Power load prediction method based on relation enhancement supervised learning

Abstract

The invention belongs to the technical field of electric power prediction, and particularly relates to an electric power load prediction method based on relation enhancement supervised learning, which comprises the following steps: acquiring historical data related to electric power load from an electric power system monitoring station, preprocessing the data, and analyzing the correlation between enterprises so as to establish an enterprise correlation graph; a neural network model GL-STGNN is established to train the neural network model GL-STGNN, the predicted data of the power load in a future set time period overcomes the defects in the prior art, frequency domain and time domain multi-dimensional global time sequence characteristics extracted by a global GRU model are fused, a local model of a node is introduced through a node embedding module, and the overall performance of the power load is improved. According to the method, global-local feature fusion is completed through splicing, so that global features and local effects are fully considered, related features are captured to the maximum extent, and the model has the accurate prediction capability.

Description

Technical Field

[0001] The present invention belongs to the technical field of power forecasting, and in particular relates to a power load forecasting method based on relationship enhanced supervised learning. Background Art

[0002] With the development of social economy and the improvement of living standards, the demand for electricity continues to grow, which brings huge challenges to the planning and operation of power systems. Nowadays, human production and life are closely related to electricity. Therefore, accurate prediction of power load is of great significance for guiding the development planning of power systems, formulating production plans, and conducting economic dispatch.

[0003] Electricity load refers to the amount of electricity demanded by a region or enterprise at a certain point in time. By using the electricity load values ​​recorded over a period of time and using mathematical methods to capture the connections and patterns between the values, it is possible to predict the electricity load for a period of time in the future. Short-term load forecasting refers to the prediction of electricity load for the next few hours, days or dozens of days. A highly accurate short-term load forecasting model can provide an approximate power generation range for power plants in daily operation, and formulate a power generation plan based on this, which is of great significance to ensuring the security of the power grid and reducing the cost of power generation.

[0004] In order to establish an effective time series forecasting model to improve the forecasting accuracy, people have made many explorations. Traditional power load forecasting is mainly based on historical load data to establish statistical models (such as AR, ARIMA). These linear statistical methods are often difficult to capture and model the complex characteristics of time series forecasting such as nonlinearity, non-stationarity, and high dimensions. Power load forecasting methods based on machine learning, such as Hidden Markov Model, Bayesian Network, Support Vector Machine (SVM), etc., have overcome the above difficulties to a certain extent, but there is still a problem of weak fitting ability for nonlinear models.

[0005] Time series forecasting methods based on deep learning use neural networks to learn complex patterns and dependencies in time series data, thereby predicting future values ​​based on historical information. Its advantage is that it can adaptively learn complex features of data without requiring too much domain knowledge and prior assumptions. It can be roughly divided into three categories: Deep learning methods can be divided into three categories: Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Encoder-Decoder Structure (Transformer) models.

[0006] Compared with the Transformer model, convolutional neural networks and recurrent neural networks are weaker in long-term modeling capabilities. However, the Transformer-based model, with its powerful global modeling capabilities, has not only shined in the field of natural language processing, but also has a wide range of applications in time series prediction due to its superb ability to process sequence data. However, most of the methods that only improve the Transformer model structure ignore the dependencies between variables. Therefore, for the processing of dependencies between variables in multivariate time series prediction problems, since graph neural networks (GNNs) have shown strong capabilities in processing relational dependencies, in recent years, methods that use graph neural networks to capture dependencies other than the time dimension, such as latent spatial features between pairs of variables, and then use other models (such as RNN, LSTM, GRU) to learn time series features have sprung up like mushrooms after rain.

[0007] Graph neural networks can handle non-Euclidean data such as transportation subway topology very well. By treating time points or variables as nodes, and their relationships as edges, modeling them using network or graph structures, and using the message propagation mechanism between graph neural network nodes to effectively learn and enhance the association relationship, the accuracy and precision of time series prediction tasks can be improved.

