Distribution network load forecasting device and method based on hybrid machine learning

Abstract

The present invention is a distribution network load prediction device and method based on hybrid machine learning, in the field of artificial intelligence prediction technology. The device includes a data acquisition module, a data storage module, a data preprocessing module and a prediction module; the method includes the following steps: S1: collecting load data of residential areas in the region; S2: preprocessing the load data of residential areas; S3: calculating node attribute characteristics using Copula entropy and fuzzy membership function; S4: constructing a weight map of the distribution station area; S5: establishing a hybrid machine learning model and loss function, and training using the weight map; S6: using the hybrid machine learning model to predict the load data of the distribution network. The method of the present invention weights, fuses and updates the load data features, which can significantly improve the accuracy, efficiency and generalization ability of the prediction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence prediction technology, and is particularly suitable for distribution network load prediction based on hybrid machine learning. Background Art

[0002] The new power system is a digital power system with new energy as the main body. It forms a safe, efficient, low-carbon and intelligent power system by promoting the integration and comprehensive allocation of power flow, information flow and value flow in various links of power production, transmission, consumption and storage. However, large-scale access to new energy will lead to changes in the structure, load characteristics and operating characteristics of the power system. As an important physical carrier and cornerstone of the new power system, the distribution network improves its operational stability and has great value and significance for the regulation and control decision-making of the new power system. At present, machine learning has been widely used in power equipment anomaly detection, load forecasting, topology identification, vulnerable route identification, online power flow optimization and other aspects and has played a significant role, providing strong support for the safe and stable operation of the new distribution network.

[0003] Load forecasting is an important part of power management. The load forecasting data it provides is extremely important for the control, operation and planning of the power system. Accurately predicting power load data is not only important for determining the operation mode of the power system, but also for determining the optimal dispatch of the power system, the power transmission plan between regions and the load dispatch plan. In addition, the accuracy of power load forecasting also directly affects the safety, reliability, economy and power quality of the power system operation, and is related to the production planning and dispatching operation of the power system. At the same time, in the past decade, the intensification of market competition, the aging of infrastructure and the integration requirements of renewable energy have made load forecasting not only more important, but also more difficult. For a long time, domestic scholars have conducted extensive and in-depth research on the theory of power system load forecasting and proposed many effective methods, such as regression analysis, time series method, neural network method, wavelet analysis method, etc. For a certain prediction problem, multiple prediction methods can be established. Different prediction methods provide different prediction information and different prediction accuracy.

[0004] None of the above prediction methods take into account the volatility of grid load, which is caused by many driving factors, such as different weather, seasonal conditions, holidays, work cycles, and economic fluctuations. The data generated by these factors have subtle time patterns, resulting in poor grid load data prediction accuracy and low robustness. Therefore, high-precision grid load forecasting based on the change law of grid load itself and modeling of its influencing factors is a key technical problem that needs to be solved in power management. Summary of the invention

[0005] In view of this, in order to solve the above technical problems, the present invention provides a distribution network load prediction device and method based on hybrid machine learning, which aims to achieve accurate prediction of power grid load by utilizing the organic combination of machine learning methods such as weighted graph convolutional networks, gated graph neural networks, and multi-layer perceptron prediction.

[0006] In order to achieve the above purpose, the distribution network area includes several distribution stations, each distribution station has several residential communities, and each residential community is equipped with sensors to collect the load data of the entire community in real time. The present invention provides a distribution network load prediction method based on hybrid machine learning, which is implemented by a distribution network load prediction device based on hybrid machine learning; combined with Figure 1 The distribution network load prediction device based on hybrid machine learning is composed of a data acquisition module, a data storage module, a data preprocessing module and a prediction module; the data acquisition module reads the load data of sensors in residential areas within the jurisdiction of the distribution network through communication; the data storage module is composed of a load data storage unit and a prediction data storage unit; the data preprocessing module is composed of a data preprocessing unit connected in series with a feature weighting unit; the load data storage unit is a memory, which is respectively connected to the data acquisition module and the data preprocessing unit, and is used to store and transmit the residential area load data collected by the data acquisition module; the data preprocessing unit is directly connected to the data acquisition module, and is used to read the residential area load data in real time; the prediction data storage unit is a memory, which is connected to the prediction module, and is used to store and transmit the data fed back by the prediction module; the feature weighting unit is connected to the prediction module; the prediction module is a processor with a hybrid machine learning model in the memory, which is used to predict the power supply of the distribution station.

