DEST-GNN-based multi-region short-term power load prediction method
By employing a multi-regional short-term power load forecasting method based on DEST-GNN and utilizing sparse attention mechanism and adaptive graph convolutional network, the problem of difficulty in mining spatiotemporal correlation in traditional methods is solved, achieving high-precision and low-complexity power load forecasting, adapting to the volatility of new energy sources, and supporting the stable operation of the power grid.
Patent Information
- Application Number
- CN202510900859.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-01
- Publication Date
- 2025-10-28
AI Technical Summary
Traditional load forecasting methods struggle to fully exploit the complex spatiotemporal correlations in power load data, especially when dealing with multi-regional load data, as they cannot effectively utilize the mutual influence information between regions. Furthermore, the models are highly complex and consume significant computational resources. The intermittency and volatility of renewable energy sources pose new challenges to the stable operation of the power grid.
A multi-regional short-term power load forecasting method based on DEST-GNN is adopted. By constructing a sparse attention mechanism and an adaptive graph convolutional network, combined with a temporal convolutional network, an adaptive adjacency matrix is built to filter low-correlation information, capture spatiotemporal features, reduce computational burden and improve forecasting accuracy.
It improves the accuracy and applicability of multi-regional power load forecasting, reduces model complexity and computational resource consumption, enhances adaptability to new energy sources, and supports stable grid operation and optimized dispatching.
Smart Images
Figure CN120855281A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power load forecasting technology, specifically a multi-regional short-term power load forecasting method based on DEST-GNN. Background Technology
[0002] With the continuous development of power systems and the transformation and upgrading of energy structures, accurate power load forecasting is of great significance for the stable operation of power grids, optimized dispatching, and efficient consumption of new energy sources. Traditional load forecasting methods often struggle to fully exploit the complex spatiotemporal correlations within power load data, especially when dealing with multi-regional load data, failing to effectively utilize inter-regional interaction information to improve forecast accuracy. In recent years, deep learning and graph neural network technologies have demonstrated enormous potential in handling complex data relationships. The DEST-GNN model, as an advanced graph neural network method, has achieved significant results in fields such as photovoltaic power forecasting.
[0003] Currently, conventional load forecasting models typically focus only on the temporal characteristics of time series data, neglecting the spatial correlation between power loads in different regions. Furthermore, while pursuing high forecasting accuracy, model complexity and computational resource consumption are also important considerations. In addition, as the proportion of new energy sources such as photovoltaic power generation in the power system continues to increase, their intermittency and volatility pose new challenges to the stable operation of the power grid. Therefore, in view of the above situation, there is an urgent need to develop a multi-regional short-term power load forecasting method based on DEST-GNN to overcome the shortcomings in current practical applications. Summary of the Invention
[0004] The purpose of this invention is to provide a multi-regional short-term power load forecasting method based on DEST-GNN to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A multi-regional short-term power load forecasting method based on DEST-GNN includes the following steps:
[0007] (1) Data preparation: Collect historical load data, meteorological data and time information from multiple regions, and perform preprocessing and dataset division;
[0008] (2) Constructing a sparse attention mechanism:
[0009] Time dimension: Calculate the correlation between different time steps in the time series and filter out low-correlation time steps by using a threshold;
[0010] Spatial dimension: Calculate the correlation between different regions and filter low-correlation regions by using a threshold;
[0011] (3) Construct an adaptive graph convolutional network: Generate an adaptive adjacency matrix to dynamically represent the spatial relationships between regions; perform graph convolution operations based on the adaptive adjacency matrix to extract spatial features;
[0012] (4) Construct a temporal convolutional network: extract time series features through convolution operations;
[0013] (5) Model integration: The outputs of steps (2)-(4) are integrated to construct a prediction model and output multi-regional load prediction values;
[0014] (6) Model training and evaluation: Optimize model parameters and verify prediction accuracy.