[0008] In recent years, the frontier research focus of graph neural networks in time series forecasting can be roughly divided into dynamic graph research and stronger and more accurate capture of spatiotemporal dependencies. However, it is worth noting that whether using recursive neural networks in the time dimension or graph neural networks in the space dimension, each node still essentially shares weights. Therefore, in solving multivariate time series problems, although mainstream spatiotemporal neural networks can capture global features, they may perform poorly for data with large differences between nodes (for example, some time series generated by different random processes) due to the inadequate capture of the personalized features of the nodes. Most previous works have not considered the personalized features of each node (hereinafter referred to as "local features"), or even if local features are considered, the fusion with global features is ignored. The key influence of local features on multivariate time series makes it necessary to capture them, and global features cannot be ignored. Therefore, a power load forecasting method that comprehensively considers local and global features is urgently needed. Summary of the invention

[0009] The purpose of the present invention is to provide a power load forecasting method based on relational enhanced supervised learning, which overcomes the shortcomings of the prior art. It not only integrates the frequency domain and time domain multi-dimensional global time series features extracted by the global GRU model, but also introduces the local model of the node through the node embedding module, and completes the global-local feature fusion by splicing. Therefore, it fully considers the global characteristics and local effects, and then maximizes the capture of relevant features, so that the model has the ability of accurate prediction.

[0010] To solve the above problems, the technical solutions adopted by the present invention are as follows:

[0011] A method for power load forecasting based on relational enhanced supervised learning, comprising the following steps:

[0012] Step 1: Obtain historical data related to power load from the power system monitoring station and pre-process the data;

[0013] Step 2: Analyze the correlation between enterprises based on the historical data related to power load preprocessed in step 1, so as to establish an enterprise correlation graph;

[0014] Step 3, establish a neural network model GL-STGNN, take the power load related historical data processed in step 1 as input, take the enterprise related graph established in step 2 as the spatial relationship basis, and train the neural network model GL-STGNN;

[0015] Step 4: Input the real-time power load data into the neural network model GL-STGNN trained in step 3 to obtain the predicted power load data within the future set time period.

[0016] Furthermore, the data preprocessing in step 1 includes data cleaning and additional data collection, wherein the data cleaning is used to eliminate null values ​​and duplicate values ​​in the historical data related to the power load, and the additional data collection is used to connect to the Internet to obtain holidays and weather data for the corresponding time period of the historical data related to the power load.

[0017] Further, the expression of the enterprise-related graph in step 2 is:

[0018] G=(V,E,A)

[0019] Among them, V represents the vertex set, E is the edge set, and A is the adjacency matrix;

[0020] The expression of the adjacency matrix A is:

[0021]

[0022] Among them, R ij is the calculated correlation value between enterprises.

[0023] Furthermore, the inter-enterprise correlation value is the correlation entropy of the data variables in one period, and the specific expression is:

[0024] V γ (X,Y)=E[κ γ (XY)]

[0025] Among them, X and Y are two variables, E represents mathematical expectation, κ γ is the kernel function with bandwidth , and the Gaussian kernel is usually chosen.

[0026] Furthermore, the κ γ The specific expression is:

[0027]

[0028] Among them, γ is the bandwidth of the kernel function, that is, the size of the Gaussian radial basis function, which controls the influence of each data point on other data points in the similarity measurement.

[0029] Furthermore, the neural network model GL-STGNN includes a node embedding module, an encoder, a STMP module and a decoder. The node embedding module is spliced ​​with the sequence information of the node in the current window and input into the encoder. The encoder is mapped to a high-dimensional space, performs preliminary feature extraction, and then inputs into the STMP module. The STMP module is used to propagate information in the graph structure and perform key parts of feature extraction. The decoder is used to process the hidden representation extracted by the model to generate prediction results for the time series.

[0030] Furthermore, the node embedding module initially defines a trainable parameter representing the embedding vectors of multiple nodes, and adjusts the embedding parameter during training to learn the representation vector of the node. The learned representation vector is handed over to subsequent modules to introduce the local model into the global model and integrate the local effect into the global feature.

[0031] Further, the mathematical expression of the encoder is:

[0032]

[0033] in, represents the power load data of the ith node at the t-1th time step, u t-1 Indicates external auxiliary information such as holidays and weather, v i Represents the embedding vector of the node.

[0034] Furthermore, the STMP module includes a TFC module and an MP module, wherein the TFC module uses GRU to capture timing features, and the MP module is used to transmit the captured timing information and update the node status.