[0007] Combination Figure 2 , characterized in that it comprises the following steps:

[0008] S1: The data acquisition module collects the load data of all residential areas within the distribution network area and stores it in the load data storage unit;

[0009] S2: The data preprocessing unit preprocesses the residential area load data in the load data storage unit within the time period T to obtain a weight coefficient;

[0010] S3: Taking each distribution station area as a unit, the feature weighting unit uses Copula entropy and fuzzy membership function to weight the feature vectors of the corresponding distribution station, the feature vectors of the residential area and other random feature vectors within the time period T to obtain the node attribute features;

[0011] S4: Taking each distribution station area as a unit and the distribution stations and residential areas within the area as nodes, the data preprocessing module constructs a weight graph within the time period T of the corresponding distribution station area using the weight data of step S2 and the node attribute characteristics of step S3;

[0012] S5: The prediction module builds a hybrid machine learning model, sets the loss function, and uses the weight map in the distribution network area in the historical time period for training;

[0013] S6: The prediction module uses the trained hybrid machine learning model to predict the load data of the distribution network at the next moment.

[0014] The load data described in step S1 includes: residential community ID, time, and load amount.

[0015] Furthermore, the preprocessing described in step S2 specifically includes: (1) using the k-nearest neighbor algorithm (KNN) to detect abnormal data and replace the abnormal data with neighboring data; (2) using the min-max normalization method to linearly transform the load so that its value is transformed to the interval [0, 1] to obtain a weight coefficient; (3) One-Hot Encoding the holidays and month information corresponding to the time period.

[0016] Furthermore, the amount of load data collected in the time period T is greater than the output dimension of the weighted graph neural network. Usually, one month is taken as a time period.

[0017] Preferably, the characteristics of the distribution station described in step S3 include: capacity, voltage level, type, geographical location, and grid structure, a total of 5; the characteristics of the residential community include: area, total population, population density, total income of residents, average income of residents, geographical location of the community, electricity price of the community, and load, a total of 8; other random characteristics include: average temperature, humidity, air pressure within a time period, and whether the time period contains holidays, a total of 4.

[0018] Further, the step S3 is specifically as follows:

[0019] S301: constructing characteristic values ​​of a distribution station, characteristic values ​​of a residential area and other random characteristic values ​​within a distribution station area into a characteristic vector according to time sequence;

[0020] S302: Calculate the Copula entropy between each characteristic vector and the power supply vector of the distribution station;

[0021] The power supply vector of the distribution station is the sum of the load characteristic vectors of all residential communities within the jurisdiction of the distribution station.

[0022] Furthermore, for any two n-dimensional feature vectors X = [X1, X2, ..., Xn ] and Y = [Y1, Y2, ..., Y n ], the Copula entropy calculation process of X and Y is as follows: First, we need to obtain the joint probability density function f of X and Y X，Y (x, y), then calculate the Copula function Among them, C(u, v) is the marginal distribution function F X and F Y Function mapped to the interval [0, 1], u = F X (x) and v = F Y (y).

[0023] S303: Calculate the corresponding membership degree using the Copula entropy corresponding to each eigenvector calculated in step S302, and then construct a membership matrix of all eigenvectors within the jurisdiction of the distribution station, which is the node attribute feature.

[0024] For the i-th eigenvector, the Copula entropy c i , the membership degree is calculated as: The membership matrix B of all M features is:

[0025] Furthermore, a weight graph is established for each distribution station area. The weight graph of the distribution station area in the time period T described in step S4 is: the distribution stations and residential areas in the area are nodes, and the node attribute characteristics are the membership matrix; the edges are the relationships between the nodes, and the adjacency matrix A is constructed; the weight of the edge is the power supply of the corresponding distribution station to the residential area, that is, the weight data of step S2.

