Demand side load regulation and control method and system based on graph neural network model

Abstract

The invention relates to the technical field of load regulation and control, in particular to a demand side load regulation and control method and system based on a graph neural network model, and the method comprises the steps: obtaining continuous load data and continuous meteorological data, and carrying out the splicing processing and cross fusion processing of the obtained continuous load data and continuous meteorological data, so as to obtain a fused feature vector; based on the fused feature vectors, constructing a load connection structure chart containing load nodes of the power demand side users and electrical connection paths, and determining connection weights of the electrical connection paths according to electrical connection closeness among the load nodes; extracting a spatio-temporal feature vector of the load connection structure diagram containing the connection weight based on a graph neural network model; and performing full connection processing on the spatial-temporal feature vector to obtain a load value set of the power demand side user, and calculating to obtain a regulation and control range of the power demand side load, thereby realizing high-precision load regulation and control of supply and demand matching.

Description

Technical Field

[0001] This invention relates to the field of load regulation technology, specifically to a demand-side load regulation method and system based on a graph neural network model. Background Technology

[0002] With the acceleration of the global energy transition, the proportion of renewable energy generation in the power system continues to increase. However, the inherent intermittent and volatile nature of renewable energy generation significantly reduces the regulation capacity on the generation side, making it increasingly difficult to balance the supply and demand of the power system. Against this backdrop, tapping the adjustable potential on the load side and achieving power supply and demand matching through demand-side load regulation has become an inevitable choice to ensure the safe and stable operation of the new power system.

[0003] Enterprise users, as the core entities on the electricity demand side, possess large-scale electricity loads and extensive adjustment space, making them key targets for tapping the adjustment potential of the load side. The adjustable space of user-side load is a core element of power system dispatching, demand response, and energy management; the accuracy of load control range calculation directly impacts the operational efficiency and stability of the power system. However, short-term loads are influenced by a combination of complex factors, including time factors (such as hours, dates, and holidays), meteorological factors (such as temperature, humidity, and sunlight), and the production behavior of enterprise users. Furthermore, the total load and the loads of individual enterprises within the same region are not independent; there is a significant spatial correlation between regional load and the electricity consumption patterns of enterprises within the region. For example, when the overall regional load increases, the loads of most enterprises within the region often exhibit similar trends. Existing methods for calculating load control ranges struggle to fully capture the complex relationships between these diverse influencing factors and the load, and fail to effectively exploit the spatiotemporal characteristics of the load. This leads to significant prediction deviations in demand-side enterprise user loads, ultimately resulting in low accuracy in load control range calculations, which fails to meet the precision requirements of demand-side load control in modern power systems. Summary of the Invention

[0004] The purpose of this invention is to construct a load connection structure diagram based on acquired continuous load data and continuous meteorological data, and to extract spatiotemporal features from the load connection structure diagram using a graph neural network model in order to calculate the control range of the power demand-side load, thereby achieving highly accurate load control for supply and demand matching.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: This invention provides a demand-side load regulation method based on a graph neural network model, comprising: acquiring continuous load data and continuous meteorological data, and performing splicing and cross-fusion processing on the acquired continuous load data and continuous meteorological data to obtain a fused feature vector; constructing a load connection structure diagram containing load nodes of electricity demand-side users and electrical connection paths between load nodes based on the fused feature vector, and determining the connection weights of electrical connection paths according to the electrical connection tightness between load nodes; extracting spatiotemporal feature vectors of the load connection structure diagram containing connection weights based on a graph neural network model; performing full connection processing on the spatiotemporal feature vectors to obtain a set of load values ​​of electricity demand-side users; and calculating the regulation range of electricity demand-side load based on the difference between the power supply capacity of the power supply side and the set of load values ​​of the power demand side.

[0006] Optionally, the connection weights of electrical connection paths can be determined by: determining the number of non-tethering switches and the number of tethering switches on the electrical connection paths in the load connection structure diagram; determining the cumulative weight value of the electrical connection paths based on the number of non-tethering switches and the number of tethering switches; determining the connection weight of the corresponding path based on the cumulative weight value of each electrical connection path, and thus determining the connection weights of all electrical connection paths in the load connection structure diagram.

