Power load forecasting method based on dynamic correlation of industrial user groups

Abstract

The present disclosure belongs to the technical field of power system operation and control, and provides a power load prediction method based on dynamic correlation of industrial user groups, which comprises: obtaining an optimal feature set of each industrial user load based on a maximum correlation minimum redundancy criterion; obtaining a correlation coefficient matrix of each industrial user load based on the optimal feature set; realizing accurate division of the industrial user groups by using a graph convolutional neural network according to the correlation coefficient matrix; capturing the time sequence relationship of the industrial user group load based on the industrial user groups; constructing a new graph structure, inputting the time sequence relationship of the industrial user group load into a dynamic graph learning model, and adaptively learning to generate a dynamic graph adjacency matrix; and training the graph convolutional neural network based on the new graph structure data by using the dynamic graph adjacency matrix, and predicting the load value of each industrial user group. The present disclosure can realize dynamic monitoring and prediction of power consumption behavior.

Description

Technical Field

[0001] This disclosure belongs to the field of power system operation and control technology, and in particular relates to a power load forecasting method based on the dynamic correlation of industrial user groups. Background Technology

[0002] With the continuous advancement of the construction of new power systems and the reform of the power market, new roles and technologies such as virtual power plants and demand response are constantly emerging, energy consumption patterns are gradually becoming more diversified, and the electricity consumption behavior of industrial users is becoming more complex. Their loads are exhibiting more complex and changeable new characteristics and forms, and traditional extensive system-level load forecasting methods are gradually becoming unable to meet the management needs of the power system.

[0003] However, due to the combined effects of industry type, weather conditions, and market electricity prices, industrial user load exhibits significant randomness and volatility. The accuracy of load forecasting for industrial user groups (sets containing multiple users) remains difficult to guarantee, posing a substantial challenge. Furthermore, existing methods for segmenting industrial user groups often employ traditional approaches such as K-means clustering and hierarchical clustering. These methods are typically sensitive to initial and outlier values, have high computational complexity, poor model generalization ability, and fail to adequately consider the varying impacts of different characteristics on the loads of different industrial users. This reduces the accuracy and stability of industrial user group segmentation results, further impacting the accuracy of industrial user group load forecasting.

[0004] Furthermore, most existing industrial user group load forecasting methods focus only on uncovering the temporal correlations of the group's own electricity load sequences, neglecting the potential correlations between different industrial user groups. Few industrial user group load forecasting methods consider the potential correlations between the electricity consumption behaviors of different industrial user groups; those that do typically only consider the overall static correlation between the loads of different industrial user groups. However, the correlations between the loads of different industrial user groups often change over time, thus limiting further improvements in the accuracy of industrial user group load forecasting. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a power load forecasting method based on the dynamic correlation of industrial user groups. This method captures the power consumption correlation among industrial user groups in real time, and further develops refined load forecasting for industrial user groups. This is of great significance for reducing the difficulty of power grid planning and scheduling and ensuring the safe and stable operation of the power system.

[0006] In a first aspect, embodiments of this disclosure provide a power load forecasting method based on the dynamic correlation of industrial user groups, including:

[0007] The optimal feature set for each industrial user load is obtained based on the maximum correlation and minimum redundancy criterion, including: constructing a feature set based on the original feature data; calculating the maximum correlation index between other relevant features and the industrial user load using the feature set; selecting normalized data of two other features with the same sequence length from the feature set; distributing the data points of the normalized data of the two other features in pairs in a 2D space to construct a data set; calculating the minimum redundancy index between other relevant features and the industrial user load based on the mutual information of the data set; obtaining the maximum correlation and minimum redundancy index (mRMR) based on the maximum correlation and minimum redundancy index; the data set that maximizes the maximum correlation and minimum redundancy index (mRMR) is the optimal feature set for the industrial user.

[0008] Based on the optimal feature set, the correlation coefficient matrix of each industrial user's load is obtained;

[0009] Based on the correlation coefficient matrix, a graph convolutional neural network is used to accurately segment industrial user groups.

