, and the average loss on all tasks is obtained by adding and averaging. Finally, the network parameters are optimized by the Adam algorithm to obtain a model that performs well on all tasks. The process is shown in formulas (21-23).

[0044] (twenty one) (twenty two) (twenty three) in, is the jth cluster to be divided, A spatiotemporal graph representation of historical power loads from different regions, is the number of clusters, It is OPTICS operation. is the cluster prediction result corresponding to the jth cluster, is the average loss, is the index.

[0045] This embodiment divides the power load data into different clusters for multi-task learning to improve the generalization ability of the graph convolutional network.

[0046] In order to verify the above technical solution, this application also gives experimental results. Two real power data sets GEFCom2012 and GEFCom 2017 are used in the experiment to verify the effectiveness of the MG-ASTL model. The GEFCom2012 data set contains 39,576 energy data records from 20 regions in the United States from January 1, 2004 to June 30, 2008. The GEFCom2012 data set contains 61,344 energy data records from 500 regions in the United States from January 1, 2005 to December 31, 2011. This application will select part of the power load data from the above two data sets for experiments.

[0047] In the experiment, the MG-ASTL model is compared with the following baseline models to verify the effectiveness of the MG-ASTL model in short-term power load forecasting: GRU: Recurrent neural network based on double gating mechanism; LSTM: a recurrent neural network based on a triple gating mechanism; CNN-LSTM: A hybrid model that combines CNN and LSTM; Graph WaveNet: A hybrid model combining diffuse convolution and adaptive graph. CNN-BiLSTM: A hybrid model that combines CNN and bidirectional LSTM; GCN-LSTM: A hybrid model that combines GCN and LSTM.

[0048] In order to evaluate the MG-ASTL model from multiple perspectives, the experiment used three evaluation indicators, namely, mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE).

[0049] In the experiments, the input window size is set to 24, and the output window size is set to 3, 7, and 12. The selected regions are set to 10, 15, and 20 on the GEFCom2012 dataset, and the selected regions are set to 20, 60, and 80 on the GEFCom2017 dataset. The weight decay is set to 0.0001, the learning rate is set to 0.001, the training epochs are set to 100, the batch size is set to 64, the loss function is set to MSE, the depth of the node embedding matrix is ​​set to 10, the optimizer is set to Adam, the number of GCN layers is set to 2, and the hidden size is set to 64. For fairness, the hyperparameter values ​​of all baseline models are from their original references.

[0050] The experiments compared the performance of the MG-ASTL model with the baseline model under different conditions on two datasets. Tables 1 and 2 show the average results of five repeated experiments for each model.

[0051] Table 1 Table 2 Among them, Table 1 is the performance comparison of the MG-ASTL model and the baseline model on the GEFCom2012 dataset, and Table 2 is the performance comparison of the MG-ASTL model and the baseline model on the GEFCom2017 dataset. From Tables 1 and 2, the following conclusions can be drawn: (1) The overall performance of MG-ASTL on both datasets is better than that of most baseline models, which verifies the effectiveness of MG-ASTL.

[0052] (2) As the number of nodes increases, the performance of GRU and LSTM on the two datasets becomes worse and worse, indicating that GRU and LSTM are insufficient in extracting spatial features.

[0053] (3) Compared with CNN-LSTM and CNN-BiLSTM which use CNN to capture Euclidean space features, GraphWaveNet and GCN-LSTM perform better on both datasets, indicating that non-Euclidean space features are more effective than Euclidean space features in short-term power load forecasting.

[0054] (4) As the number of nodes increases, the performance of MG-ASTL on the two datasets gets better and better, indicating that MG-ASTL can still maintain good spatial feature extraction capabilities when the number of nodes increases.

[0055] (5) When the output window increases from 3 hours to 12 hours, MG-ASTL still performs well on both datasets, indicating that MG-ASTL has good robustness.

[0056] At the same time, an ablation experiment was also conducted to compare the performance of MG-ASTL and its different variant models to verify the necessity of each module. The variant models are as follows:

[0057] Variant 1: MG-ASTL without the ASTL module, i.e., the weighted component is not used as an additional feature; Variant 2: MG-ASTL of DBSCAN is removed, that is, OPTICS is not used to divide the power load data into different clusters; Variant 3: MG-ASTL without the gated TCN module, i.e., gated TCN is not used to capture temporal features; Variant 4: MG-ASTL without the MG module, that is, the MG module is not used to capture spatial features.

[0058] The performance of MG-ASTL and its variant models under different conditions are compared on two datasets. Tables 3 and 4 show the average results of each model after five repeated experiments.

[0059] Table 3 Table 4 Among them, Table 3 is the performance comparison of the MG-ASTL model and the variant model on the GEFCom2012 dataset, and Table 4 is the performance comparison of the MG-ASTL model and the variant model on the GEFCom2017 dataset. The following conclusions can be drawn from Tables 3 and 4: (1) Variant 1 performs worse than MG-ASTL on both datasets, indicating that the ASTL module allows the model to focus on more important components to effectively extract temporal features.

