Power load prediction method based on multi-scale space-time diagram attention network

Abstract

The invention relates to the technical field of power system load prediction, and discloses a power load prediction method based on a multi-scale space-time diagram attention network, which comprises a space-time load information diagram construction unit, a diagram attention modeling unit and a gating feature fusion unit. According to the power load prediction method based on the multi-scale space-time diagram attention network, the regional space is divided by using cells, a topological graph considering cellular spatial correlation is constructed, a prediction architecture of three-dimensional feature fusion is constructed, spatial topological correlation is analyzed through the diagram attention network, a load feature deep mode is mined through a one-dimensional convolutional network, and the power load prediction method based on the multi-scale space-time diagram attention network is realized. The bidirectional gating cycle unit captures dynamic characteristics of a time sequence, and a gating mechanism performs multi-scale feature weighted fusion, so that complex space-time dependence can be better modeled, the generalization ability of the model is improved, the method adapts to a high-frequency fluctuation and multi-source heterogeneous data environment, the accuracy of power load prediction is improved, and the power load prediction efficiency is improved. And powerful support is provided for building a clean, low-carbon, safe and efficient modernized electric power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system load prediction, in particular to a power load prediction method based on a multi-scale spatio-temporal graph attention network. BACKGROUND

[0002] Under the top-level design framework of strategic objectives, a new energy system with renewable energy as the core is being accelerated to build, and power system load prediction plays an important foundational role in the process of building a new power system. Accurate prediction of power load is particularly important for the construction of a clean, low-carbon, safe, and efficient modern power system.

[0003] However, the current mainstream prediction methods have obvious shortcomings in modeling complex spatio-temporal dependencies and improving model generalization capabilities, making it difficult to adapt to high-frequency fluctuations and multi-source heterogeneous data environments.

[0004] Therefore, it is necessary to propose a power load prediction method based on a multi-scale spatio-temporal graph attention network. SUMMARY

[0005] In view of the shortcomings of the prior art, the present application provides a power load prediction method based on a multi-scale spatio-temporal graph attention network, which can better model complex spatio-temporal dependencies, improve model generalization capabilities, adapt to high-frequency fluctuations and multi-source heterogeneous data environments, and improve the accuracy of power load prediction, solving the problems raised in the background art.

[0006] The present application provides the following technical solution: a power load prediction method based on a multi-scale spatio-temporal graph attention network, comprising: spatio-temporal load information graph construction, graph attention modeling, and a gated feature fusion unit. Step one: spatio-temporal load information graph construction: using a point-line graph structure, nodes and edges are used to represent the feature information of nodes and the relationship between nodes, a neighbor matrix construction method based on geographical adjacency relationship is proposed, only determining whether the cell has a spatial adjacent attribute without pre-assigning weights, and the real spatial correlation features are mined through model self-learning; Step two: graph attention modeling: the input of the graph attention layer is a set of node feature vectors, and the output is a new set of node feature vectors after processing, the attention coefficients are calculated and normalized, and then all the attention coefficients are weighted and summed; Step three: gated feature fusion unit: output features through one-dimensional convolution, extract time series information using GRU neural network and bidirectional gated recurrent unit, linearly project the 24-hour and 72-hour spatio-temporal features extracted by GAT, 1DCNN, and BiGRU modules to the same dimension, and then fuse them through the gated fusion unit.

[0007] Preferably, in step one, an adjacency matrix is ​​constructed based on geospatial proximity, with adjacent cells assigned a value of 1 and non-adjacent cells assigned a value of 0. The contiguous connection matrix is ​​set using the following formula: .

[0008] Preferably, in step two: the formula for a node feature vector set is: , The formula for the new node feature vector set is: .

[0009] Preferably, in step two, the formula for calculating the attention coefficient is: , In the formula: For nodes For nodes Attention coefficient The attention function maps the concatenated high-dimensional features to a real number. and They are nodes and The input feature vector; This is the shared parameter matrix for linear transformations; nodes The total number of first-order neighboring nodes is given by the formula:

[0010] In the formula: It is an activation function. It determines the slope of the function when the input is negative; here we take... =0.01, As the independent variable; It is a normalization function; Attention coefficient at the junction Attention weights after normalization; It is a node The For each contact point, the formula for summing all attention coefficients by feature weight is: In the formula: It is the sigmoid activation function.

[0011] Preferably, in step three: one-dimensional convolution output features The formula is: The calculation formula for the GRU neural network is: ; ; ; .