[0026] Furthermore, the hybrid machine learning model described in step S5 is composed of a weighted graph neural network connected in series with a gated graph neural network and then connected in series with a multi-layer perceptron (MLP). The weighted graph neural network is an effective convolutional neural network that can directly act on the graph, and is used to achieve feature fusion; the gated graph neural network updates the feature information of the node through a message passing mechanism, so that the node can use the information of its neighboring nodes to update itself, and is used to achieve fusion feature update; the multi-layer perceptron is used to achieve prediction of the fused and updated features.

[0027] Furthermore, for a weighted graph of a distribution substation area with N nodes and M-dimensional features, the working principle of the weighted graph convolutional network is as follows:

[0028]

[0029] in, is the feature matrix, which is the matrix composed of all the feature vectors in step S3; IN is the N-order identity matrix; is an approximate graph convolution filter, Weight Matrix The corresponding coefficients are generated by the weight data of step S2; D is the dimension of the output fusion feature X of the weighted graph convolutional network; f(·) is the activation function.

[0030] Furthermore, the gated graph neural network, whose input is the output of the weighted graph convolutional network, works as follows:

[0031]

[0032] in, represents the feature update vector of the area under the jurisdiction of the l-th distribution station i. When l = 0, is the output fusion feature X of the corresponding weighted graph convolutional network; represents the feature update vector of the lth layer corresponding to the distribution station j adjacent to the distribution station i; and is the parameter matrix to be learned; σ(·) is the sigmoid function; represents element-by-element multiplication; [·] represents a vector concatenation operation; N(i) represents the set of neighboring substations of substation i.

[0033] Furthermore, the input of the multilayer perceptron is the output of the gated graph neural network, and its output and output dimension are the same as the number of distribution stations in the distribution network area. Its working principle is:

[0034]

[0035] in, is the output of the jth neuron in the lth layer. When l = 0, is the feature update vector of the j-th distribution station output by the gated graph neural network; n (l-1) is the number of neurons in the previous layer; is the weight to be learned between the jth neuron in the lth layer and the ith neuron in the l-1th layer; is the bias term to be learned; f(·) is the activation function.

[0036] In particular, the activation function of the middle layer of the multilayer perceptron is a nonlinear function such as ReLU, Sigmoid or Tanh; the activation function of the output layer is a linear function.

[0037] Further, the loss function of the hybrid machine learning model described in step S5 is: Loss=Loss1+Loss2, wherein Loss1=CosineEmbeddingLoss(i, j) is the sum of the cosine similarities between any two nodes i and j in the weighted graph convolutional network; Loss2=MSELoss(y_predicted, y_label) is the mean square error between the actual distribution station power transmission y_label and the predicted value y_predicted.

[0038] Compared with ordinary graph convolutional networks, weighted graph convolutional networks add a diagonal membership matrix, which can weight different features according to their importance to power supply, so that features of different importance contribute differently during convolution. The more important the feature, the greater its weight during convolution, and the greater its impact on the result represented by the node vector.

[0039] In particular, if time period T is not the first time period, the weighted graph convolutional network is first used to fuse the node features in the distribution network in the current time period, and the feature vector representation of the node is set to current_embedding. The feature vector representation update_embedding obtained in the previous time period is put into the gated graph neural network to fuse the time information to obtain a new update_embedding, and then the new update_embedding is fused with the current_embedding of the current time period to obtain the feature vector representation of the node in the next time period.

[0040] Preferably, when the input and output dimensions of the weighted graph neural network, gated graph neural network and multilayer perceptron of the hybrid machine learning model are the same, the trained multilayer perceptron can be directly used to execute step S6 to predict the load data of the distribution network at the next moment, thereby improving the prediction efficiency.