[0035] Further, the expression of the decoder is:

[0036]

[0037] in, is the representation learned by the i-th node in the STMP module, vi Represents the node embedding vector.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The power load forecasting method based on relational enhanced supervised learning described in the present invention not only integrates the frequency domain and time domain multi-dimensional global time series features extracted by the global GRU model, but also introduces the local model of the node through the node embedding module, and uses splicing to complete the global-local feature fusion, thereby fully considering the global features and local effects, and then maximizing the capture of relevant features, so that the model has the ability of accurate prediction.

[0040] The present invention takes into account both accuracy and economic benefits and time cost. The comparative experiment with the baseline proves that the proposed method is superior to the traditional method in terms of accuracy, and has fewer training parameters and shorter training time. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a principle structure diagram of the neural network model GL-STGNN in a power load forecasting method based on relational enhanced supervised learning.

[0042] Figure 2 This is a principle structure diagram of the STMP module in a power load forecasting method based on relational enhanced supervised learning.

[0043] Figure 3 It is a bar chart of the trainable parameter amount and training time of each model in the embodiment. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] As shown in FIG1 , the method for power load forecasting based on relationship-enhanced supervised learning of the present invention comprehensively considers spatial, temporal characteristics, frequency domain, and time domain characteristics, and integrates global and local characteristics to complete the task of power load forecasting for the next few days; specifically, the following steps are included:

[0046] Step 1) Perform data analysis and data processing. First, since the prediction target is the peak power load of the next few days, and the sampling frequency of the original data is 15 minutes, the original data needs to be resampled so that the sampling frequency is unified to the unit of day.

[0047] In order to make full use of the original data as much as possible, this embodiment selects the maximum value (peak value) and the corresponding time point, the minimum value (trough value) and the corresponding time point, the average value, and the variance of the sampled data on the day according to the 96 sampling points of the day as the signal description quantity with a frequency of day after resampling. Then the data set is cleaned, and the null values ​​are filled or discarded. Next, the data distribution is observed with the help of histograms and box plots to understand the data characteristics. In addition, in order to select a time slice length suitable for model training, Fourier analysis and autocorrelation coefficient are combined to analyze the periodicity of the data set. In this embodiment, the time slice length used is 6.

[0048] Finally, holiday information and wind cooling index within the time period of the data set used were collected through the Internet and introduced as additional information to enrich the data features and provide more prediction basis for the model.

[0049] In summary, in step 1), a clean dataset X = {χ 1 , χ 2 , ..., χ n}, where χ i Represents the power load data vector of the i-th enterprise, with a vector dimension of d x . Additional data collection was performed to obtain data U = {u1, u2, ..., u t}={T,C}, where T represents the meteorological data during this period (i.e., wind cooling index WCI), C represents the holiday data during this period, and the vector u i The dimension is d u .

[0050] Step 2) Establishing the enterprise correlation graph: Based on the data set X and the additional data U obtained in step 1), through correlation analysis, the adjacency matrix A of the enterprise correlation graph G = (V, E, A) is determined according to the correlation coefficient.

[0051] The expression of the adjacency matrix A is:

[0052]

[0053] Among them, R ij is the calculated correlation value between enterprises.

[0054] Correlation R ij It can be obtained by calculating the correlation entropy of variables in a period. Given two different enterprises' power load vectors χ i and χ j , calculate the correlation entropy correlation (CorrentropyCoefficient) of two variables as shown in the following formula:

[0055]

[0056] Among them, E[·] represents the mathematical expectation, κ γ (·) is a kernel function with a bandwidth of γ, and a Gaussian kernel is usually selected (as shown in the following formula).

[0057] κ γ The specific expression is:

[0058] Among them, γ is the bandwidth of the kernel function, that is, the size of the Gaussian radial basis function, which controls the influence of each data point on other data points in the similarity measurement.

[0059] Thus, the similarity matrix is ​​obtained by calculating the correlation entropy between the time series, and then the adjacency matrix is ​​constructed by extracting the k nearest neighbor graph from the similarity matrix. In this embodiment, k=30.

[0060] For each enterprise's historical power load time series, find the top k time series with the highest similarity through k nearest neighbors. For its k nearest neighbors, replace the element a in the adjacency matrix with ij Set it to 1 to establish an edge, and set the rest to 0, indicating that the correlation is weak and can be ignored.