[0010] Based on the industrial user group, capture the time-series relationship of the industrial user group's load;

[0011] A new graph structure is constructed, and the temporal relationship of the load of each industrial user group is input into the dynamic graph learning model, which adaptively learns and generates a dynamic graph adjacency matrix.

[0012] A graph convolutional neural network based on new graph structure data is trained using a dynamic graph adjacency matrix to predict the load values ​​of various industrial user groups.

[0013] Furthermore,

[0014] The original feature data is normalized to the interval [0, 1] according to the following formula;

[0015] All normalized data are constructed into feature sets according to each industrial user and time series.

[0016] Furthermore,

[0017] Based on the optimal feature set, the correlation coefficient matrix of each industrial user's load is obtained, including:

[0018] Select the non-repeating features contained in the optimal feature set of all industrial user loads, form a new feature set, and determine the correlation coefficient vector;

[0019] The correlation coefficient matrix is ​​obtained by concatenating the new correlation coefficient vectors of all industrial user loads.

[0020] Furthermore,

[0021] Based on the correlation coefficient matrix, a graph convolutional neural network is used to accurately segment industrial user groups, including:

[0022] The K-means algorithm was used to initially cluster the industrial user load, and the number of initial clusters was determined by combining the silhouette coefficient formula, and the initial cluster centers and nearby samples were obtained.

[0023] Based on the initial cluster centers and nearby samples, label the correlation coefficient matrix;

[0024] Construct graph data structures and graph adjacency matrices;

[0025] The labeled correlation coefficient matrix is ​​input into a graph convolutional neural network model based on the graph structure to achieve accurate segmentation of all industrial user groups.

[0026] Furthermore,

[0027] Based on the industrial user group, the time-series relationship of the industrial user group's load is captured, including:

[0028] At each data acquisition time, the normalized load data of each industrial user in each industrial user group are summed to obtain the load of each industrial user group;

[0029] Multilayer dilated causal convolution is used to capture the temporal relationships of the load for each industrial user group.

[0030] Furthermore,

[0031] A new graph structure is constructed, and the temporal relationships of the loads of various industrial user groups are input into a dynamic graph learning model. The model adaptively learns and generates a dynamic graph adjacency matrix, including:

[0032] The loads of each industrial user group are used as nodes in the new graph structure, and the temporal relationships of the loads of each industrial user group are used as the associations between the nodes in the new graph structure.

[0033] In the dynamic graph learning layer, Mahalanobis distance is used to measure the correlation between nodes in each new graph structure, and the covariance matrix in Mahalanobis distance is set as a trainable weight parameter to ensure the feasibility of updating the adjacency matrix of the dynamic graph in real time.

[0034] Secondly, this disclosure also provides a power load forecasting system based on the dynamic correlation of industrial user groups, including an optimal feature set determination module, a correlation coefficient matrix construction module, an industrial user group segmentation module, a time series relationship calculation module, a dynamic graph adjacency matrix generation module, and a load value forecasting module.

[0035] The optimal feature set determination module is used to obtain the optimal feature set for each industrial user load based on the maximum correlation and minimum redundancy criterion. This includes: constructing a feature set based on the original feature data; calculating the maximum correlation index between other relevant features and the industrial user load using the feature set; selecting normalized data of two other features with the same sequence length from the feature set; discretizing the data points in the normalized data of the two other features in pairs in a 2D space to construct a data set; calculating the minimum redundancy index between other relevant features and the industrial user load based on the mutual information of the data set; obtaining the maximum correlation and minimum redundancy (mRMR) index based on the maximum correlation and minimum redundancy index; and determining the data set that maximizes the maximum correlation and minimum redundancy (mRMR) index as the optimal feature set for the industrial user.

[0036] The correlation coefficient matrix construction module is used to obtain the correlation coefficient matrix of each industrial user's load based on the optimal feature set;

[0037] The industrial user group segmentation module is used to accurately segment industrial user groups based on the correlation coefficient matrix using a graph convolutional neural network.

[0038] The time-series relationship calculation module is used to capture the time-series relationship of the load of industrial user groups.

