A method for repairing user load data considering locality and globality

Abstract

The present application relates to a kind of considering local and global user load data repair method, belong to load data interpolation field, local interpolation model utilizes the local similarity between data points, based on local similarity to carry out interpolation to missing value;Global interpolation model utilizes the feature and distribution information of entire data set, based on global information to carry out interpolation to missing value.Based on the interpolation method of similar data point is simple and easy to use, but is limited by local similarity, lack of global information of data set;The interpolation method based on global model can utilize the feature and distribution information of entire data set, but the calculation complexity is higher, is greatly influenced by extreme data point.By combining the two, first, the local similarity of data is mined using GCN, and local interpolation is carried out, and then the results of local interpolation are used for the adversarial training of GAN, and global interpolation is carried out.Experiment proves that, compared with existing interpolation algorithm, the interpolation method proposed in the present application is more accurate and stable.

Description

Technical Field

[0001] This invention belongs to the field of load data interpolation technology, specifically relating to a method for repairing user load data that considers both local and global factors. Background Technology

[0002] Driven by the new round of power system reform, the retail electricity market has been further liberalized, giving rise to retail electricity companies as a new market entity. Accurate user-level load forecasting is crucial for companies to understand user demand, reduce performance risks associated with deviation assessments, and improve economic efficiency. However, the load of a single user has greater uncertainty than the system load, making traditional statistical forecasting methods difficult to obtain accurate results. Deep learning methods, with their stronger nonlinear fitting and generalization capabilities, can achieve more accurate forecasts. Deep learning models are data-driven, and data quality significantly impacts model performance. However, during the collection, transmission, and conversion of user electricity consumption data by retail electricity companies, issues such as equipment failures, communication equipment malfunctions, and attacks can lead to incomplete data and poor data quality.

[0003] Currently, methods for handling missing data can be divided into direct deletion and imputation. While direct deletion is simple and easy to use, it is only suitable for cases where the proportion of missing values ​​is small. When the proportion is large, a significant amount of crucial and useful information is lost, leading to poor model performance or even model training failure. Imputation methods can be divided into two categories. The first category infers missing values ​​based on similar data points, mainly including methods using simple statistics (such as mean, median, etc.) and k-nearest neighbors (KNN). The second category builds a global model based on information from the entire dataset for imputation, mainly including multiple imputation and generative adversarial networks (GANs). Currently, k-nearest neighbor-based imputation methods are used. This method is simple and easy to use, but the modeling is limited to similar data points and does not build a global model, resulting in lower imputation accuracy. Lagrange interpolation is also used for missing value imputation, mathematically constructing a model to capture local similarities in the data for imputation. A chain-rule-based multiple interpolation method (MICE) is proposed. This method traverses the entire dataset multiple times to obtain data association rules and uses these rules to impute missing values, making it a popular interpolation method. Due to the multi-layered nonlinear structure of deep learning, it has advantages in capturing complex correlations in data and building global models. Generative Adversarial Networks (GANs) are deep learning generative models that can improve the quality of original data by generating samples that resemble the original data and follow the same probability distribution when the original dataset is of poor quality. This forces the reconstructed data to approximate the natural distribution of the original data. Using GANs to reconstruct missing measurement data in a power system has achieved good results.

[0004] The above-mentioned interpolation methods based on similar data points are simple and easy to use, but they are limited by local similarity and lack global information of the dataset. Interpolation methods based on global models can utilize the features and distribution information of the entire dataset, but they have higher computational complexity and are more affected by extreme data points.

[0005] Therefore, at this stage, it is necessary to design a user load data repair method that considers both local and global factors to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a user load data repair method that considers both local and global aspects, in order to solve the technical problems existing in the prior art. First, a graph convolutional network (GCN) is used to mine the potential connections between similar data points to construct a local interpolation model; then, a global interpolation model is constructed through a GAN. The two are combined to improve the accuracy of data interpolation.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A method for repairing user load data that considers both local and global factors includes the following steps:

[0009] S1. First, GCN is used to mine the local similarity of the data and perform local imputation; that is, the local imputation model uses the local similarity between data points to imput missing values ​​based on the local similarity.

[0010] S2. Then, based on the results of local imputation, GAN adversarial training is used to perform global imputation; that is, the global imputation model uses the features and distribution information of the entire dataset to imput missing values ​​based on global information.

[0011] S3. Finally, conduct experimental simulations to verify the effectiveness of combining local interpolation with global interpolation in improving interpolation performance.