[0041] The beneficial effects of the present invention are as follows: the present invention provides a distribution network load prediction device and method based on hybrid machine learning, which uses Copula entropy and fuzzy membership function to weight the preprocessed feature variables to improve the accuracy and efficiency of the prediction model; introduces weighted graph convolutional networks and gated graph neural networks to achieve feature fusion and feature update between distribution stations and distribution stations, and then uses a multi-layer perceptron to predict the power load, which significantly improves the accuracy, speed and generalization ability of power load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to make the purpose and technical solution of the present invention, the present invention provides the following drawings for explanation:

[0043] Figure 1It is a framework diagram of the distribution network load prediction device based on hybrid machine learning of the present invention; wherein the solid line is the data flow direction, and the dotted line is the training data interaction process;

[0044] Figure 2 It is a flow chart of the distribution network load forecasting method based on hybrid machine learning of the present invention;

[0045] Figure 3 This is a schematic diagram of the distribution network area in Example 1 of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose and technical solution of the present invention more clearly understood, the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0047] Example 1: Combination Figure 3 , the power supply system in a certain distribution network area is monitored as a whole by the master station system, which has k=8 distribution station areas under its jurisdiction, and each distribution station supplies power to several residential communities. Each residential community is equipped with an electric energy meter, and the load data is uploaded to the collector through RS485 communication, and then the data is uploaded to the concentrator through power line carrier communication, and finally the data is sent to the master station system through wireless communication. The characteristic data of the distribution station and the characteristic data of the residential community are stored in the master station system. The characteristics of the distribution station include: capacity, voltage level, type, geographical location, and grid structure. The master station system is connected to the Internet to obtain other random features, including: average temperature, humidity, air pressure within a time period, and whether the time period contains holidays. In order to achieve optimal configuration of power supply in the region and realize accurate power distribution, the present invention provides a "distribution network load prediction device based on hybrid machine learning."

[0048] Combination Figure 1, specifically comprising: a data acquisition module (1), a data storage module (2), a data preprocessing module (3) and a prediction module (4); the data acquisition module (1) is a corresponding concentrator, which reads the data of the sensor of the residential area through RS485 communication and power line carrier communication to realize the collection of the load of the residential area, wherein the sensor is an electric energy meter. The data storage module (2) is composed of a load data storage unit (21) and a prediction data storage unit (22); the data preprocessing module (3) is composed of a data preprocessing unit (31) connected in series with a feature weighting unit (32), and the data preprocessing unit (31) is directly connected to the data acquisition module (1) for real-time data processing; the load storage unit (21) is part of the memory of the main station system, and is respectively connected to the data acquisition module (1) and the data preprocessing unit (31) through wireless communication, and is used to store and transmit the load data collected by the data acquisition module (1); the prediction data storage unit (22) is part of the memory of the main station system, and is connected to the prediction module (4) for storing data fed back by the prediction module (4); the feature weighting unit (32) is connected to the prediction module (4) for preprocessing; the prediction module (4) is a processor with a hybrid machine learning model in the main station system, and is used to predict the power supply of the distribution station. The hybrid machine learning model is composed of a weighted graph neural network connected in series with a gated graph neural network and then connected in series with a multi-layer perceptron (MLP). The weighted graph neural network is an effective convolutional neural network that can directly act on the graph, and is used to achieve feature fusion; the gated graph neural network updates the feature information of the node through a message passing mechanism, so that the node can use the information of its neighboring nodes to update itself, and is used to achieve fusion feature update; the multi-layer perceptron is used to achieve the prediction of the fused and updated features.

[0049] Embodiment 2: Based on the device of embodiment 1, the present invention also provides a "distribution network load forecasting method based on hybrid machine learning" to predict the power supply trend of the distribution station.

[0050] Combination Figure 2 , which specifically includes the following steps:

[0051] Step 1: The data collection module (1) collects load data of all residential communities within the jurisdiction of the distribution network and stores the data in a load data storage unit (21).

[0052] The load data includes: residential area ID, time, and load amount.

[0053] Step 2: The data preprocessing unit (31) preprocesses the residential area load data in the load data storage unit (21) within each month to obtain a weight coefficient.

[0054] The preprocessing specifically includes: (i) detecting abnormal data using the k-nearest neighbor algorithm (KNN) and replacing the abnormal data with neighboring data; (ii) performing a linear transformation on the load using the min-max normalization method so that its value is transformed to the interval [0, 1] to obtain a weight coefficient; and (iii) performing One-Hot Encoding on the holidays and month information corresponding to the time period.