[0061] Therefore, after step 2), the enterprise correlation graph G = (V, E, A) is obtained, which serves as the basis for capturing the spatial relationship in step 3).

[0062] Step 3) Build the GL-STGNN model. The GL-STGNN model is divided into four parts: node embedding, encoder, STMP module, and encoder.

[0063] In the node embedding module, the local features of each enterprise are obtained based on the datasets X and U obtained in step 1) using the idea of ​​Embedding. Initially, a Trainable parameters, where N represents the number of enterprises, d v =d x +d u , representing the embedding vectors of multiple nodes. The embedding parameters are adjusted during training to learn the local features of the node. In order to enhance the stability and robustness of the model, the embedding space is limited to a certain range while capturing local effects as much as possible, ensuring the model training capability and efficiency, and introducing clustering-based regularization to structure the embedding space. In addition, since the distribution process of cluster centers is discrete, in order to make this process gradient calculable, thereby guiding the embedded representation to approach the cluster center during the end-to-end training process, the Gumble-Softmax technique is adopted here. Finally, clustering-based regularization will use the loss function Lreg , during training, the distance between the node embedding matrix V and the cluster center C is minimized to achieve embedding aggregation to the corresponding cluster center.

[0064] L reg =E M ||V-MC||2

[0065]

[0066] Among them, p(M ij ) represents the probability of node i being assigned to cluster j, and τ is a hyperparameter used to control the smoothness of the softmax function. reg It measures the distance between each node and a certain centroid. ij It is the noise extracted from the Gumbel distribution Gumbel(0,1). First, the standard uniform distribution random variable uni is extracted, where uni~Uniform(0,1), and then the Gumbel noise g is generated by the inverse transformation method, where g=-log(-log(uni)).

[0067] Therefore, in the node embedding module, the power load vector of the enterprise And the additional information vector u at this time step t , and obtain the embedding vector v representing the local features of the enterprise i .

[0068] The encoder then converts the input time series data into a hidden representation or feature representation, preliminarily captures the features and patterns in the time series data, and fuses the input data of each enterprise u t-1 With the embedding vector v i The encoder is a fully connected linear layer, namely:

[0069]

[0070] Thus, in the encoder, the features after preliminary extraction and fusion are obtained

[0071] Next, the STMP module will be used to complete the extraction of temporal features and spatial features. This layer first uses the TFC module to capture temporal features, and then uses the MP module to transmit the captured temporal information based on the enterprise correlation graph defined in step 2) and update the node status.

[0072] The TFC module mainly uses GRU to capture time series features. After the time series signal is transformed by DFT, a complex number will be obtained. The real part and imaginary part represent the amplitude of the cosine and sine components of the corresponding frequency components in the frequency domain. In the Spe-Sqe unit, the real part and imaginary part will be sent to two GRU units respectively, and then the output of the GRU unit will be recombined into a complex number for IDFT (inverse discrete Fourier transform). Finally, the result is activated by tanh nonlinearity to further capture nonlinear features. The TFC module will also send the original time series data to another GRU unit, and use the sigmoid activation function to map the output to between 0 and 1, and then multiply its output by the output of the Spe-Sqe unit element by element, so as to screen the features.

[0073] The features captured by the TFC module will be input into the MP module to continue to transmit information in the spatial dimension and retain the modeling ability of spatial dependencies. The MP module will rely on the pre-defined enterprise correlation graph and adopt an isotropic message passing mechanism, that is, the message passing function only depends on the sending node. It is implemented using GCN, and jump connections are used between layers of GCN to avoid gradient disappearance. The calculation method of the MP module is as shown in the formula:

[0074]

[0075] in, That is v i Neighbor nodes, σ is the activation function, is a trainable parameter. The MP layer transmits messages so that the node representation is updated by aggregating the messages of neighboring nodes.

[0076] Therefore, after being processed by the STMP module, the model obtains global temporal and spatial features.

