Power system multi-site load prediction method and system based on graph attention network
By using a graph attention network-based approach, a graph structure is constructed using Kalman filtering and geographic distance to capture the spatiotemporal correlation of multiple sites. This solves the problems of capturing spatiotemporal correlation, integrating heterogeneous data, and adapting across time ranges in multi-site load forecasting, achieving high-precision and efficient load forecasting.
Patent Information
- Application Number
- CN202511110256.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-11-21
AI Technical Summary
Existing load forecasting methods suffer from insufficient capture of spatiotemporal correlations across multiple sites, difficulty in integrating heterogeneous data, lack of adaptability across time ranges, and scalability bottlenecks in large-scale systems, leading to a decline in forecast accuracy and efficiency.
A graph attention network-based approach is adopted, which uses Kalman filtering to smooth heterogeneous input data, constructs a graph structure based on geographic distance, uses attention networks to capture the spatiotemporal correlation between sites, and performs load prediction through a fully connected neural network.
It improves the accuracy and efficiency of multi-site load forecasting, enhances the adaptability and robustness of the model, adapts to forecasting needs of different scales and time ranges, and optimizes computational efficiency and forecasting accuracy.
Smart Images
Figure CN120996267A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power grid load forecasting technology, specifically relating to a multi-site load forecasting method and system for power systems based on graph attention networks. Background Technology
[0002] With the increasing proportion of renewable energy in the power system and the gradual opening of the electricity market, accurate load forecasting is of great significance for the stable operation and economic dispatch of the power system. However, traditional load forecasting methods have the following limitations:
[0003] (1) Insufficient capture of spatiotemporal correlation across multiple sites:
[0004] Existing methods are mainly based on single-point prediction, which makes it difficult to effectively model the complex spatiotemporal correlations between multiple regions / sites. In particular, when dealing with large-scale distributed energy integration, the prediction accuracy and efficiency drop significantly.
[0005] (2) Difficulty in integrating heterogeneous data:
[0006] Load forecasting requires the integration of heterogeneous data types such as historical load data, weather data (temperature / humidity / air pressure, etc.), and date and time information. Existing methods lack the ability to handle these data flexibly.
[0007] (3) Adaptive deficit across time spans:
[0008] Load forecasting demand covers both short-term (minute- to hourly) and long-term (day- to weekly) forecasts, and existing methods cannot provide a unified adaptive framework.
[0009] (4) Scalability bottleneck of large-scale systems:
[0010] When the number of sites expands to thousands, existing methods suffer from low computational efficiency and high resource consumption. Summary of the Invention
[0011] The purpose of this invention is to overcome the shortcomings of existing methods and provide a method and system for multi-site load forecasting in power systems based on graph attention networks.
[0012] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0013] The first objective of this invention is to provide a multi-site load forecasting method for power systems based on graph attention networks, comprising:
[0014] (1) Collect heterogeneous input data from multiple stations, smooth the collected data using Kalman filtering, normalize and standardize the smoothed data and perform periodic feature transformation, and map the heterogeneous input data to a feature vector of fixed length.
[0015] (2) Construct a graph structure based on the geographical distance of the sites, and define the construction rules of the graph structure and the edge weight calculation rules;
[0016] (3) Use attention networks to capture the spatiotemporal correlation between sites and aggregate the feature vectors of neighboring sites to generate aggregated features for each site;
[0017] (4) Output multi-site load prediction values through a fully connected neural network.
[0018] Furthermore, the heterogeneous input data includes historical load data, weather data, and date and time information;
[0019] Mapping heterogeneous input data to fixed-length feature vectors includes:
[0020] The historical load data is normalized by using the MinMaxScaler function to map it to the range [0,1].
[0021] The weather data is standardized to obtain a vector of fixed length;
[0022] Convert date and time information into periodic features and add additional time information as a supplement;
[0023] Historical output power data, weather data, and date and time information are concatenated into a fixed-length feature vector.
[0024] Furthermore, in step (2), the construction rule for the graph structure is: The edges exist under the following conditions:
[0025]
[0026] in, ε is the set of vertices, corresponding to each load station; ε is the set of edges, representing the connection relationships between stations; It is the set of edge weights; d ij δ is the geographical distance between stations i and j, and δ is the distance threshold. It is the set of nearest neighbors of station i (ensuring that each station is connected to at least (K) nearest neighbors), where δ and K are adjustable to control the sparsity of the graph.
[0027] Furthermore, in step (2), the formula for calculating the edge weight is:
[0028]
[0029] Here, S is the scaling operation used to normalize the weights.