[0007] Optionally, the weight accumulation value of the electrical connection path is determined by the following method: setting an initial weight accumulation value for the electrical connection path; decreasing the initial weight accumulation value of the electrical connection path by a first weight value based on the number of non-tethering switches on the electrical connection path, and decreasing the initial weight accumulation value of the electrical connection path by a second weight value based on the number of tie switches on the electrical connection path, to obtain the decreased weight accumulation value of the electrical connection path; the first weight value and the second weight value are in a preset ratio, and the second weight value is greater than the first weight value.

[0008] The graph neural network model consists of a GCN layer, a GRU layer, and a concatenation layer connected in sequence. The spatiotemporal feature vector is obtained in the following ways: the GCN layer performs convolution processing on the load connection structure diagram based on the connection weights and activation functions of the electrical connection paths to obtain spatial features; the GRU layer is used to extract temporal features from the spatial features output by the GCN layer and the fused feature vector to obtain temporal features; the concatenation layer is used to concatenate the output spatial features and temporal features to obtain the spatiotemporal feature vector.

[0009] Optionally, continuous load data and continuous meteorological data can be obtained by: acquiring discrete load data and discrete meteorological data of electricity demand-side users, and using the least squares method to fit the discrete load data and discrete meteorological data based on a preset period to obtain continuous load data and continuous meteorological data.

[0010] Optionally, discrete load data and discrete meteorological data can be obtained in the following ways: Select electricity demand-side users, obtain the load data corresponding to the electricity demand-side users from the power grid database, and obtain the meteorological data corresponding to the electricity demand-side users from the meteorological database; perform time alignment processing on the load data and meteorological data based on the load acquisition time and meteorological acquisition time to obtain time-aligned discrete load data and discrete meteorological data; the time alignment processing is used to align the acquisition time of the load data and meteorological data.

[0011] Optionally, before performing splicing and cross-fusion processing on the continuous load data and continuous meteorological data, the following steps are also included: performing quantity alignment and normalization processing on the continuous load data and continuous meteorological data to obtain the processed continuous load data and continuous meteorological data; the quantity alignment processing is used to align the sample quantity and sample points of the continuous load data and continuous meteorological data.

[0012] Optionally, the discrete load data and discrete meteorological data can be fitted in the following ways: a preset period is set, and the preset period is divided into several preset time periods; for the discrete load data within the preset time period, the first piecewise linear expression of the load data within that time period is obtained by fitting using the least squares method, and the first piecewise linear expression is determined by the slope and intercept of the discrete load data within that time period; for the discrete meteorological data within the preset time period, the second piecewise linear expression of the meteorological data within that time period is obtained by fitting using the least squares method, and the second piecewise linear expression is determined by the slope and intercept of the discrete meteorological data within that time period.

[0013] This invention also provides a demand-side load control system based on a graph neural network model, comprising a data feature fusion module, a spatiotemporal feature extraction module, and an output module. The data feature fusion module acquires discrete load data and discrete meteorological data from electricity demand-side users, and uses the least squares method to fit the discrete load data and discrete meteorological data based on a preset period to obtain continuous load data and continuous meteorological data. It then performs splicing and cross-fusion processing on the continuous load data and continuous meteorological data to obtain a fused feature vector. The spatiotemporal feature extraction module constructs a load connection structure diagram containing load nodes of electricity demand-side users and electrical connection paths between load nodes based on the fused feature vector, and determines the connection weights of the electrical connection paths according to the electrical connection tightness between load nodes. It extracts the spatiotemporal feature vector of the load connection structure diagram containing the connection weights based on a graph neural network model. The output module performs fully connected processing on the spatiotemporal feature vector to obtain a set of load values ​​for electricity demand-side users. Based on the difference between the power supply capacity of the power supply side and the set of load values ​​on the power demand side, it calculates the control range of the electricity demand-side load.

[0014] The present invention also provides a storage medium storing a computer program, which is invoked and executed by a computer to implement the method described above.

[0015] The beneficial effects of this invention are as follows: 1. By constructing a load connection structure diagram with user loads as nodes, electrical connection paths as edges, and the number of non-tethered switches and the number of tethered switches as weights, the actual electrical connections between users on the power demand side are accurately reproduced, and the spatial coupling relationship of node loads can be captured more accurately. By determining the path weights through the weight decreasing rule of the number of non-tethered switches and the number of tethered switches, the electrical tightness between users is accurately quantified, and a control range that conforms to the actual topology of the power grid is obtained.