[0012] Preferably, the hidden state feature vector fusion of the bidirectional gated loop unit in step three is formally represented as follows: .

[0013] Preferably, in step three, the spatiotemporal characteristics of 24 hours and 72 hours are expressed by formulas as follows: and Due to its feature dimensions and The differences exist, so the formula for linearly projecting the spatiotemporal features of 24 hours and 72 hours is: , .

[0014] Preferably, in step three, the formula for the fusion feature of the gated fusion unit is: .

[0015] Preferably, the method further includes: executing the solution process: collecting basic information, dividing into cells, processing data, constructing a spatiotemporal information graph, using the STGAT model for prediction, obtaining the prediction result through feature fusion and a fully connected layer, and conducting experiments to verify the model performance.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This power load forecasting method based on multi-scale spatiotemporal graph attention networks utilizes cells to divide the spatial region, constructs a topological graph considering the spatial correlation of cells, and builds a three-dimensional feature fusion prediction architecture. It uses graph attention networks to parse spatial topological relationships, one-dimensional convolutional networks to mine deep patterns of load features, bidirectional gated recurrent units to capture dynamic characteristics of time series, and a gating mechanism to perform multi-scale feature weighted fusion. This approach can better model complex spatiotemporal dependencies, improve model generalization ability, adapt to high-frequency fluctuations and multi-source heterogeneous data environments, and enhance the accuracy of power load forecasting, providing strong support for building a clean, low-carbon, safe, and efficient modern power system. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the solution process of this invention. Detailed Implementation

[0019] 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.

[0020] Please see Figure 1 A power load forecasting method based on a multi-scale spatiotemporal graph attention network includes: spatiotemporal load information graph construction, graph attention modeling, and gated feature fusion unit; Step 1: Spatiotemporal load information graph construction: A dot-line graph structure is adopted, and nodes and edges are used to represent the feature information of nodes and the relationship between nodes. An adjacency matrix construction method based on geographic adjacency is proposed. Only the spatial adjacency attribute of cells is determined without pre-assigning weights. The real spatial correlation features are mined through model autonomous learning. Step 2: Graph Attention Modeling: The input to the graph attention layer is a set of node feature vectors. After processing, a new set of node feature vectors is output. Attention coefficients are calculated and normalized. Then, all attention coefficients are summed using feature weighting. Step 3: Gated Feature Fusion Unit: Features are output through one-dimensional convolution, and temporal information is extracted using a GRU neural network and a bidirectional gated recurrent unit. The 24-hour and 72-hour spatiotemporal features extracted by GAT, 1DCNN, and BiGRU modules are linearly projected to the same dimension and then fused through a gated fusion unit.

[0021] Furthermore, a regional-level cellular topology graph is constructed, and the power load space is modeled. A graph attention network (GAT) is used to mine spatial dimension information. A one-dimensional convolutional neural network (1DCNN) is introduced to enhance the model's ability to extract information in the feature dimension and to increase the dimensionality. A BiGRU layer extracts temporal information in the time dimension. Then, a gating mechanism is used to perform multi-scale feature weighted fusion. The fused state is input into a fully connected network to predict the power load value for the next 24 hours.

[0022] In a preferred embodiment of the present invention, step one involves constructing an adjacency matrix based on geographic spatial adjacency relationships, assigning a value of 1 to adjacent cells and a value of 0 to non-adjacent cells. The contiguous connection matrix is ​​set using the following formula: In the formula: 0 is a zero matrix; 1 is a matrix of all 1s; I is the identity matrix; A3 determines the adjacency matrix based on the specific adjacency situation, taking 1 for adjacent cells and 0 for non-adjacent cells. In step one, the adjacency matrix is ​​constructed based on the geographic spatial adjacency relationship, with adjacent cells assigned a value of 1 and non-adjacent cells assigned a value of 0. The contiguous connection matrix is ​​set using the following formula: In the formula: 0 is the zero matrix; 1 is the all-1 matrix; I is the identity matrix; A3 is the adjacency matrix determined according to the specific adjacency situation, taking 1 for adjacent matrices and 0 otherwise.

[0023] As a preferred embodiment of the present invention, in step two: the formula for a node feature vector set is: , In the formula: N is the number of nodes, F is the number of node features, and the size of matrix t is... , representing the characteristics of all nodes, and R, representing the characteristics of a single node, with a size of . The formula for the new node feature vector set is: , In the formula: This represents the dimension of the feature vector of the new node.