[0055] Step 3: Taking each distribution station area as a unit, the feature weighting unit (32) uses Copula entropy and fuzzy membership function to weight the feature vectors of the corresponding distribution stations, residential communities and other random feature vectors in each month to obtain node attribute features.

[0056] Specifically:

[0057] 1) According to the time sequence, the characteristic values ​​of the distribution stations, the characteristic values ​​of the residential areas and other random characteristic values ​​within the jurisdiction of a distribution station are constructed into a characteristic vector;

[0058] 2) Calculate the Copula entropy between each eigenvector and the power supply vector of the distribution station;

[0059] The power supply vector of the distribution station is the sum of the load characteristic vectors of all residential communities within the jurisdiction of the distribution station.

[0060] Furthermore, for any two n-dimensional feature vectors X = [X1, X2, ..., X n ] and Y = [Y1, Y2, ..., Y n ], the Copula entropy calculation process of X and Y is as follows: First, we need to obtain the joint probability density function f of X and Y X，Y (x, y), then calculate the Copula function Among them, C(u, v) is the marginal distribution function F X and F Y Function mapped to the interval [0, 1], u = F X (x) and v = F Y (y).

[0061] It should be noted that the frequency of data collection within each month is much greater than 1. The more times data is collected, the more accurate the prediction is. This frequency corresponds to the dimension of the feature vector.

[0062] 1) Use the Copula entropy corresponding to each eigenvector calculated in step 2) to calculate the corresponding membership degree, and then construct the membership matrix of all eigenvectors within the jurisdiction of the distribution station, which is the node attribute feature.

[0063] For the i-th eigenvector, the Copula entropy c i, the membership degree is calculated as: The membership matrix B of all M features is:

[0064] Step 4: Take the jurisdiction of each distribution station as a unit, the distribution stations and residential communities within the jurisdiction as nodes, and the node attribute characteristics as the membership matrix; the edge is the relationship between nodes, corresponding to the adjacency matrix A, and the weight of the edge is the power supply of the corresponding distribution station to the residential community. Construct a weight graph for the distribution station jurisdiction every month.

[0065] Step 5: The prediction module (4) establishes a hybrid machine learning model, sets a loss function, and uses a weight map within the distribution network area in a historical time period for training, and stores the results in a prediction data storage unit (22).

[0066] Furthermore, the hybrid machine learning model consists of a weighted graph neural network connected in series with a gated graph neural network and then connected in series with a multi-layer perceptron (MLP).

[0067] Furthermore, for a weighted graph of a distribution substation area with N nodes and M=16-dimensional features, the working principle of the weighted graph convolutional network is as follows:

[0068]

[0069] in, is the feature matrix, which is the matrix composed of all the feature vectors in step S3; I N is the N-order identity matrix; is an approximate graph convolution filter, Weight Matrix The corresponding coefficients are generated by the weight data of step S2; D is the dimension of the output fusion feature X of the weighted graph convolutional network; f(·) is the activation function.

[0070] Furthermore, the gated graph neural network, whose input is the output of the weighted graph convolutional network, works as follows:

[0071]

[0072] in, represents the feature update vector of the area under the jurisdiction of the l-th distribution station i. When l = 0, is the output fusion feature X of the corresponding weighted graph convolutional network; represents the feature update vector of the lth layer corresponding to the distribution station j adjacent to the distribution station i; and is the parameter matrix to be learned; σ(·) is the sigmoid function; ⊙ represents element-by-element multiplication; [·] represents the vector concatenation operation; N(i) represents the set of adjacent substations of substation i.

[0073] Furthermore, the input of the multilayer perceptron is the output of the gated graph neural network, and the dimension of its output is the same as the number of distribution stations in the distribution network area, which is k=8. Its working principle is:

[0074]

[0075] in, is the output of the jth neuron in the lth layer. When l = 0, is the feature update vector of the j-th distribution station output by the gated graph neural network; n (l-1) is the number of neurons in the previous layer; is the weight to be learned between the jth neuron in the lth layer and the ith neuron in the l-1th layer; is the bias term to be learned; f(·) is the activation function.