[0077] Finally, the decoder fuses the global features extracted from the STMP module with the local features corresponding to the node embedding, and maps the learned fusion features back to the original time series space to achieve prediction of future time steps. The decoder is a multi-layer perceptron (MLP), that is:

[0078]

[0079] in, is the representation learned by the i-th node in the STMP module, v i Represents the node embedding vector. The MLP layer consists of a hidden layer and an output layer. The activation function uses elu. Finally, the prediction result

[0080] The calculation is shown in the formula:

[0081]

[0082]

[0083] In the experiment, the batch size is 16, the embedding vector dimension is 64, and the clustering-based regularization attempts to cluster the embedding vector into 5 categories. The hidden layer dimension is 128. There are two layers of GRU modules and two layers of GCN modules, and a Dropout layer with a ratio of 0.05 is added between the two layers of GRU modules to prevent overfitting.

[0084] The training used MAE (MeanAbsolute Error) as the loss function, the Adam optimizer, and dynamically adjusted the learning rate. The initial learning rate was 0.003, and the learning rate was increased by 0.25 times every 50 epochs. In addition, the early-stopping mechanism was used. If the loss value of the model validation set did not decrease after 50 epochs, the training was terminated. A total of 200 epochs were trained, and the training set, validation set, and test set were divided into a ratio of 6:2:2.

[0085] In order to evaluate the effect of the algorithm on power load forecasting, three evaluation indicators are used: RMSE, MAE, and MAPE. The smaller the value of the three evaluation indicators, the higher the prediction accuracy. The calculation method of each evaluation indicator is as follows:

[0086]

[0087]

[0088]

[0089] In order to prove that the GL-STGNN model proposed in this invention is more suitable for power load forecasting in multi-enterprise situations than previous time series forecasting methods, the comparison results of the model with other algorithms on the test set in predicting the power load peaks in the next 1 day, 5 days and 10 days are shown in Table 1. Among them, LocalGRU is a method of assigning a GRU model to each enterprise.

[0090]

[0091] It can be seen from Table 1 that for predicting the power load for the next day, the GraphWaveNet model performs best in the MAE evaluation index, followed by the proposed GL-STGNN model; while in the RMSE and MAPE evaluation indicators, the proposed GL-STGNN model is the best, which shows that the model is less sensitive to outliers and extreme errors and can provide stable and accurate predictions under various load levels.

[0092] For the prediction of power load in the next five and ten days, the proposed GL-STGNN model performs best in terms of MAE, RMSE and MAPE.

[0093] Whether it is short-term or long-term prediction, the local model LocalGRU has obvious improvements over the traditional GRU model, which verifies the importance of capturing local effects. In addition, AGCRN also considers local effects by replacing shared parameters in graph convolution with node-independent parameters. It is worth noting that the performance of AGCRN and DCRNN on this dataset is not ideal, but AGCRN is better than DCRNN, which is also a reflection of local effects.

[0094] As for why the performance of AGCRN and DCRNN is not ideal, or even weaker than the GRU model, the researchers believe that the unsatisfactory performance of AGCRN and DCRNN may be related to the small number of time steps in the dataset itself. Fewer time steps will cause more complex models to be unable to fully learn effective features, which in turn affects their prediction performance, making it difficult for AGCRN and DCRNN to deeply learn their temporal patterns. GraphWaveNet and GL-STGNN, which are also graph-based learning methods, perform better. For GraphWaveNet, the researchers believe that this may be related to its use of TCN rather than traditional RNN to capture time dependencies. When the amount of data is small, TCN performs convolutions at multiple scales to ensure that the model extracts information from multiple angles and levels. For GL-STGNN, the method of introducing local effects in this model is different from that of AGCRN. Although GL-STGNN also adopts a recursive neural network, it introduces encoding, decoding and node embedding modules. Node embedding is involved in both the encoder and decoder modules. The local information it carries not only participates as a feature, but also is a kind of historical information because it is continuously updated during training. This provides the model with multi-level information, which is conducive to the learning of temporal patterns.

[0095] The experiment draws a bar chart of the trainable parameters of each model when the number of nodes is 285. The results are as follows Figure 3As shown on the left. As can be seen from the figure, since LocalGRU assigns a model to each enterprise, its training parameter volume is in significant contrast with the other models. In fact, when all the nodes of the entire data set (a total of 1725) are sent to training, the parameter volume of the LocalGRU model is as high as 227MB, and this parameter volume will continue to increase with the increase of nodes. Therefore, it is impractical to blindly assign a model to each node to capture local effects. Although the model AGCRN, which also captures local effects, uses matrix decomposition to reduce the number of parameters, it is still much higher than the GL-STGNN model in terms of parameters (2MB and 715K respectively). The study drew a bar graph of the training time of each model, and the results are as follows: Figure 3 As shown on the right, GL-STGNN also has an outstanding advantage in terms of training time. It takes less training time while ensuring prediction accuracy. Although GraphWaveNet also has a high accuracy in predicting the next day, its training time cost is the highest among all models.