[0015] As a further aspect of the present invention: in step (2), the time dimension correlation is calculated using the following formula:
[0016]
[0017] in, This is the normalized time correlation value; The elements in the original time correlation matrix represent the correlation between time steps i and i′, reflecting the correlation between time step i. ′ The importance of the data to the load value of prediction time step i; θ is the correlation threshold, used to determine whether to retain the correlation of this time step;
[0018] The formula for calculating spatial dimensional correlation is as follows:
[0019]
[0020] in, This is the normalized spatial correlation matrix; θ represents the element in the original spatial correlation matrix, indicating the correlation between regions i and i′; θ is the correlation threshold, used to determine whether to retain the correlation of this region.
[0021] As a further aspect of the present invention: In step (3), the adaptive adjacency matrix is used to replace the static adjacency matrix based on distance or Pearson correlation, and the adaptive adjacency matrix is generated as follows:
[0022] A adapt =softmax(δ(W·W) T ));
[0023] Among them, A adapt δ is an adaptive adjacency matrix; W is a learnable matrix, initially randomized and continuously optimized through training; δ is the ReLU activation function, used to handle complex graph structure relationships.
[0024] As a further aspect of the present invention: the calculation method for the graph convolution operation is as follows:
[0025]
[0026] Among them, H (l+1) Θ is the node feature matrix of the (l+1)th layer; σ is the activation function used to introduce nonlinearity; (k) The Chebyshev polynomial coefficient vector is initialized using Glorot at the start of model training; H (l) is the node feature matrix of the l-th layer; K is the order of the Chebyshev polynomial, which can take any value from 2 to 5; This is the normalized graph Laplace matrix.
[0027] As a further aspect of the present invention: in step (4), the temporal convolutional network is calculated as follows:
[0028] O (t) =Conv1D(I (t) W conv ,b conv );
[0029] Among them, O (t) This is the output of the temporal convolutional network; I (t) W is the input to the temporal convolutional network. conv b is the convolution kernel weight matrix; conv This is the kernel bias term.
[0030] As a further aspect of the present invention: In step (5), the sparse attention mechanism, adaptive graph convolutional network, and temporal convolutional network are integrated to construct a complete load prediction model:
[0031]
[0032] in, X is the model's predicted output; A is the model's input; X is the model's predicted output. adapt W is an adaptive adjacency matrix. att W is the learnable weight matrix in the sparse attention mechanism. conv b is the kernel weight matrix in a temporal convolutional network; conv is the kernel bias term in the temporal convolutional network; f is the overall structure and computational logic of the model.
[0033] As a further aspect of the present invention: in step (1), the time information includes date type, holiday markers, and day / night markers;
[0034] Preprocessing includes missing value imputation, outlier correction, and normalization.
[0035] As a further aspect of the present invention: in step (6), mean square error and mean absolute error are used as loss functions;
[0036] The evaluation metrics include mean absolute error, root mean square error, and mean absolute percentage error.
[0037] As a further aspect of the present invention, the evaluation indicators satisfy: MAE≤3.35kW, RMSE≤3.48kW, MAPE≤3.85%.
[0038] Compared with the prior art, the beneficial effects of the present invention are:
[0039] 1. By constructing a graph neural network and combining sparse attention mechanism and adaptive graph convolutional network, the spatiotemporal correlation in multi-regional power load data is fully explored to improve prediction accuracy;
[0040] 2. By introducing sparse attention mechanism and adaptive graph convolutional network, it is possible to reduce model complexity and computational burden while maintaining high prediction accuracy, thereby improving the applicability and practicality of the model.
[0041] 3. By drawing on the experience of the DEST-GNN model in handling photovoltaic power prediction problems, we can improve its adaptability to new energy access and provide stronger support for the stable operation and optimized scheduling of the power grid. Attached Figure Description
[0042] Figure 1 This is a flowchart of the multi-regional short-term power load forecasting method based on DEST-GNN in an embodiment of the present invention.
[0043] Figure 2 This is a schematic diagram illustrating the application differences of various algorithms when a random device is selected in an embodiment of the present invention. Detailed Implementation
[0044] 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.
[0045] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.