[0039] The dynamic graph adjacency matrix generation module is used to construct a new graph structure. It inputs the time-series relationship of the load of each industrial user group into the dynamic graph learning model and adaptively learns to generate a dynamic graph adjacency matrix.

[0040] The load prediction module is used to train a graph convolutional neural network based on new graph structure data using a dynamic graph adjacency matrix to predict the load values ​​of various industrial user groups.

[0041] Thirdly, based on the same inventive concept, this disclosure also provides an electronic device, including at least one processor and at least one memory electrically connected;

[0042] The memory is electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the power load forecasting method based on the dynamic correlation of industrial user groups as described in any of the preceding claims.

[0043] Fourthly, based on the same inventive concept, this disclosure also provides a computer storage medium on which a computer program is stored;

[0044] When the computer program is executed by the processor, it implements the power load forecasting method based on the dynamic correlation of industrial user groups as described in the preceding one.

[0045] Fifthly, based on the same inventive concept, this disclosure also provides a computer program product, which is stored in at least one storage medium;

[0046] The computer program product includes several instructions to cause at least one computer device to execute the power load forecasting method based on the dynamic correlation of industrial user groups as described in the preceding one.

[0047] Compared with existing technologies, this invention provides a power load forecasting method based on the dynamic correlation of industrial user groups, which has the following beneficial effects:

[0048] The embodiments disclosed herein can adapt to the continuous changes in the predicted scenario, fully consider the differences in the impact of different characteristics on the load of each industrial user and the topological information between different industrial users, and provide more real-time and flexible decision support for industrial controllable loads to participate in demand response. This has important practical significance for improving the ability of industrial users to participate in demand response and optimizing energy management.

[0049] Other features and advantages of this disclosure will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A schematic diagram of a power load forecasting method based on the dynamic correlation of industrial user groups according to an embodiment of the present disclosure is shown;

[0052] Figure 2 A schematic diagram of an electronic device structure according to an embodiment of the present disclosure is shown. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0054] Appendix Figure 1 A flowchart illustrating a power load forecasting method based on the dynamic correlation of industrial user groups according to an embodiment of this disclosure is shown. Figure 1 As shown, the power load forecasting method based on the dynamic correlation of industrial user groups in this disclosure includes:

[0055] S1, based on the maximum correlation and minimum redundancy criterion, obtains the optimal feature set of each industrial user load.

[0056] S11, collect raw feature data.

[0057] The raw characteristic data includes historical data on industrial user load and historical data on other characteristics related to industrial user load. Historical data on other characteristics related to industrial user load includes weather data, electricity prices, etc.

[0058] S12 normalizes the original feature data to construct a feature set.

[0059] First, normalize the original feature data to the interval [0, 1] according to the following formula:

[0060] (1)

[0061] In the formula, X o It is the original feature data; X min and X max These are the original feature data and X o Minimum and maximum values ​​of similar feature data; X yes X o Normalized data, i.e., 0≤ X ≤1.

[0062] Secondly, all the normalized data are constructed into feature sets according to each industrial user and time series.

[0063] S13, using the feature set, calculate the maximum correlation index between other relevant features and industrial user load.

[0064] For each industrial user, the Pearson correlation coefficient is used to represent the correlation between other relevant characteristics and the industrial user's load. The closer the absolute value is to 1, the stronger the correlation; the closer it is to 0, the weaker the correlation. The formula for calculating the Pearson correlation coefficient is as follows:

[0065] (2)

[0066] In the formula, X i For the first feature set i Normalized data for other features; Y Normalized data for industrial user loads in the feature set; f cov ( X i ,Y The covariance is the normalized data for other characteristics and the normalized data for industrial user load. σ ( X i )and σ ( Y ) are the normalized data for other features. X i Normalized data with industrial user load Y Standard deviation; r ( X i ,Y ) represents the Pearson correlation coefficient.

[0067] Further, the maximum correlation index can be obtained. D ( S,Y )for:

[0068] (3)

[0069] In the formula, S and N S These represent the optimal feature set of industrial user loads and the number of other features it contains.

[0070] S14, determine the minimum redundancy index.