[0096] In summary, the proposed GL-STGNN model shows better comprehensive performance than other baseline models at different time granularities of short-term prediction, and has excellent performance with fewer parameters and shorter training time.

[0097] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for power load forecasting based on relational enhanced supervised learning, characterized in that: The following steps are involved: Step 1: Obtain historical data related to power load from the power system monitoring station and pre-process the data; Step 2: Analyze the correlation between enterprises based on the historical data related to power load preprocessed in step 1, so as to establish an enterprise correlation graph; Step 3, establish a neural network model GL-STGNN, take the power load related historical data processed in step 1 as input, take the enterprise related graph established in step 2 as the spatial relationship basis, and train the neural network model GL-STGNN; Step 4: Input the real-time power load data into the neural network model GL-STGNN trained in step 3 to obtain the predicted power load data within the future set time period.

2. The method for power load forecasting based on relational enhanced supervised learning according to claim 1, characterized in that: The data preprocessing described in step 1 includes data cleaning and additional data collection. The data cleaning is used to eliminate null values ​​and duplicate values ​​in the historical data related to the power load. The additional data collection is used to connect to the Internet to obtain holidays and weather data for the corresponding time period of the historical data related to the power load.

3. The method for power load forecasting based on relational enhanced supervised learning according to claim 1, characterized in that: The expression of the enterprise-related graph in step 2 is: G=(V,E,A) Among them, V represents the vertex set, E is the edge set, and A is the adjacency matrix; The expression of the adjacency matrix A is: Among them, R ij is the calculated correlation value between enterprises.

4. The method for power load forecasting based on relational enhanced supervised learning according to claim 2 is characterized in that: The inter-enterprise correlation value is the correlation entropy of the data variables in one period, and the specific expression is: V γ (X,Y)=E[κ γ (XY)] Among them, X and Y are two variables, E represents mathematical expectation, κ γ It is a kernel function with bandwidth γ, and a Gaussian kernel is usually chosen.

5. The method for power load forecasting based on relational enhanced supervised learning according to claim 4 is characterized in that: The κ γ The specific expression is: Among them, γ is the bandwidth of the kernel function, that is, the size of the Gaussian radial basis function, which controls the influence of each data point on other data points in the similarity measurement.

6. The method for power load forecasting based on relational enhanced supervised learning according to claim 1, characterized in that: The neural network model GL-STGNN includes a node embedding module, an encoder, an STMP module and a decoder. The node embedding module is spliced ​​with the sequence information of the node in the current window and input into the encoder. The encoder is mapped to a high-dimensional space, performs preliminary feature extraction, and then inputs into the STMP module. The STMP module is used to propagate information in the graph structure and perform key parts of feature extraction. The decoder is used to process the hidden representation extracted by the model to generate prediction results for the time series.

7. The method for power load forecasting based on relational enhanced supervised learning according to claim 6 is characterized in that: The node embedding module initially defines a trainable parameter representing the embedding vectors of multiple nodes, and adjusts the embedding parameter during training to learn the representation vector of the node. The learned representation vector is handed over to subsequent modules to introduce the local model into the global model and integrate the local effect into the global feature.

8. The method for power load forecasting based on relational enhanced supervised learning according to claim 6, characterized in that: The mathematical expression of the encoder is: in, represents the power load data of the ith node at the t-1th time step, u t-1 Indicates external auxiliary information such as holidays and weather, v i Represents the embedding vector of the node.

9. The method for power load forecasting based on relational enhanced supervised learning according to claim 6, characterized in that: The STMP module includes a TFC module and an MP module. The TFC module uses GRU to capture timing features, and the MP module is used to transmit the captured timing information and update the node status.

10. The method for power load forecasting based on relational enhanced supervised learning according to claim 6, characterized in that: The expression of the decoder is: in, is the representation learned by the i-th node in the STMP module, v i Represents the node embedding vector.

Images