[0024] As a preferred embodiment of the present invention, in step two: the formula for calculating the attention coefficient is: , In the formula: For nodes For nodes Attention coefficient The attention function maps the concatenated high-dimensional features to a real number. and They are nodes and The input feature vector; Z is the shared parameter matrix of the linear transformation; nodes The total number of first-order neighboring nodes is given by the formula: ; In the formula: It is an activation function. It determines the slope of the function when the input is negative; here we take... =0.01, As the independent variable; It is a normalization function; Attention coefficient at the junction Attention weights after normalization; It is a node For the nth contact point, the formula for summing all attention coefficients by feature weight is: In the formula: It is the sigmoid activation function.

[0025] As a preferred embodiment of the present invention, in step three: one-dimensional convolution output features The formula is: In the formula: the one-dimensional convolution input tensor is The convolution kernel matrix is , The offset is calculated using the following formula in the GRU neural network: ; ; ; In the formula To update the gate output; To reset the gate output; This is represented as a candidate hidden state; It is in a hidden state; For activation functions; ' and These are the parameter weights between the input layer and the hidden layer; , and The parameter weights are those between hidden layers; , and This is the offset.

[0026] As a preferred technical solution of the present invention, in the structural design of the bidirectional gated recurrent unit in step three, the forward GRU layer parses the input sequence along the temporal direction to extract historical dependency features, while the reverse GRU layer reconstructs the contextual semantic information through reverse temporal reconstruction. After the two independent networks complete the temporal modeling respectively, the fusion of the hidden state feature vectors of the bidirectional gated recurrent unit is formally represented as follows: In the formula: and These represent the hidden state outputs of the forward and backward GRUs at time t, respectively. This represents the feature fusion operator.

[0027] Furthermore, to fully leverage the complementary advantages of spatiotemporal features at different time scales in power load forecasting, this paper designs a gated feature fusion unit (GFF) to dynamically integrate spatiotemporal joint features extracted from multi-scale branches, thereby enhancing the model's ability to analyze complex dynamics of regional loads. Unlike traditional simple feature splicing methods, the gated feature fusion unit can adaptively adjust the fusion ratio based on the contribution of features at different time scales to load forecasting, effectively avoiding information redundancy and conflicts, and improving the fusion effect.

[0028] Furthermore, to fully leverage the complementary advantages of spatiotemporal features at different time scales in power load forecasting, this paper designs a gated feature fusion unit (GFF) to dynamically integrate spatiotemporal joint features extracted from multi-scale branches, thereby enhancing the model's ability to analyze complex dynamics of regional loads. Unlike traditional simple feature splicing methods, the gated feature fusion unit can adaptively adjust the fusion ratio based on the contribution of features at different time scales to load forecasting, effectively avoiding information redundancy and conflicts, and improving the fusion effect.

[0029] As a preferred embodiment of the present invention, in step three, the spatiotemporal features of 24 hours and 72 hours extracted after passing through the GAT, 1DCNN, and BiGRU modules are expressed by the following formulas: and Due to its feature dimensions and The differences exist, so the formula for linearly projecting the spatiotemporal features of 24 hours and 72 hours is: , In the formula: , These are mapping matrices, , It is the offset; To achieve a unified feature dimension after fusion, It is the sigmoid activation function. and These are the feature information for 24 hours and 72 hours after unifying the dimensions, respectively.

[0030] As a preferred embodiment of the present invention, in step three, the formula for the fusion feature of the gated fusion unit is: In the formula: and These are the weight matrices corresponding to the gating units. and These are the characteristic information of the obtained 24-hour and 72-hour power load forecasts, respectively. The final spatiotemporal joint features are represented by b, where b is the bias term. It is the sigmoid activation function.

[0031] As a preferred technical solution of the present invention, it further includes: executing the solution process: collecting basic information, dividing into cells, processing data, constructing a spatiotemporal information graph, using the STGAT model for prediction, obtaining the prediction result through feature fusion and a fully connected layer, and conducting experiments to verify the model performance.