[0046] Please see Figure 1 and Figure 2 The present invention provides a multi-regional short-term power load forecasting method based on DEST-GNN, the specific steps of which are as follows:
[0047] Step 1: Problem Definition and Data Preparation;
[0048] 1.1 Data Collection: Collect historical load data, weather data, and historical time information for all devices within the same area, with a time resolution of 15 minutes. In addition to the basic time, the time information includes the day of the week (1-7 representing Monday to Sunday), whether it is a holiday (0 for no, 1 for yes), and whether it is nighttime (0 for no, 1 for yes).
[0049] 1.2 Data Preprocessing:
[0050] Missing value handling: interpolating or filling in missing data points;
[0051] Outlier handling: Identify and correct outliers;
[0052] Normalization: Standardize the data to a certain range ([-1,1]) to facilitate model training;
[0053] Dataset splitting: The dataset is split into a training set (60%), a validation set (20%), and a test set (20%).
[0054] Step 2: Construct a sparse attention mechanism;
[0055] By employing a sparse attention mechanism, the influence of low-correlation or unrelated information in the input data is reduced, thus lowering the computational burden and improving model performance. This involves calculating the correlation between different time steps in the time series data and ignoring time steps that have little impact on the prediction results—essentially introducing sparse temporal attention. The specific calculation formula is as follows:
[0056]
[0057] in, This is the normalized time correlation value; The elements in the original time correlation matrix represent the correlation between time steps i and i′, reflecting the correlation between time step i. ′ The data represents the importance of the load value at prediction time step i; θ is the correlation threshold, used to determine whether to retain the correlation of that time step, and the optimal value is determined by methods such as cross-validation.
[0058] Then, sparse spatial attention is added to assess the spatial correlation between different regions, filtering out other regions with weak load correlation with the target region. The specific calculation formula is as follows:
[0059]
[0060] in, This is the normalized spatial correlation matrix; θ represents the element in the original spatial correlation matrix, indicating the correlation between regions i and i′; θ is the correlation threshold, used to determine whether to retain the correlation of this region.
[0061] Step 3: Construct an adaptive graph convolutional network;
[0062] An adaptive graph convolutional network is used to learn the complex spatial correlations between different regions and construct a dynamic graph structure. The graph structure is built using a learnable adjacency matrix, replacing the traditional static adjacency matrix based on distance or Pearson correlation (also called adaptive adjacency matrix learning). The specific formula is as follows:
[0063] A adapt =softmax(δ(W·W) T ));
[0064] Among them, A adapt δ is an adaptive adjacency matrix; W is a learnable matrix, initially randomized and continuously optimized through training; δ is the ReLU activation function, used to handle complex graph structure relationships.
[0065] Then, graph convolution is performed based on the adaptive adjacency matrix to extract the feature information of the nodes. The specific calculation formula is as follows:
[0066]
[0067] Among them, H (l+1) Θ is the node feature matrix of the (l+1)th layer; σ is the activation function used to introduce nonlinearity; (k) The Chebyshev polynomial coefficient vector is initialized using Glorot at the start of model training; H (l) is the node feature matrix of the l-th layer; K is the order of the Chebyshev polynomial, which can take any value from 2 to 5; This is the normalized graph Laplace matrix.
[0068] Step 4: Construct the temporal convolutional network and ensemble model;
[0069] This paper describes a method to capture the temporal dependencies in time-series data using a temporal convolutional network. It integrates sparse attention mechanisms, adaptive graph convolutional networks, and a temporal convolutional network to construct a complete load prediction model. First, one-dimensional convolution operations are used to extract temporal features from the time-series data, capturing the changing patterns of load data over time. The specific formula is as follows:
[0070] O (t) =Conv1D(I (t) W conv ,b conv );
[0071] Among them, O(t) This is the output of the temporal convolutional network; I (t) W is the input to the temporal convolutional network. conv b is the convolution kernel weight matrix; conv This is the kernel bias term.