[0030] Furthermore, in step (3), features of neighboring sites are aggregated based on attention scores:
[0031]
[0032] in, Let h be the set of neighbors of station i, σ be the activation function, and h be the set of neighbors of station i. i It is the aggregated feature vector of site i.
[0033] Furthermore, in step (4), the formula for calculating the load forecast value is as follows:
[0034]
[0035] Where H is the feature matrix output by the graph attention network, and P is a fully connected neural network containing a multi-layer perceptron structure. It is the predicted load value vector.
[0036] Another object of the present invention is to provide a multi-site load forecasting system for power systems based on graph attention networks, comprising:
[0037] The data preprocessing module processes and integrates heterogeneous input data, mapping it to a fixed-length feature vector. The heterogeneous input data includes historical output power data, weather data, and date and time information.
[0038] The graph building module is used to build a graph structure based on the geographical distance of sites, ensuring the spatial relevance of each site to other sites;
[0039] The feature aggregation module is used to capture the spatiotemporal correlation between sites using an attention network and aggregate the features of neighboring sites to generate a feature representation for each site.
[0040] The prediction result output module is used to output multi-site load prediction values through a fully connected neural network.
[0041] Another object of the present invention is to provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the multi-site load forecasting method for power systems based on graph attention networks provided by the first object of the present invention.
[0042] Another object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-site load forecasting method for power systems based on graph attention networks provided by the first object of the present invention.
[0043] Another object of the present invention is to provide a server comprising at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the processor to cause the at least one processor to perform the power system multi-site load forecasting method based on graph attention network provided in the first object of the present invention.
[0044] In combination with the above technical solutions, the beneficial effects of the present invention compared with the prior art are as follows:
[0045] This invention utilizes graph attention networks (GAT) to capture the spatiotemporal correlations between multiple load sites, thereby improving prediction accuracy. By constructing a graph structure, geographical distance information is incorporated into the model, enhancing the ability to model the mutual influence between sites.
[0046] This invention effectively removes noise from the original data through Kalman filtering smoothing, improving data quality and providing more reliable input for subsequent model training. It can flexibly handle different data types and resolutions, enhancing the adaptability and robustness of the model.
[0047] This invention constructs a graph structure based on geographical distance, ensuring minimum connectivity and adjustable sparsity for each site. By adjusting the distance threshold and the number of nearest neighbors, the sparsity of the graph is controlled to adapt to systems of different sizes.
[0048] This invention adapts to prediction needs across different time ranges (from minutes to days) by adjusting model parameters and input features, providing high-precision predictions while maintaining good scalability.
[0049] This invention can effectively integrate heterogeneous data, capture the spatiotemporal correlation between sites through graph structure, and adapt to different prediction time ranges. By adjusting the sparsity of the graph and feature selection, prediction accuracy and computational efficiency can be optimized. Attached Figure Description
[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0051] Figure 1 This is a flowchart of a multi-site load forecasting method for power systems based on graph attention networks provided in an embodiment of the present invention;
[0052] Figure 2 This is a flowchart illustrating the principle of the multi-site load forecasting method for power systems based on graph attention networks provided in this embodiment of the invention.
[0053] Figure 3This is an architecture diagram of a power system multi-site load forecasting system based on graph attention network provided in an embodiment of the present invention. Detailed Implementation
[0054] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0055] Example 1:
[0056] like Figure 1 The image shows an embodiment of the power system multi-site load forecasting method based on graph attention networks provided by the present invention, which includes the following steps:
[0057] S1: Collect heterogeneous input data from multiple sites, smooth the collected data using Kalman filtering, normalize and standardize the smoothed data, and perform periodic feature transformation to map the heterogeneous input data to a feature vector of fixed length.
[0058] S2: Construct a graph structure based on the geographical distance of the sites, and define the construction rules and edge weight calculation rules for the graph structure;
[0059] S3: Use attention networks to capture the spatiotemporal correlation between sites, and aggregate the feature vectors of neighboring sites to generate aggregated features for each site;
[0060] S4: Outputs multi-site load prediction values through a fully connected neural network.
[0061] Specifically, step S1 is mainly responsible for mapping heterogeneous input data (such as historical load data, weather data, date and time information, etc.) to a feature vector of fixed length.
[0062] (1) Input data type:
[0063] Historical load data: Load values at past times for each station, with a time precision of 15 minutes. The historical load data is normalized and mapped to the range [0,1] using the MinMaxScaler function.
[0064] Weather data, including temperature, humidity, air pressure, dew point, snow point, etc., is obtained from meteorological data sources and is also standardized.
[0065] Date and time information: Convert date and time information into periodic features (such as using sine and cosine functions to represent hours and dates). In addition, additional time information needs to be added as supplements, such as day of the week, whether it is a weekday, whether it is a public holiday, and whether it is night.