[0076] In particular, the activation function of the middle layer of the multilayer perceptron is ReLU and the activation function of the output layer is the linear function Identity.

[0077] Furthermore, the loss function of the hybrid machine learning model is: Loss=Loss1+Loss2, wherein Loss1=CosineEmbeddingLoss(i, j) is the sum of the cosine similarities between any two nodes i and j in the weighted graph convolutional network; Loss2=MSELoss(y_predicted, y_label) is the mean square error between the actual distribution station power transmission y_label and the predicted value y_predicted.

[0078] S6: The prediction module (4) uses the trained hybrid machine learning model to predict the power supply of the distribution station at the next moment, which corresponds to the load data of the distribution network.

[0079] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A distribution network load forecasting method based on hybrid machine learning, characterized in that: The method comprises the following steps: S1: The data collection module (1) collects load data of all residential areas within the distribution network area and stores the data in the load data storage unit (21); S2: the data preprocessing unit (31) preprocesses the residential area load data in the load data storage unit (21) within the time period T to obtain a weight coefficient; S3: Taking each distribution station area as a unit, the feature weighting unit (32) uses Copula entropy and fuzzy membership function to perform weighted processing on the feature vectors of the corresponding distribution stations, the feature vectors of the residential areas and other random feature vectors within the time period T to obtain node attribute features; S4: Taking each distribution station area as a unit and the distribution stations and residential areas within the area as nodes, the data preprocessing module (3) constructs a weight graph within the distribution station area time period T using the weight data of step S2 and the node attribute characteristics of step S3; S5: The prediction module (4) establishes a hybrid machine learning model, sets a loss function, and uses a weight map within the distribution network area in a historical time period for training; S6: The prediction module (4) uses the trained hybrid machine learning model to predict the load data of the distribution network at the next moment; The load data includes: residential area ID, time, load amount; The preprocessing described in step S2 specifically includes: (1) using the k-nearest neighbor algorithm (KNN) to detect abnormal data and replace the abnormal data with neighboring data; (2) using the min-max normalization method to linearly transform the load so that its value is transformed to the interval [0,1] to obtain a weight coefficient; (3) One-Hot Encoding the holiday and month information corresponding to the time period; The step S3 is specifically as follows: S301: constructing characteristic values ​​of a distribution station, characteristic values ​​of a residential area and other random characteristic values ​​within a distribution station area into a characteristic vector according to time sequence; S302: Calculate the Copula entropy between each characteristic vector and the power supply vector of the distribution station; the power supply vector of the distribution station is the sum of the load characteristic vectors of all residential communities within the jurisdiction of the distribution station; Furthermore, for any two n-dimensional feature vectors X = [X1, X2, ..., X n ] and Y=[Y1,Y2,...,Y n ], the Copula entropy calculation process of X and Y is as follows: First, we need to obtain the joint probability density function f of X and Y X,Y (x,y), and then calculate the Copula function Among them, C(u,v) is the marginal distribution function F X and F Y Function mapped to the interval [0,1], u = F X (x) and v = F Y (y); S303: Calculate the corresponding membership degree using the Copula entropy corresponding to each eigenvector calculated in step S302, and then construct the membership degree matrix of all eigenvectors within the jurisdiction of the distribution station, which is the node attribute feature; Furthermore, for the i-th eigenvector, the Copula entropy c i , the membership degree is calculated as: The membership matrix B of all M features is: A weight graph is established for each distribution station area. The weight graph of the distribution station area in the time period T described in step S4 is: the distribution stations and residential areas in the area are nodes, and the node attribute characteristics are the membership matrix; the edges are the relationship between the nodes, and the adjacency matrix A is constructed; the weight of the edge is the power supply from the corresponding distribution station to the residential area, that is, the weight data of step S2; The hybrid machine learning model described in step S5 is composed of a weighted graph neural network connected in series with a gated graph neural network and then connected in series with a multi-layer perceptron; Furthermore, for a weighted graph of a distribution station area with N nodes and M-dimensional features, the working principle of the weighted graph convolutional network is as follows: in, is the feature matrix, which is the matrix composed of all the feature vectors in step S3; I N is the N-order identity matrix; is an approximate graph convolution filter, Weight Matrix The corresponding coefficients are generated by the weight data of step S2; D is the dimension of the output fusion feature X of the weighted graph convolutional network; f(·) is the activation function; Furthermore, the gated graph neural network, whose input is the output of the weighted graph convolutional network, works as follows: in, represents the feature update vector of the area under the jurisdiction of the l-th distribution station i. When l = 0, h i (0) is the output fusion feature X of the corresponding weighted graph convolutional network; represents the feature update vector of the lth layer corresponding to the distribution station j adjacent to the distribution station i; and W r (l) is the parameter matrix to be learned; σ(·) is the sigmoid function; ⊙ represents element-by-element multiplication; [·] represents the vector concatenation operation; N(i) represents the set of adjacent substations of substation i; Furthermore, the input of the multilayer perceptron is the output of the gated graph neural network, and its output and output dimension are the same as the number of distribution stations in the distribution network area. Its working principle is: in, is the output of the jth neuron in the lth layer. When l = 0, is the feature update vector of the j-th distribution station output by the gated graph neural network; n (l-1) is the number of neurons in the previous layer; is the weight to be learned between the jth neuron in the lth layer and the ith neuron in the l-1th layer; is the bias term to be learned; f(·) is the activation function; The loss function of the hybrid machine learning model described in step S5 is: Loss=Loss1+Loss2, wherein Loss1=CosineEmbeddingLoss(i,j) is the sum of the cosine similarities between any two nodes i and j in the weighted graph convolutional network; Loss2=MSELoss(y_predicted,y_label) is the mean square error between the actual distribution station power transmission y_label and the predicted value y_predicted.