[0071] The Maximum Information Coefficient (MIC) is used to measure the redundancy between different features. It adds standardization to mutual information and characterizes the amount of information contained in one random variable from another. The calculation process is as follows:

[0072] First, assume X i and Xj They are respectively the first in the feature set i The and the first j Normalized data for each of the other features, and the sequence length is [missing information]. n .Will X i and X j The data points in the dataset are discretely distributed in pairs in a 2D space, denoted as a data set. D ( X i ,X j Mesh the 2D space, transforming the 2D space... x and y The directions are divided into A Individual and B A range.

[0073] Data set D ( X i ,X j In the grid G ( A,B The calculation method for mutual information is as follows:

[0074] (4)

[0075] In the formula, p ( D a,b ) is a data set D ( X i ,X j The data points in the region () are located in the region () a,b The probability of ). p ( D a )and p ( D b The data points are located in the respective regions. a and b The probability of each region is proportional to the number of data points located in that region.

[0076] Secondly, take the data set D ( X i ,X j In the grid G ( A,B Mutual information under different partitioning methods M ( D,G The maximum value of ) is normalized to obtain:

[0077] (5)

[0078] Thus obtain X i and X j The formula for calculating MIC is:

[0079] (6)

[0080] In the formula, MIC takes the value range of [0,1]. The closer its value is to 1, the greater the correlation between the two data; the closer it is to 0, the smaller the correlation.

[0081] Furthermore, the minimum redundancy index can be obtained as:

[0082] (7).

[0083] S15, obtain the Maximum Relevance Minimum Redundancy (mRMR) index, and determine the optimal feature set.

[0084] Combining equations (3) and (7), the final maximum correlation minimum redundancy (mRMR) index is obtained as follows:

[0085] (8)

[0086] Make I Data set that yields the maximum value D ( X i ,X j This is the optimal feature set for industrial users. S .

[0087] Through the above process, the optimal feature set can be selected for each industrial user load. S .

[0088] S2, based on the optimal feature set, obtains the correlation coefficient matrix of the load for each industrial user.

[0089] S21, Filter out the optimal feature set of all industrial user loads. S Contained in N Each non-repeating feature is used to form a new feature set. F Determine the correlation coefficient vector.

[0090] For each industrial user, the load and their respective optimal feature sets are obtained by combining equation (2). SThe Pearson correlation coefficients among the features included are used to obtain the correlation coefficient vector for each industrial user load. The resulting correlation coefficient vectors are then padded as follows:

[0091] For a specific industrial user, if a new set of features F A certain feature in the load is in the optimal feature set of the industrial user. S In the middle, the corresponding correlation coefficient is filled in at the corresponding position in the correlation coefficient vector. If the new feature set F A certain feature in the load is not in the optimal feature set of the industrial user load. S In the middle, 0 is filled into the corresponding position of the correlation coefficient vector, so that the dimension of each industrial user load is obtained. N The new correlation coefficient vector: , F For the new feature set, Y' For the new feature set F Industrial user load in the middle.

[0092] S22, concatenate the new correlation coefficient vectors of all industrial user loads to obtain a dimension of ( M,N The correlation coefficient matrix H :

[0093] (9).

[0094] In the formula, M N represents the dimension of matrix H.

[0095] S3, based on the correlation coefficient matrix, uses a graph convolutional neural network to accurately segment industrial user groups.

[0096] S31. The K-means algorithm is used to initially cluster the industrial user load, and the initial number of clusters is determined by combining the silhouette coefficient formula, and the initial cluster centers and nearby samples are obtained.

[0097] Specifically, the K-means algorithm is used to classify industrial users based on their load data. The optimal number of clusters k is determined according to the silhouette coefficient, and the industrial users are finally divided into k industrial user groups.

[0098] S32, label the correlation coefficient matrix based on the initial cluster centers and nearby samples.

[0099] In the correlation coefficient matrix, data related to the initial cluster center and nearby samples are used as labeled data, while data unrelated to the initial cluster center and nearby samples are used as unlabeled data.

[0100] S33, construct the graph data structure and the graph adjacency matrix.