[0016] 2. By using techniques such as time-of-collection alignment, sample number alignment, and least squares time-segmented fitting, discrete load data and discrete meteorological data are transformed into continuous and matched feature vectors, avoiding subsequent analysis errors caused by data misalignment. Furthermore, normalization processing unifies the data dimensions and improves the stability of data feature fusion and graph neural network model output.

[0017] 3. The graph neural network model consists of a GCN layer and a GRU layer connected in sequence. The GCN layer is used to extract the spatial features of electrical correlation between users, and the GRU layer is used to capture temporal features, achieving full coverage of features in both spatial and temporal dimensions. Compared with a single-dimensional model, the calculation results of the load control range of electricity demand-side users are more in line with the actual electricity consumption patterns.

[0018] 4. The control range is calculated based on the difference between the power supply capacity on the power supply side and the load value set on the power demand side. At the same time, the actual electrical connection relationship of the user's power consumption node is combined. The resulting control range not only meets the capacity constraints of the power supply side, but also takes into account the rationality of the user's power consumption, thus improving the feasibility of demand-side load control. Attached Figure Description

[0019] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0020] Figure 1 This is a flowchart of a demand-side load control method based on a graph neural network model in this invention; Figure 2 This is a schematic diagram illustrating the steps of a demand-side load control method based on a graph neural network model in this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0022] As one implementation method, such as Figure 1 As shown, this invention provides a demand-side load regulation method based on a graph neural network model. Utilizing the ability of graph neural networks to effectively process sequential topological data, it mines the spatial correlations and temporal fluctuation patterns between user production loads. Through precise understanding of user production load patterns, it further achieves precise regulation of user-side load resources. The method includes: S1. Obtain discrete load data and discrete meteorological data from electricity demand-side users, and use the least squares method to fit the discrete load data and discrete meteorological data based on a preset period to obtain continuous load data and continuous meteorological data.

[0023] Furthermore, S11, discrete load data and discrete meteorological data are obtained in the following ways: Select electricity demand-side users, obtain the load data corresponding to the electricity demand-side users from the power grid database, and obtain the meteorological data corresponding to the electricity demand-side users from the meteorological database; perform time alignment processing on the load data and meteorological data based on the load acquisition time and meteorological acquisition time to obtain time-aligned discrete load data and discrete meteorological data; the time alignment processing is used to align the acquisition time of the load data and meteorological data.

[0024] S12. Fit discrete load data and discrete meteorological data in the following ways: Set a preset period and divide the preset period into several preset time periods; for discrete load data within the preset time periods, fit the first piecewise linear expression of the load data within the time period using the least squares method, the first piecewise linear expression being determined by the slope and intercept of the discrete load data within the time period; for discrete meteorological data within the preset time periods, fit the second piecewise linear expression of the meteorological data within the time period using the least squares method, the second piecewise linear expression being determined by the slope and intercept of the discrete meteorological data within the time period.

[0025] S2. Perform splicing and cross-fusion processing on continuous load data and continuous meteorological data to obtain the fused feature vector.

[0026] S21. Before performing splicing and cross-fusion processing on continuous load data and continuous meteorological data, the following steps are also included: performing quantity alignment and normalization processing on continuous load data and continuous meteorological data to obtain processed continuous load data and continuous meteorological data; quantity alignment processing is used to align the sample quantity and sample points of continuous load data and continuous meteorological data.

[0027] S3. Based on the fused feature vectors, construct a load connection structure diagram that includes load nodes of power demand-side users and electrical connection paths between load nodes, and determine the connection weight of the electrical connection path according to the electrical connection tightness between load nodes.

[0028] The connection weights of electrical connection paths are determined as follows: the number of non-tethering switches and the number of tethering switches on the electrical connection paths in the load connection structure diagram are determined; the cumulative weight value of the electrical connection paths is determined based on the number of non-tethering switches and the number of tethering switches; the connection weight of the corresponding path is determined based on the cumulative weight value of each electrical connection path, and then the connection weights of all electrical connection paths in the load connection structure diagram are determined.

[0029] Furthermore, regarding and For all possible connected paths between the two nodes, calculate their cumulative weights and take the maximum value as the final weight between the two nodes. When the cumulative weights of the two types of switches reach the preset maximum value... At this point, the path is the path with the maximum weight in the entire connected graph, where the accumulated weight value... No more than 1.