[0060] (2) Variant 2 performs worse than MG-ASTL on both datasets, indicating that the MG module can utilize OPTICS to capture spatial features and enhance generalization ability.

[0061] (3) Variant 3 performs worse than MG-ASTL on both datasets, indicating that temporal features are very important for MG-ASTL to perform short-term power load forecasting.

[0062] (4) Variant 4 performs worse than MG-ASTL on both datasets, indicating that spatial features are very important for MG-ASTL to perform short-term power load forecasting.

[0063] This application proposes an MG-ASTL model for short-term power load forecasting, which uses ASTL and MG to capture temporal and spatial features respectively. Specifically, ASTL uses an attention mechanism to weight different components, which helps MG-ASTL focus on more important components and thus capture temporal features more effectively. MG uses DBSCAN to divide power load data into different clusters, which helps MG-ASTL learn simple spatial patterns and thus improve generalization capabilities. A large number of comparative experiments show that MG-ASTL performs better than the baseline model.

[0064] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A multi-task short-term power load forecasting method based on attention time series decomposition, characterized in that: The multi-task short-term power load forecasting method based on attention time series decomposition includes: Constructing and training a prediction network model, wherein the prediction network model includes a spatiotemporal embedding module, an attention time series decomposition module, a gated temporal convolution module, and a multi-task graph convolution module; In the spatiotemporal embedding module, the historical power loads from different regions are converted into corresponding spatiotemporal graph representations through spatiotemporal embedding; In the attention time series decomposition module, the spatiotemporal graph representation is decomposed into trend component, seasonal component and residual component, and each component obtained by decomposition is weighted, and each weighted component is spliced ​​into weighted power load data; In the gated time convolution module, the time characteristics of the weighted power load data are captured to obtain a feature expression containing the time characteristics; The feature expression containing temporal features is input into the multi-task graph convolution module, and the non-Euclidean spatial features are extracted based on the graph convolution network with a learnable adjacency matrix to obtain the feature expression containing spatial features. The feature expression containing spatial features is then linearly transformed to obtain the prediction result.

2. The multi-task short-term power load forecasting method based on attention time series decomposition as claimed in claim 1 is characterized in that: The time-space diagram is decomposed into a trend component, a seasonal component and a residual component, and each component obtained by the decomposition is weighted, and each weighted component is spliced ​​into weighted power load data, including: The spatiotemporal graph representation is decomposed into trend component, seasonal component and residual component by using STL time series decomposition method; Global average pooling is used to compress the feature map of each component into a scalar; Use the gating mechanism to process the scalar corresponding to each component to obtain the weight of each component; Multiply each component by the weight and concatenate them to obtain the weighted power load data.

3. The multi-task short-term power load forecasting method based on attention time series decomposition as claimed in claim 2 is characterized in that: The gating mechanism is used to process the scalar corresponding to each component to obtain the weight of each component, including: in, is a scalar generated by the trend component, is a scalar generated by the seasonal component, is a scalar generated by the residual component, is the weight of the trend component, is the weight of the seasonal component, is the weight of the residual component, and is the learnable weight matrix, is the ReLU function, is the Sigmoid function.

4. The multi-task short-term power load forecasting method based on attention time series decomposition as claimed in claim 1, characterized in that: The feature expression containing the spatial feature is linearly transformed to obtain the prediction result, and the following operations are performed: in, is the prediction result, represents the learnable weight matrix, is the bias, Represents a feature expression containing spatial features.

5. The multi-task short-term power load forecasting method based on attention time series decomposition as claimed in claim 1, characterized in that: The multi-task short-term power load forecasting method based on attention time series decomposition also includes: When training the prediction network model, the spatiotemporal graph representation after spatiotemporal embedding is divided into different clusters, and then the prediction results are divided into cluster prediction results corresponding to each cluster according to the number of clusters. The average loss corresponding to each cluster is calculated, and the parameters of the prediction network model are optimized to obtain a trained prediction network model.

6. The multi-task short-term power load forecasting method based on attention time series decomposition as claimed in claim 5, characterized in that: The average loss calculation formula is as follows: in, is the jth cluster to be divided, is the number of clusters, is the cluster prediction result corresponding to the jth cluster, is the average loss.

Citation Information

Patent Citations

  • Short-term power load prediction method, device, equipment and medium

    CN115293046A

  • Traffic flow prediction method based on multi-scale spatial feature mining

    CN116070778A

  • Short-term power load prediction method and system based on time-space diagram attention

    CN116703000A

  • Probability-oriented power load prediction method based on graph convolutional network model

    CN116960991A

  • Electrical load prediction method and device based on space-time correlation

    CN117175588A

Cited By

  • Mountain area expressway traffic flow prediction method based on STL decomposition and iTransform model

    CN121579878A