[0066] Mapping heterogeneous input data to fixed-length feature vectors includes:
[0067] 1) Using historical load data from the past n time points, first normalize the historical load data and represent it as a vector, h = [h1, h2, ..., h...]. n ], where h i This represents the historical load value at the i-th time point;
[0068] 2) Standardize the weather data (such as temperature, cloud cover, solar parameters, etc.) to obtain a fixed-length vector, w = [w1, w2, ..., w m ], where w m This represents the standardized value of the m-th weather feature;
[0069] 3) Convert date and time information into periodic features, for example, using sine and cosine functions to represent hours and dates, t = [sin(2πt)]. min / 60),cos(2πt min / 60),sin(2πt hour / 24),cos(2πt hour / 24),sin(2πt day /
[0070] 365),cos(2πt day / 365)], where t min t hour t day These represent the current time in minutes, hours, and days, respectively.
[0071] 4) Concatenate the historical output power data, weather data, and date and time information into a fixed-length feature vector, x = [h; w; t], where h, w, and t represent vectors of historical output power data, weather data, and date and time information, respectively, and x is the final feature vector;
[0072] (2) Kalman filtering smoothing:
[0073] Kalman filtering is used to smooth the original data and remove noise. The formula is:
[0074]
[0075] P k|k &=(IK k H k )P k|k-1
[0076] in: It is a priori estimate of the current state, P k|k-1 It is a priori estimate of covariance, K k It is Kalman gain. It is the posterior estimate of the current state, P k|k It is the posterior estimate of the covariance, F k It is the state transition matrix, B k It is the control input matrix, u k It is the control input, Q k It is the process noise covariance, z k It is a measured value, H k It is a measurement matrix, R k It measures the noise covariance.
[0077] Preferably, in step S2, the construction rule for the graph structure is graph The edges exist under the following conditions:
[0078]
[0079] in, ε is the set of vertices, corresponding to each load station; ε is the set of edges, representing the connection relationships between stations; It is the set of edge weights; d ij δ is the geographical distance between stations i and j, and δ is the distance threshold. It is the set of nearest neighbors of station i (ensuring that each station is connected to at least (K) nearest neighbors), where δ and K are adjustable to control the sparsity of the graph.
[0080] The formula for calculating edge weight is:
[0081]
[0082] Here, S is the scaling operation used to normalize the weights.
[0083] Specifically, step S3 uses GAT to capture the spatiotemporal correlation between sites and calculates the feature representation of each site, which is a new feature vector after aggregating neighbor information through a graph attention network. Neighbor sites are defined by the graph structure in step S2, and the edge weights are determined by geographical distance or other spatiotemporal similarity measures. Therefore, sites with spatiotemporal correlation are considered neighbor sites.
[0084] (1) Attention mechanism calculation:
[0085] For each station i, calculate its attention score with its neighboring station j:
[0086]
[0087] W is a learnable weight matrix used for feature transformation; a is an attention vector used to calculate the attention score; LeakyReLU is the activation function; softmax is used to normalize the attention score into a probability distribution.
[0088] (2) Feature aggregation: Aggregate features of neighboring sites based on attention scores:
[0089]
[0090] in, Let h be the set of neighbors of station i, σ be the activation function, and h be the set of neighbors of station i. i It is the aggregated feature vector of site i, a ij Here, W represents the attention weights, and W is the learnable linear transformation matrix.
[0091] Preferably, in step (4), the formula for calculating the load forecast value is:
[0092]
[0093] Where H is the feature matrix output by the graph attention network, and P is a fully connected neural network containing a multi-layer perceptron structure. It is the predicted load value vector.
[0094] The technical effects of the present invention will be further explained below with reference to specific experiments.
[0095] Experimental setup:
[0096] Dataset: Using actual load data for a certain region, covering 100 sites, with a time range from January 2020 to December 2020, and a time resolution of 15 minutes.
[0097] Evaluation metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE).
[0098] Comparison method:
[0099] 1. ARIMA: A traditional time series method.
[0100] 2. LSTM: Long Short-Term Memory Network.
[0101] 3. GAT (Baseline): Basic Graph Attention Network.
[0102] 4. Proposed Method: The improved GAT framework proposed in this invention.
[0103] Experimental results:
[0104]
[0105] The improved GAT framework proposed in this invention outperforms existing methods on all evaluation metrics, reducing MAE, RMSE, and MAPE by 13.4%, 11.4%, and 17.6%, respectively.