[0030] The weight accumulation value of the electrical connection path is determined in the following manner: an initial weight accumulation value is set for the electrical connection path; the initial weight accumulation value of the electrical connection path is reduced by a first weight value based on the number of non-tethering switches on the electrical connection path, and the initial weight accumulation value of the electrical connection path is reduced by a second weight value based on the number of tie switches on the electrical connection path, so as to obtain the reduced weight accumulation value of the electrical connection path; the first weight value and the second weight value are in a preset ratio, and the second weight value is greater than the first weight value.

[0031] S4. Extract the spatiotemporal feature vector of the load connection structure graph containing connection weights based on the graph neural network model.

[0032] Furthermore, a graph neural network model is constructed, comprising a GCN layer, a GRU layer, and a concatenation layer connected sequentially. The spatiotemporal feature vector is obtained as follows: the GCN layer performs convolution processing on the load connection structure diagram based on the connection weights and activation functions of the electrical connection paths to obtain spatial features; the GRU layer extracts temporal features from the spatial features output by the GCN layer and the fused feature vector to obtain temporal features; the concatenation layer concatenates the output spatial and temporal features to obtain the spatiotemporal feature vector.

[0033] S5. Perform fully connected processing on the spatiotemporal feature vectors to obtain the load value set of electricity demand-side users. Based on the difference between the power supply capacity on the power supply side and the load value set on the power demand side, calculate the control range of the power demand-side load.

[0034] This invention constructs a load connection structure diagram based on collected load data and meteorological data, and extracts spatiotemporal features from the load connection structure diagram using a graph neural network model to calculate the control range of the power demand-side load, thereby achieving highly accurate load control for supply and demand matching.

[0035] As one implementation method, electricity demand-side users are selected. Electricity demand-side users can be a set of enterprise users in the target area, represented by US=[U1,U2,...,Un]. Load (power) data for each user is obtained from the power grid database, as shown in Table 1 below. The data time span is 1 year, and the number of users is 200.

[0036] Table 1. Load-based information table for a type of electricity demand-side user Referring to Table 2 below, the data are labeled according to the time characteristics of the load. Then, meteorological labels are added to the data based on the time labels and meteorological data. The labeled data is shown in Table 2. Time alignment is also needed to ensure that the collection time of the load and temperature data is consistent. Since the collection period and collection time of meteorological data and load data are different, data alignment needs to take into account the differences in data collection time.

[0037] Table 2. Load and temperature alignment results for a type of electricity demand-side user. The least squares method was used to fit the discrete load data, with one day as a cycle, and the data collected within one day were fitted. As shown in Table 1, a total of n load data points and m meteorological data points were collected on March 11, 2025. According to the load fluctuation pattern, the day was divided into four time periods: [0-6], [6-12], [12-18], and [18-24]. Taking the load data fitting within the time period [0-6] as an example, the load data within this time period was fitted into a broken line, and the slope of the broken line was... and intercept As shown below.

[0038] In the formula, It is a moment The corresponding load value, where 'a' is the number of load data collected within the time period [0-6], then the piecewise linear expression for the load data within this time period is: .

[0039] Based on the changing characteristics of meteorological data, a day is divided into four time periods: [21-7], [7-11], [11-16], and [16-21]. The number of meteorological data collected in each of these four time periods is as follows: Taking the fitting of meteorological data for a specific time period as an example, the meteorological data for that period is fitted into a broken line, and the slope of the broken line is... and intercept As shown below.

[0040] The line graph representation of the meteorological data for that period is: .

[0041] After linear fitting, the load data and meteorological data for each time period become continuous data. Taking a uniform data collection quantity of n, the two types of data are aligned. The alignment results are shown in Table 2. The combinations of load data and meteorological data for times 1 to n are: [Load1,Tmp1], [Load2,Tmp2], [Load...] n ,Tmp n ].

[0042] Continuous load data and continuous meteorological data are standardized, using the load data [Load1, Load2, ..., Load] for March 11, 2025. n Let's take [the example] as an example for processing.

[0043] In the formula, It is [Load1, Load2, ..., Load] n The mean of ] It is a set [Load1, Load2, ..., Load n The standard deviation of ] yes The corresponding value after normalization.

[0044] Concatenate all standardized continuous features to obtain the input vector. .

[0045] Different types of features can be cross-fused using a multilayer perceptron (MLP). The input vector is placed in the input layer of the MLP. The model outputs a fused multivariate feature vector.