[0101] Specifically, each industrial user is treated as a node (1-dimensional element) in the graph data structure, and the Pearson correlation coefficient between the industrial user loads is used as the corresponding 2-dimensional element in the graph adjacency matrix to finely quantify the association between nodes in the graph data structure and consider the topological information between nodes in different graph data structures.

[0102] The Pearson correlation coefficient between industrial user loads is calculated based on the normalized data of industrial user loads at each data collection time.

[0103] S34. The labeled correlation coefficient matrix is ​​input into the graph convolutional neural network (GCNNs) model based on the graph structure in S33 to achieve accurate segmentation of the entire industrial user group.

[0104] S4 captures the time-series relationships of the load of industrial user groups.

[0105] S41, at each data acquisition time, the normalized data of the load of each industrial user in each industrial user group are added together to obtain the load of each industrial user group.

[0106] S42 uses multilayer dilated causal convolution to capture the temporal relationships of the load for each industrial user group.

[0107] First, dilated causal convolution enables... t The output at time +1 is only related to t The input prior to the given time point aligns with the practical logic of analyzing industrial user load data containing time series sequences. Furthermore, it increases the receptive field of the convolution operation by injecting holes into the standard convolution kernel. The formula for calculating the receptive field size of the convolution is:

[0108] (10)

[0109] In the formula, z The size of the original convolution kernel. d It is an expansion factor.

[0110] Then, skip connections are used to concatenate the outputs of each dilated causal convolutional layer to obtain the temporal relationship (causal relationship and long-term dependency relationship) of the industrial user group load, avoiding degradation problems caused by excessive stacking.

[0111] S5. Construct a new graph structure by inputting the temporal relationship of the load of each industrial user group into the Dynamic Graph Learning model, and adaptively learn to generate a dynamic graph adjacency matrix.

[0112] S51, the loads of each industrial user group are used as nodes in the new graph structure, and the temporal relationship of the loads of each industrial user group is used as the association between the nodes in the new graph structure.

[0113] S52 uses Mahalanobis distance to measure the correlation between nodes in each new graph structure in the dynamic graph learning layer, and sets the covariance matrix in Mahalanobis distance as a trainable weight parameter to ensure the feasibility of updating the adjacency matrix of the dynamic graph in real time.

[0114] First, the correlation between nodes in each new graph structure is measured using Mahalanobis distance, and the calculation formula is as follows:

[0115] (11)

[0116] In the formula, x p t and x q t These are the nodes of the new graph structure. p and q exist t The temporal characteristic sequence at each moment, W Let be the covariance matrix.

[0117] Secondly, the covariance matrix in Mahalanobis distance is set as a trainable weight parameter to ensure the feasibility of real-time updating of the adjacency matrix of the dynamic graph. Then, the correlation between nodes in the new graph structure is calculated using Mahalanobis distance, and its calculation formula is as follows:

[0118] (12)

[0119] (13)

[0120] In the formula, G t ( x i t , x j t )for t time x p t and x q t The correlation; σ D Standard deviation; for t The normalized correlation matrix between nodes of each new graph structure at each time step is... t The adjacency matrix of the dynamic graph at time step.

[0121] S6. Finally, a graph convolutional neural network based on the new graph structure data is trained using a dynamic graph adjacency matrix to predict the load values ​​of each industrial user group.

[0122] The embodiments of this disclosure employ a graph convolutional neural network model to finely segment industrial user groups, fully considering the differences in the impact of various characteristics on the load of each industrial user and the topological information among different industrial users. Through end-to-end modeling, a fine and stable segmentation of industrial user groups is achieved. The dynamic graph learning method for industrial user group correlation can capture the electricity consumption correlation between groups in real time, thereby enabling dynamic monitoring and prediction of electricity consumption behavior. This method can adapt to the constantly changing prediction scenarios, providing more real-time and flexible decision support for industrial controllable loads to participate in demand response, which has significant practical implications for improving the ability of industrial users to participate in demand response and optimizing energy management.

[0123] Based on the above method, this disclosure also provides a power load forecasting system based on the dynamic correlation of industrial user groups, corresponding to the above method, including an optimal feature set determination module, a correlation coefficient matrix construction module, an industrial user group segmentation module, a time series relationship calculation module, a dynamic graph adjacency matrix generation module, and a load value forecasting module.