2. The method for distribution network load forecasting based on hybrid machine learning according to claim 1 is characterized in that: The amount of load data collected within the time period T is greater than the output dimension of the weighted graph neural network.

3. The method for distribution network load forecasting based on hybrid machine learning according to claim 1, characterized in that: The characteristics of the distribution station described in step S3 include: capacity, voltage level, type, geographical location, and grid structure, a total of 5; the characteristics of the residential community include: area, total population, population density, total income of residents, average income of residents, geographical location of the community, electricity price of the community, and load, a total of 8; other random characteristics include: average temperature, humidity, air pressure within a time period, and whether the time period contains holidays, a total of 4.

4. The method for distribution network load forecasting based on hybrid machine learning according to claim 1, characterized in that: The activation function of the middle layer of the multi-layer perceptron is a nonlinear function; the activation function of the output layer is a linear function.

5. The method for distribution network load forecasting based on hybrid machine learning according to claim 1, characterized in that: When the input and output dimensions of the weighted graph neural network, the gated graph neural network and the multilayer perceptron in the hybrid machine learning model are the same, the trained multilayer perceptron is directly used to execute step S6 to predict the load data of the distribution network at the next moment, thereby improving the prediction efficiency.

6. A distribution network load forecasting device based on hybrid machine learning based on the method according to any one of claims 1 to 5, characterized in that: The device comprises a data acquisition module (1), a data storage module (2), a data preprocessing module (3) and a prediction module (4); the data acquisition module (1) reads load data of sensors in residential areas within the jurisdiction of the distribution network through communication; the data storage module (2) is composed of a load data storage unit (21) and a prediction data storage unit (22); the data preprocessing module (3) is composed of a data preprocessing unit (31) connected in series with a feature weighting unit (32); the load data storage unit (21) is a memory, which is respectively connected to the data acquisition module (1) and the data storage module (21). The data preprocessing unit (31) is directly connected to the data acquisition module (1) and is used to store and transmit the residential area load data collected by the data acquisition module (1); the data preprocessing unit (31) is directly connected to the data acquisition module (1) and is used to read the residential area load data in real time; the predicted data storage unit (22) is a memory, connected to the prediction module (4), and is used to store and transmit the data fed back by the prediction module (4); the feature weighting unit (32) is connected to the prediction module (4); the prediction module (4) is a processor with a hybrid machine learning model in its memory, which is used to predict the power supply of the distribution station.

Images