[0046] To calculate the spatial correlation of loads, a connection structure diagram between loads needs to be constructed. Each load location is designated as a node, and node characteristics are derived from historical load data and fused multivariate features of the region. The edge weights between nodes are determined based on the tightness of the electrical connection between loads. If a tie switch exists in the connection path between two load nodes, the weight value decreases. If a non-handling switch exists, the weight value decreases. . and There is a linear relationship between the values ​​of , that is Solve The value can be used to obtain the weight of all connecting edges in the entire topology graph.

[0047] In the formula, For nodes and The connecting path between them, which contains A non-tethering switch, One contact switch. Parameters and These are the decreasing weight values ​​corresponding to non-tethering switches and tethering switches, respectively. When the cumulative weight value of the two types of switches reaches the preset maximum value... At this point, the path is the path with the maximum weight in the entire connected graph. The accumulated weight value... No more than 1.

[0048] In conclusion After determining the value, the weights of the connecting edges of all nodes are assigned to obtain weighted graph structure data.

[0049] A Graph Convolutional Network (GCN) is used to process graph-structured data and extract spatial features. For the node feature matrix at time t... The output after graph convolution is: in, The normalized adjacency matrix, The graph convolution weight matrix, The activation function can be ReLU.

[0050] The spatial features output from the Graph Convolutional Network (GCN) layer are input into the paper-based GRU layer to capture temporal correlations. The output of the GRU layer is: in, This represents the hidden state from the previous time step. The output of the multi-feature fusion module... Output of spatiotemporal graph convolution unit The vectors are then concatenated to obtain the final spatiotemporal feature vector. .

[0051] Using fully connected layers to process spatiotemporal feature vectors Process the data and output the load value at the corresponding time: In the formula, Indicates a fully connected layer. This indicates the output value.

[0052] .

[0053] Get the moment The set of load values ​​for all enterprise customers is used to calculate the adjustable space of the user-side load, i.e., the upper and lower limits of the user's load resources.

[0054] As one implementation method, refer to Figure 2This invention provides a schematic diagram of the steps of a demand-side load control method based on a graph neural network model. The first stage is data feature fusion: A set of enterprise users within the target area is selected, and user load data is obtained from the power grid database and corresponding meteorological data from the meteorological database. Time markers are added to the data based on the temporal characteristics of the load, and then meteorological markers are added to the data based on the time markers combined with the meteorological data. The load data is fitted using the least squares method, with a one-day period, fitting the data collected within one day. The meteorological data is also fitted using the least squares method, with a one-day period, fitting the data collected within one day. The load data and meteorological data are aligned on the same time scale, and then normalized. The standardized continuous features are concatenated to obtain the model input vector x.

[0055] Phase 2: Spatiotemporal Feature Extraction: Based on the electrical connection tightness between loads, the connection weights between load nodes are determined, and a load connection structure graph is constructed (with users as nodes, electrical connection paths as edges, and edge weights corresponding to connection tightness). The graph structure data is processed using a GCN (Graph Convolutional Network) to extract the spatial features of electrical associations between users; the spatial features output from the graph convolutional layer are input into a GRU (Gated Recurrent Unit) to capture the temporal correlation of the data.

[0056] The third stage: Output load prediction: By processing the spatiotemporal feature vector through a fully connected layer, the load value at the target time is output, thereby obtaining the load value set of all enterprise users. Based on the difference between the power supply capacity on the power supply side and the predicted load value set, the adjustable space of the user-side load is finally obtained, realizing load regulation and improving the accuracy of load regulation.

[0057] This invention regulates load resources on the power demand side in three stages, improving the applicability of graph neural network models in power demand-side response resource regulation applications. In the data feature fusion stage, the complex correlations between time factors, meteorological factors, and enterprise user loads are captured through identifier addition and data fitting. In the spatiotemporal feature extraction stage, the spatial and temporal correlations of the load are extracted based on the graph convolutional network (GCN) and gated recurrent unit (GRU) in the graph neural network model. During model output, the extracted spatiotemporal features are integrated, and the calculated production load results are output through a fully connected layer. Based on the load calculation results, the method of this invention calculates the regulation range of user-side controllable resources, i.e., the upper and lower limits of regulation, providing data support for improving the flexibility of user-side resource regulation.