[0124] The optimal feature set determination module is used to obtain the optimal feature set of each industrial user's load based on the maximum correlation and minimum redundancy criterion.

[0125] The correlation coefficient matrix construction module is used to obtain the correlation coefficient matrix of each industrial user's load based on the optimal feature set;

[0126] The industrial user group segmentation module is used to accurately segment industrial user groups based on the correlation coefficient matrix using a graph convolutional neural network.

[0127] The time-series relationship calculation module is used to capture the time-series relationship of the load of industrial user groups.

[0128] The dynamic graph adjacency matrix generation module is used to construct a new graph structure. It inputs the time-series relationship of the load of each industrial user group into the dynamic graph learning model and adaptively learns to generate a dynamic graph adjacency matrix.

[0129] The load prediction module is used to train a graph convolutional neural network based on new graph structure data using a dynamic graph adjacency matrix to predict the load values ​​of various industrial user groups.

[0130] Based on the same inventive concept as the above-disclosed content, this disclosure also provides an electronic device. For example... Figure 2As shown, the electronic device of this disclosure includes at least one processor and at least one memory electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the power load forecasting method based on the dynamic correlation of industrial user groups as described above.

[0131] It should be noted that the electrical connection between the above-mentioned units does not necessarily mean the connection between lines. The indirect connection method can be applied to the embodiments of this disclosure as long as it achieves the purpose of this disclosure.

[0132] Based on the same inventive concept, this disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the power load forecasting method based on the dynamic correlation of industrial user groups as described above.

[0133] Based on the same inventive concept, this disclosure also provides a computer program product stored in at least one storage medium; the computer program product includes several instructions to cause at least one computer device to execute the power load forecasting method based on the dynamic correlation of industrial user groups as described above.

[0134] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A power load prediction method based on dynamic correlation of industrial user groups, characterized in that the method comprises: obtaining an optimal feature set of each industrial user load based on a maximum correlation minimum redundancy criterion, including: constructing a feature set according to original feature data, calculating a maximum correlation index between the industrial user load and other related features using the feature set; selecting normalized data of two other features with the same sequence length in the feature set, discretely distributing the data points in the normalized data of the two other features in pairs in a 2-dimensional space, constructing a data set, calculating a minimum redundancy index between the industrial user load and other related features according to the mutual information of the data set; obtaining a maximum correlation minimum redundancy mRMR index according to the maximum correlation index and the minimum redundancy index; and making the data set that makes the maximum correlation minimum redundancy mRMR index maximum as the optimal feature set of the industrial user; obtaining a correlation coefficient matrix of each industrial user load based on the optimal feature set; realizing accurate division of the industrial user groups by using a graph convolutional neural network according to the correlation coefficient matrix; capturing the time series relationship of the industrial user group load based on the industrial user groups, including: adding the normalized data of each industrial user load in each industrial user group at each data collection time to obtain each industrial user group load; and capturing the time series relationship of each industrial user group load using a multi-layer dilation causal convolution; constructing a new graph structure, inputting the time series relationship of each industrial user group load into a dynamic graph learning model, and adaptively learning to generate a dynamic graph adjacency matrix, including: taking each industrial user group load as a node of the new graph structure, and taking the time series relationship of each industrial user group load as the association between the nodes of the new graph structure; measuring the correlation between the nodes of each new graph structure by Mahalanobis distance in the dynamic graph learning layer, and setting the covariance matrix in the Mahalanobis distance as a trainable weight parameter to ensure the feasibility of real-time updating of the dynamic graph adjacency matrix; training a graph convolutional neural network based on new graph structure data using the dynamic graph adjacency matrix to predict the load value of each industrial user group.

2. The power load prediction method based on dynamic correlation of industrial user groups according to claim 1, characterized in that the original feature data is normalized to construct a feature set, including: normalizing the original feature data to the [0, 1] interval according to the following formula; and constructing the normalized data of all industrial users and time series as a feature set.