[0072] Then, the sparse attention mechanism, adaptive graph convolutional network, and temporal convolutional network are integrated to construct a complete load prediction model, as shown in the following formula:
[0073]
[0074] in, X represents the model's predicted output, i.e., the predicted electricity load for different regions over multiple future time periods; X represents the model's input, including historical load data and other feature information; A adapt W is an adaptive adjacency matrix. att W is the learnable weight matrix in the sparse attention mechanism. conv b is the kernel weight matrix in a temporal convolutional network; conv is the kernel bias term in the temporal convolutional network; f is the overall structure and computational logic of the model, integrating sparse attention mechanism, adaptive graph convolutional network and temporal convolutional network.
[0075] Step 5: Model training and evaluation;
[0076] The model is trained using training data, and its parameters are adjusted by optimizing the loss function (using a combination of mean squared error and mean absolute error) so that the model can accurately predict power load.
[0077] Example: Data was collected in a residential area from January to April 2025. The specific data is as follows:
[0078] Historical load data: Electricity load data for each household in this residential area was collected, with a sampling frequency of once every 15 minutes. The data includes the active and reactive power of each household.
[0079] Meteorological data: Temperature, humidity, dew point, and cloud cover data were obtained from local weather stations, with sampling frequency once per hour. These meteorological factors have a significant impact on residents' electricity consumption behavior.
[0080] Holiday and Special Event Information: This section collects information on holidays (such as May Day) and special events (such as large-scale local events) during this period, as these factors can cause abnormal fluctuations in power load.
[0081] Model application and validation:
[0082] The preprocessed data is input into a DEST-GNN-based load forecasting model. The model's prediction results are recorded and compared with actual load data. Prediction error metrics, such as mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), are calculated to evaluate the model's prediction accuracy. Simultaneously, this model is compared with other common forecasting models, such as LSTM (Long Short-Term Memory), CNN (Convolutional Neural Network), and ARIMA (Autoregressive Moving Average), to verify its superiority.
[0083] Model Name MAE(kW) RMSE (kW) MAPE (%) This plan 3.35 3.48 3.85 LSTM 7.52 7.71 9.12 CNN 7.61 7.83 9.78 ARIMA 13.75 12.92 14.36
[0084] like Figure 2 The figure shows the differences in the application of various algorithms on a randomly selected device (during two consecutive random days in mid-April). It is clear that the present invention is closer to the true value.
[0085] In summary, this invention constructs a graph neural network to represent electricity load data from different regions as a graph structure, and utilizes a graph convolutional network to capture the spatial correlation between regions. By introducing a sparse attention mechanism, load data is modeled from both temporal and spatial dimensions, effectively filtering out low-relevance or unrelated information, reducing computational burden, and improving model performance.
[0086] An adaptive adjacency matrix learning method is proposed, which uses a learnable adjacency matrix to construct a dynamic graph structure, replacing the traditional static adjacency matrix based on distance or Pearson correlation. This method can better capture the complex spatial correlations between different regions and adapt to dynamic changes in the data.
[0087] Furthermore, sparse temporal attention and sparse spatial attention modules were designed to calculate the correlation between different time steps and the spatial correlation between different regions in the time series data, respectively. By setting a correlation threshold, time steps and regions with little impact on the prediction results are filtered out, thereby reducing noise and redundant information in the input data.
[0088] Simultaneously, a sparse attention mechanism, adaptive graph convolutional network, and temporal convolutional network are integrated to construct a complete load forecasting model. This integrated approach can fully utilize the advantages of each component to achieve comprehensive mining of the spatiotemporal characteristics of load data. A multi-regional short-term power load forecasting framework based on graph neural networks is proposed. Through model training and optimization, high-precision forecasting of multi-regional power load is achieved.