[0058] As one implementation, the present invention also provides a demand-side load control system based on a graph neural network model, including a data feature fusion module, a spatiotemporal feature extraction module, and an output module. The data feature fusion module acquires discrete load data and discrete meteorological data from electricity demand-side users, and uses the least squares method to fit the discrete load data and discrete meteorological data based on a preset period to obtain continuous load data and continuous meteorological data. It then performs splicing and cross-fusion processing on the continuous load data and continuous meteorological data to obtain a fused feature vector. The spatiotemporal feature extraction module constructs a load connection structure diagram containing load nodes of electricity demand-side users and electrical connection paths between load nodes based on the fused feature vector, and determines the connection weights of the electrical connection paths according to the electrical connection tightness between load nodes. It extracts the spatiotemporal feature vector of the load connection structure diagram containing connection weights based on a graph neural network model. The output module performs fully connected processing on the spatiotemporal feature vector to obtain a set of load values ​​for electricity demand-side users. Based on the difference between the power supply capacity of the power supply side and the set of load values ​​on the power demand side, it calculates the control range of the electricity demand-side load.

[0059] The present invention also provides a terminal device, including a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method described above.

[0060] The present invention also provides a storage medium storing a computer program, which is invoked and executed by a computer to implement the method described above.

[0061] Compared with the prior art, the present invention has the following beneficial effects based on the above embodiments: 1. By constructing a load connection structure diagram with user loads as nodes, electrical connection paths as edges, and the number of non-tethered switches and the number of tethered switches as weights, the actual electrical connections between users on the demand side of the power grid are accurately reproduced, enabling more precise capture of the spatial coupling relationship between node loads. Path weights are determined by a decreasing weighting rule for the number of non-tethered switches and the number of tethered switches, accurately quantifying the electrical tightness between users and obtaining a control range that closely matches the actual topology of the power grid.

[0062] 2. By using techniques such as time-of-collection alignment, sample number alignment, and least squares time-segmented fitting, discrete load data and discrete meteorological data are transformed into continuous and matched feature vectors, avoiding subsequent analysis errors caused by data misalignment. Furthermore, normalization processing unifies the data dimensions and improves the stability of data feature fusion and graph neural network model output.

[0063] 3. The graph neural network model consists of a GCN layer and a GRU layer connected in sequence. The GCN layer is used to extract the spatial features of electrical correlation between users, and the GRU layer is used to capture temporal features, achieving full coverage of features in both spatial and temporal dimensions. Compared with a single-dimensional model, the calculation results of the load control range of electricity demand-side users are more in line with the actual electricity consumption patterns.

[0064] 4. The control range is calculated based on the difference between the power supply capacity on the power supply side and the load value set on the power demand side. At the same time, the actual electrical connection relationship of the user's power consumption node is combined. The resulting control range not only meets the capacity constraints of the power supply side, but also takes into account the rationality of the user's power consumption, thus improving the feasibility of demand-side load control.

[0065] The specific embodiments described above are preferred embodiments of a demand-side load control method and system based on a graph neural network model according to this application. They are not intended to limit the specific scope of this application. The scope of this application includes but is not limited to these specific embodiments. All equivalent changes made in accordance with the shape and structure of this application are within the protection scope of this application.

Claims

1. A demand-side load regulation method based on a graph neural network model, characterized in that, include: Continuous load data and continuous meteorological data are acquired, and the acquired continuous load data and continuous meteorological data are spliced ​​and cross-fused to obtain the fused feature vector; Based on the fused feature vectors, a load connection structure diagram is constructed that includes load nodes of electricity demand-side users and electrical connection paths between load nodes, and the connection weight of the electrical connection path is determined according to the electrical connection tightness between load nodes. Spatiotemporal feature vectors of the load connection structure graph containing connection weights are extracted based on a graph neural network model; The spatiotemporal feature vectors are processed by a fully connected method to obtain the load value set of electricity demand-side users; Based on the differences between the power supply capacity on the power supply side and the load value set on the power demand side, the control range of the power demand side load is calculated.

2. The demand-side load regulation method based on a graph neural network model according to claim 1, characterized in that, The connection weights of electrical connection paths are determined as follows: the number of non-tethering switches and the number of tethering switches on the electrical connection paths in the load connection structure diagram are determined; the cumulative weight value of the electrical connection paths is determined based on the number of non-tethering switches and the number of tethering switches; the connection weight of the corresponding path is determined based on the cumulative weight value of each electrical connection path, and then the connection weights of all electrical connection paths in the load connection structure diagram are determined.