[0032] Basic information includes meteorological and population information, geographic information, and historical load data of the power supply area; data processing includes handling missing and outlier values, and normalization processing; The data collection frequency (e.g., hourly, daily), data accuracy (e.g., temperature retained to one decimal place), and acquisition channels (e.g., meteorological bureau public API, local monitoring stations) for meteorological information (such as temperature, humidity, precipitation, wind speed, etc.); the statistical granularity (e.g., street level, community level), update cycle (e.g., monthly / annual), and matching rules with power supply areas (e.g., cell association by administrative division) for population information; and the time granularity (e.g., 15 minutes / 30 minutes / 1 hour), metering point coverage (e.g., transformer area, substation), and data storage format (e.g., CSV, database table structure) for historical load data. Missing values: Distinguish between short-term missing values ​​(e.g., 1-3 consecutive time steps) and long-term missing values ​​(e.g., more than 24 hours). Short-term missing values ​​are filled using linear interpolation or the mean of nearest neighbor cells, while long-term missing values ​​are filled using a combination of historical data from the same period and data from similar meteorological days. Outliers: Define rules for identifying outliers (e.g., load values ​​exceeding the historical average by ±3 standard deviations, high load occurring during non-electricity consumption periods). Handling methods include replacing with nearby normal values ​​and detecting and correcting based on the isolated forest algorithm. Normalization: Specify the normalization method (such as min-max normalization) to avoid normalization bias caused by extreme values.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A power load forecasting method based on a multi-scale spatiotemporal graph attention network, characterized in that, include: Spatiotemporal load information graph construction, graph attention modeling, and gating feature fusion unit; Step 1: Spatiotemporal load information graph construction: A dot-line graph structure is adopted, and nodes and edges are used to represent the feature information of nodes and the relationship between nodes. An adjacency matrix construction method based on geographic adjacency is proposed. Only the spatial adjacency attribute of cells is determined without pre-assigning weights. The real spatial correlation features are mined through model autonomous learning. Step 2: Graph Attention Modeling: The input to the graph attention layer is a set of node feature vectors. After processing, a new set of node feature vectors is output. Attention coefficients are calculated and normalized. Then, all attention coefficients are summed using feature weighting. Step 3: Gated Feature Fusion Unit: Features are output through one-dimensional convolution, and temporal information is extracted using a GRU neural network and a bidirectional gated recurrent unit. The 24-hour and 72-hour spatiotemporal features extracted by GAT, 1DCNN, and BiGRU modules are linearly projected to the same dimension and then fused through a gated fusion unit.

2. The power load forecasting method based on a multi-scale spatiotemporal graph attention network according to claim 1, characterized in that: In step one, an adjacency matrix is ​​constructed based on geographic spatial adjacency relationships, with adjacent cells assigned a value of 1 and non-adjacent cells assigned a value of 0. The contiguous connection matrix is ​​set using the following formula: .

3. The power load forecasting method based on a multi-scale spatiotemporal graph attention network according to claim 1, characterized in that: In step two: the formula for a node feature vector set is: , The formula for the new node feature vector set is: , .

4. The power load forecasting method based on a multi-scale spatiotemporal graph attention network according to claim 1, characterized in that: In step two, the formula for calculating the attention coefficient is: , In the formula: For nodes For nodes Attention coefficient The attention function maps the concatenated high-dimensional features to a real number. and They are nodes and The input feature vector; This is the shared parameter matrix for linear transformations; nodes The total number of first-order neighboring nodes is given by the formula: ; In the formula: It is an activation function. It determines the slope of the function when the input is negative; here we take... =0.01, As the independent variable; It is a normalization function; Attention coefficient at the junction Attention weights after normalization; It is a node The For each contact point, the formula for summing all attention coefficients by feature weight is: In the formula: It is the sigmoid activation function.

5. The power load forecasting method based on a multi-scale spatiotemporal graph attention network according to claim 1, characterized in that: In step three: one-dimensional convolution outputs features The formula is: , and the calculation formula for the GRU neural network is: ; ; ; .

6. The power load forecasting method based on a multi-scale spatiotemporal graph attention network according to claim 1, characterized in that: The formal representation of the hidden state feature vector fusion of the bidirectional gated loop unit in step three is as follows: .

7. The power load forecasting method based on a multi-scale spatiotemporal graph attention network according to claim 1, characterized in that: In step three, the spatiotemporal characteristics of 24 hours and 72 hours are expressed by formulas as follows: and Due to its feature dimensions and The differences exist, so the formula for linearly projecting the spatiotemporal features of 24 hours and 72 hours is: , .

8. The power load forecasting method based on a multi-scale spatiotemporal graph attention network according to claim 1, characterized in that: In step three, the formula for the fusion features of the gated fusion unit is: .

9. The power load forecasting method based on a multi-scale spatiotemporal graph attention network according to claim 1, characterized in that: The process also includes: executing the solution process: collecting basic information, dividing the data into cells, processing the data, constructing a spatiotemporal information graph, using the STGAT model for prediction, obtaining the prediction results through feature fusion and fully connected layers, and conducting experiments to verify the model performance.

Images