[0089] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A multi-regional short-term power load forecasting method based on DEST-GNN, characterized in that, Includes the following steps: (1) Data preparation: Collect historical load data, meteorological data and time information from multiple regions, and perform preprocessing and dataset division; (2) Constructing a sparse attention mechanism: Time dimension: Calculate the correlation between different time steps in the time series and filter out low-correlation time steps by using a threshold; Spatial dimension: Calculate the correlation between different regions and filter low-correlation regions by using a threshold; (3) Construct an adaptive graph convolutional network: Generate an adaptive adjacency matrix to dynamically represent the spatial relationships between regions; perform graph convolution operations based on the adaptive adjacency matrix to extract spatial features; (4) Construct a temporal convolutional network: extract time series features through convolution operations; (5) Model integration: The outputs of steps (2)-(4) are integrated to construct a prediction model and output multi-regional load prediction values; (6) Model training and evaluation: Optimize model parameters and verify prediction accuracy.
2. The multi-regional short-term power load forecasting method based on DEST-GNN according to claim 1, characterized in that, In step (2), the time dimension correlation is calculated using the following formula: in, This is the normalized time correlation value; The elements in the original time correlation matrix represent the correlation between time steps i and i′, reflecting the correlation between time step i. ′ The importance of the data to the load value of prediction time step i; θ is the correlation threshold, used to determine whether to retain the correlation of this time step; The formula for calculating spatial dimensional correlation is as follows: in, This is the normalized spatial correlation matrix; θ represents the element in the original spatial correlation matrix, indicating the correlation between regions i and i′; θ is the correlation threshold, used to determine whether to retain the correlation of this region.
3. The multi-regional short-term power load forecasting method based on DEST-GNN according to claim 1, characterized in that, In step (3), the adaptive adjacency matrix is used to replace the static adjacency matrix based on distance or Pearson correlation. The adaptive adjacency matrix is generated as follows: A adapt =softmax(δ(W·W T )); Among them, A adapt δ is an adaptive adjacency matrix; W is a learnable matrix, initially randomized and continuously optimized through training; δ is the ReLU activation function, used to handle complex graph structure relationships.
4. The multi-regional short-term power load forecasting method based on DEST-GNN according to claim 3, characterized in that, The graph convolution operation is calculated as follows: Among them, H (l+1) Θ is the node feature matrix of the (l+1)th layer; σ is the activation function used to introduce nonlinearity; (k) The Chebyshev polynomial coefficient vector is initialized using Glorot at the start of model training; H (l) is the node feature matrix of the l-th layer; K is the order of the Chebyshev polynomial, which can take any value from 2 to 5; This is the normalized graph Laplace matrix.
5. The multi-regional short-term power load forecasting method based on DEST-GNN according to claim 4, characterized in that, In step (4), the temporal convolutional network is calculated as follows: O (t) =Conv1D(I (t) ,W conv ,b conv ); Among them, O (t) This is the output of the temporal convolutional network; I (t) W is the input to the temporal convolutional network. conv b is the convolution kernel weight matrix; conv This is the kernel bias term.
6. The multi-regional short-term power load forecasting method based on DEST-GNN according to claim 5, characterized in that, In step (5), the sparse attention mechanism, adaptive graph convolutional network, and temporal convolutional network are integrated to construct a complete load prediction model: in, X is the model's predicted output; A is the model's input; X is the model's predicted output. adapt W is an adaptive adjacency matrix. att W is the learnable weight matrix in the sparse attention mechanism. conv b is the kernel weight matrix in a temporal convolutional network; conv is the kernel bias term in the temporal convolutional network; f is the overall structure and computational logic of the model.
7. The multi-regional short-term power load forecasting method based on DEST-GNN according to claim 1, characterized in that, In step (1), the time information includes date type, holiday flags, and day / night flags; Preprocessing includes missing value imputation, outlier correction, and normalization.
8. The multi-regional short-term power load forecasting method based on DEST-GNN according to claim 1, characterized in that, In step (6), mean square error and mean absolute error are used as loss functions; The evaluation metrics include mean absolute error, root mean square error, and mean absolute percentage error.
9. The multi-regional short-term power load forecasting method based on DEST-GNN according to claim 8, characterized in that, The evaluation indicators must meet the following requirements: MAE ≤ 3.35kW, RMSE ≤ 3.48kW, and MAPE ≤ 3.85%.