3. The power load prediction method based on dynamic correlation of industrial user groups according to claim 1, characterized in that the correlation coefficient matrix of each industrial user load is obtained based on the optimal feature set, including: screening out non-redundant features contained in the optimal feature set of all industrial user loads to form a new feature set, and determining a correlation coefficient vector; and splicing the new correlation coefficient vectors of all industrial users to obtain a correlation coefficient matrix.

4. The power load prediction method based on dynamic correlation of industrial user groups according to claim 1, characterized in that ​ ​ ​ ​ ​ ​ ​ ​ According to the correlation coefficient matrix, the industrial user group is accurately divided by using a graph convolutional neural network, including: Using the K-means algorithm to initially cluster the industrial user load, combining the silhouette coefficient formula to determine the initial cluster number, and obtaining the initial cluster center and the nearby samples; According to the initial cluster center and the nearby samples, the correlation coefficient matrix is labeled; Constructing a graph data structure and a graph adjacency matrix; The labeled correlation coefficient matrix is input into the graph convolutional neural network model based on the graph structure to realize accurate division of all industrial user groups.

5. A power load prediction system based on dynamic correlation of industrial user groups, characterized in that: The system comprises an optimal feature set determination module, a correlation coefficient matrix construction module, an industrial user group division module, a time series relationship calculation module, a dynamic graph adjacency matrix generation module, and a load value prediction module; The optimal feature set determination module is used to obtain the optimal feature set of each industrial user load based on the maximum correlation minimum redundancy criterion, including: constructing a feature set according to original feature data, calculating the maximum correlation index between the related other features and the industrial user load by using the feature set; selecting the normalized data of two other features with the same sequence length in the feature set, and discretely distributing the data points in the normalized data of the two other features in pairs in a 2-dimensional space to construct a data set, calculating the minimum redundancy index between the related other features and the industrial user load according to the mutual information of the data set; obtaining the maximum correlation minimum redundancy mRMR index according to the maximum correlation index and the minimum redundancy index; and making the data set whose maximum correlation minimum redundancy mRMR index reaches the maximum value as the optimal feature set of the industrial user; The correlation coefficient matrix construction module is used to obtain the correlation coefficient matrix of each industrial user load based on the optimal feature set; The industrial user group division module is used to accurately divide the industrial user group by using a graph convolutional neural network according to the correlation coefficient matrix; The time series relationship calculation module is used to capture the time series relationship of the industrial user group load based on the industrial user group, including: adding the normalized data of each industrial user load in each industrial user group at each data acquisition time to obtain each industrial user group load; Using a multi-layer expansion causal convolution to capture the time series relationship of each industrial user group load; The dynamic graph adjacency matrix generation module is used to construct a new graph structure, input the time series relationship of each industrial user group load into a dynamic graph learning model, and adaptively learn to generate a dynamic graph adjacency matrix, including: taking each industrial user group load as a node of the new graph structure, and taking the time series relationship of each industrial user group load as the association between the nodes of the new graph structure; measuring the correlation between the nodes of each new graph structure by Mahalanobis distance in the dynamic graph learning layer, and setting the covariance matrix in the Mahalanobis distance as a trainable weight parameter to ensure the feasibility of real-time updating of the dynamic graph adjacency matrix; The load value prediction module is used to train a graph convolutional neural network based on new graph structure data by using the dynamic graph adjacency matrix to predict the load value of each industrial user group.

6. An electronic device, comprising: at least one processor and at least one memory electrically connected; the memory is electrically connected with the processor, wherein the memory stores instructions executable by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to perform the power load prediction method based on dynamic correlation of industrial user groups according to any one of claims 1-4.

7. A computer readable storage medium, characterized in that, the computer readable storage medium stores a computer program; the computer program is executed by a processor to implement the power load prediction method based on dynamic correlation of industrial user groups according to any one of claims 1-4.

8. A computer program product, characterized in that, the computer program product is stored in at least one storage medium; the computer program product includes a plurality of instructions to enable at least one electronic device to perform the power load prediction method based on dynamic correlation of industrial user groups according to any one of claims 1-4.

Images