3. The demand-side load regulation method based on a graph neural network model according to claim 2, characterized in that, The weight accumulation value of the electrical connection path is determined in the following manner: an initial weight accumulation value is set for the electrical connection path; the initial weight accumulation value of the electrical connection path is reduced by a first weight value based on the number of non-tethering switches on the electrical connection path, and the initial weight accumulation value of the electrical connection path is reduced by a second weight value based on the number of tie switches on the electrical connection path, so as to obtain the reduced weight accumulation value of the electrical connection path; the first weight value and the second weight value are in a preset ratio, and the second weight value is greater than the first weight value.

4. The demand-side load regulation method based on a graph neural network model according to claim 1, characterized in that, The graph neural network model consists of a GCN layer, a GRU layer, and a splicing layer connected in sequence. The spatiotemporal feature vector is obtained in the following way: the GCN layer performs convolution processing on the load connection structure graph based on the connection weights and activation functions of the electrical connection paths to obtain spatial features. The GRU layer is used to perform temporal feature extraction on the spatial features and fused feature vectors output by the GCN layer to obtain temporal features; The splicing layer is used to splice the output spatial features with temporal features to obtain a spatiotemporal feature vector.

5. The demand-side load regulation method based on a graph neural network model according to claim 1, characterized in that, Continuous load data and continuous meteorological data are obtained by acquiring discrete load data and discrete meteorological data from electricity demand-side users, and then using the least squares method to fit the discrete load data and discrete meteorological data based on a preset period to obtain continuous load data and continuous meteorological data.

6. The demand-side load regulation method based on a graph neural network model according to claim 5, characterized in that, Discrete load data and discrete meteorological data are obtained in the following ways: Select electricity demand-side users, obtain the load data corresponding to the electricity demand-side users from the power grid database, and obtain the meteorological data corresponding to the electricity demand-side users from the meteorological database. The load data and meteorological data are time-aligned based on the load acquisition time and meteorological acquisition time to obtain time-aligned discrete load data and discrete meteorological data. Time alignment is used to align the acquisition times of load data and meteorological data.

7. The demand-side load regulation method based on a graph neural network model according to claim 1, characterized in that, Before performing splicing and cross-fusion processing on continuous load data and continuous meteorological data, the following steps are also included: performing quantity alignment and normalization processing on continuous load data and continuous meteorological data to obtain processed continuous load data and continuous meteorological data; the quantity alignment processing is used to align the sample number and sample points of continuous load data and continuous meteorological data.

8. The demand-side load regulation method based on a graph neural network model according to claim 1, characterized in that, Discrete load data and discrete meteorological data are fitted using the following method: a preset period is set, and the preset period is divided into several preset time periods; for the discrete load data within the preset time periods, the first piecewise linear expression of the load data within that time period is obtained by fitting using the least squares method, and the first piecewise linear expression is determined by the slope and intercept of the discrete load data within that time period; for the discrete meteorological data within the preset time periods, the second piecewise linear expression of the meteorological data within that time period is obtained by fitting using the least squares method, and the second piecewise linear expression is determined by the slope and intercept of the discrete meteorological data within that time period.

9. A demand-side load control system based on a graph neural network model, characterized in that, It includes a data feature fusion module, a spatiotemporal feature extraction module, and an output module; The data feature fusion module is used to acquire discrete load data and discrete meteorological data of electricity demand-side users. It uses the least squares method to fit the discrete load data and discrete meteorological data based on a preset period to obtain continuous load data and continuous meteorological data. Continuous load data and continuous meteorological data are spliced ​​and cross-fused to obtain a fused feature vector; The spatiotemporal feature extraction module is used to construct a load connection structure diagram containing load nodes of electricity demand-side users and electrical connection paths between load nodes based on the fused feature vectors, and to determine the connection weight of the electrical connection path according to the electrical connection tightness between load nodes. Spatiotemporal feature vectors of the load connection structure graph containing connection weights are extracted based on a graph neural network model; The output module is used to perform fully connected processing on the spatiotemporal feature vectors to obtain a set of load values ​​for electricity demand-side users; Based on the differences between the power supply capacity on the power supply side and the load value set on the power demand side, the control range of the power demand side load is calculated.

10. A storage medium, characterized in that, The storage medium stores a computer program, which is invoked and executed by a computer to implement the method as described in any one of claims 1 to 8.

Images