[0106] Example 2:
[0107] like Figure 3 As shown, this embodiment of the invention provides a multi-site load forecasting system for power systems based on graph attention networks, comprising:
[0108] The data preprocessing module processes and integrates heterogeneous input data, mapping it to a fixed-length feature vector. The heterogeneous input data includes historical output power data, weather data, and date and time information.
[0109] The graph building module is used to build a graph structure based on the geographical distance of sites, ensuring the spatial relevance of each site to other sites;
[0110] The feature aggregation module is used to capture the spatiotemporal correlation between sites using an attention network and aggregate the features of neighboring sites to generate a feature representation for each site.
[0111] The prediction result output module is used to output multi-site load prediction values through a fully connected neural network.
[0112] Example 3:
[0113] This invention provides an electronic device, including a processor and a memory storing a computer program. When the processor executes the computer program, it implements the power system multi-site load forecasting method based on graph attention network provided in Embodiment 1 of this invention.
[0114] Example 4:
[0115] The present invention provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the power system multi-site load forecasting method based on graph attention network provided in Embodiment 1 of the present invention.
[0116] Example 5:
[0117] This invention provides a server, including at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the at least one processor. The instructions are executed by the processor to cause the at least one processor to perform the power system multi-site load forecasting method based on graph attention network provided in Embodiment 1 of this invention.
[0118] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in the present invention, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.
[0119] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.
[0120] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A multi-site load forecasting method for power systems based on graph attention networks, the method comprising: (1) Collect heterogeneous input data from multiple stations, smooth the collected data using Kalman filtering, normalize and standardize the smoothed data and perform periodic feature transformation, and map the heterogeneous input data to a feature vector of fixed length. (2) Construct a graph structure based on the geographical distance of the sites, and define the construction rules of the graph structure and the edge weight calculation rules; (3) Use attention networks to capture the spatiotemporal correlation between sites and aggregate the feature vectors of neighboring sites to generate aggregated features for each site; (4) Output multi-site load prediction values through a fully connected neural network.
2. The power system multi-site load forecasting method based on graph attention network according to claim 1, characterized in that, The heterogeneous input data includes historical load data, weather data, and date and time information; Mapping heterogeneous input data to fixed-length feature vectors includes: The historical load data is normalized by using the MinMaxScaler function to map it to the range [0,1]. The weather data is standardized to obtain a vector of fixed length; Convert date and time information into periodic features and add additional time information as a supplement; Historical output power data, weather data, and date and time information are concatenated into a fixed-length feature vector.
3. The power system multi-site load forecasting method based on graph attention network according to claim 1, characterized in that, In step (2), the construction rules for the graph structure are as follows: The edges exist under the following conditions: in, It is a set of vertices, corresponding to each load site; It is a set of edges, representing the connections between stations; It is the set of edge weights; d ij δ is the geographical distance between stations i and j, and δ is the distance threshold. It is the set of nearest neighbors of station i, where δ and K are adjustable to control the sparsity of the graph.
4. The power system multi-site load forecasting method based on graph attention network according to claim 1, characterized in that, In step (2), the formula for calculating edge weights is: Here, S is the scaling operation used to normalize the weights.
5. The power system multi-site load forecasting method based on graph attention network according to claim 1, characterized in that, In step (3), features of neighboring sites are aggregated based on attention scores: in, Let h be the set of neighbors of station i, σ be the activation function, and h be the set of neighbors of station i. i It is the aggregated feature vector of site i.
6. The power system multi-site load forecasting method based on graph attention network according to claim 1, characterized in that, In step (4), the formula for calculating the load forecast value is: Where H is the feature matrix output by the graph attention network, and P is a fully connected neural network containing a multi-layer perceptron structure. It is the predicted load value vector.
7. A multi-site load forecasting system for power systems based on graph attention networks, characterized in that, The system includes: The data preprocessing module processes and integrates heterogeneous input data, mapping it to a fixed-length feature vector. The heterogeneous input data includes historical output power data, weather data, and date and time information. The graph building module is used to build a graph structure based on the geographical distance of sites, ensuring the spatial relevance of each site to other sites; The feature aggregation module is used to capture the spatiotemporal correlation between sites using an attention network and aggregate the features of neighboring sites to generate a feature representation for each site. The prediction result output module is used to output multi-site load prediction values through a fully connected neural network.
8. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the power system multi-site load forecasting method based on graph attention network as described in any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-site load forecasting method for power systems based on graph attention networks as described in any one of claims 1 to 6.
10. A server, characterized in that: The method includes at least one processor and a memory communicatively connected to the processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the processor to cause the at least one processor to perform the power system multi-site load forecasting method based on graph attention networks as described in any one of claims 1 to 6.