[0012] Furthermore, step S1 is detailed as follows:

[0013] (1) Calculate the similarity matrix;

[0014] Each feature vector in the original dataset is represented as a node in the graph. The similarity between nodes is then calculated to form a similarity matrix. This similarity matrix is ​​processed to obtain an adjacency matrix, from which the graph structure data can be derived. The similarity calculation formula based on Euclidean distance is as follows:

[0015]

[0016] In the formula: d represents the Euclidean distance; Let M represent the Hadamard product; M is a binary mask matrix used to indicate whether data is missing. If M = 0, it indicates that data is missing; otherwise, it indicates that data is not missing. i It is the i-th column of matrix M;

[0017] The similarity of missing points is obtained by processing the similarity of their K nearest non-missing points; let the missing points be x. m Its nearest K non-missing points are x ik The similarity expression for missing points calculated using the Gaussian kernel function is as follows:

[0018]

[0019] In the formula: σ is the bandwidth parameter of the Gaussian kernel function;

[0020] This yields a relatively complete similarity matrix S. ij A threshold truncation operation is performed on it to obtain a sparse matrix; for the similarity matrix S ij Sort each row from largest to smallest, and specify a quantile p. Only keep the top p% of values ​​in each row; as shown in the following formula:

[0021]

[0022] (2) Construct a GCN autoencoder;

[0023] An autoencoder consists of an encoder and a decoder. The encoder maps the original input x to a low-dimensional space h = encode(x) for intermediate representation, while the decoder maps the encoded input back to the original dimensional space to reconstruct the input. Reduce x through training The error between them; using a denoising autoencoder; the denoising autoencoder takes the noisy original sample as input and reconstructs the original sample as output; in the local interpolation model, 50% of the input is randomly deleted by using a dropout layer and then used as the input of the denoising autoencoder to complete the reconstruction of the original missing input;

[0024] A local interpolation model is constructed using GCN as the encoder and decoder. In the encoding stage, the graph convolution only involves first-order neighbor nodes, while in the decoding stage, the graph convolution is extended to second-order neighbor nodes. A skip layer and global information are added to the decoder. In the skip layer, only first-order neighbor nodes are considered. A global information vector g is added to the decoder, and the global information is combined with each node in a weighted manner to enhance the expressive power of the node representation.

[0025] The definitions of GCN-based encoders and decoders are as follows:

[0026]

[0027]

[0028] In the formula: It is a symmetric normalized Laplacian matrix that does not consider self-connection; g is the global information vector, which combines global information with each node in a weighted manner;

[0029] Using MSE as the loss function for the autoencoder:

[0030]

[0031] Furthermore, step S2 is detailed as follows:

[0032] Consider introducing adversarial training on top of using GCN for local interpolation to form a global interpolation model; GAN consists of a generator and a discriminator. The generator is responsible for generating fake data, while the discriminator is responsible for distinguishing between the generated fake data and real data. The two are trained adversarially, which ultimately forces the probability distribution of the generated fake data to be close to the probability distribution of the real data; in the global interpolation model, the generator uses a GCN-based autoencoder, and the discriminator uses a multilayer perceptron.

[0033] The global imputation model first uses a generator to perform local imputation to obtain the filled data. Then, X and the data are input into a discriminator, which outputs the probability values ​​of their authenticity. The discriminator and the generator are updated through training feedback. Finally, the corresponding values ​​are used to fill the missing parts of X to complete the global imputation of the data.

[0034] The GAN model is in the form of WGAN-GP; the loss function of WGAN-GP is:

[0035]

[0036] In the formula: E(·) is the expected value; P x (x) and The probability distributions of the real data x and the generated data are respectively. The probability distribution of ; D(·) is the discriminator function; λ is the weight coefficient of the gradient penalty term; where ε is a random number;

[0037] The generator's loss function needs to be modified from L1, becoming:

[0038]

[0039] During the training process, the discriminator weights are updated every 5 times, and the generator is updated once.

[0040] Furthermore, step S3 is as follows:

[0041] User data from multiple enterprises was selected, including only historical load characteristics, with a sampling interval of 1 hour. The training, validation, and test sets were divided in a 7:1:2 ratio. The root mean square error (RMSE) was used as the evaluation metric for the experimental results. RMSE measures the similarity between predicted and actual values. Two missing load scenarios were manually set: random missing and fragment missing. In the random missing scenario, six different random missing rates ranging from 10% to 60% were set. In the fragment missing scenario, nine different fragment missing rates ranging from 1 to 9 days were set. GCIN, KNN, MICE, Mean, and MF baseline models were used for comparison.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] One of the beneficial effects of this solution is that...

[0044] Local imputation models utilize the local similarity between data points to impute missing values. Global imputation models utilize the features and distribution information of the entire dataset to impute missing values ​​based on global information. Imputation methods based on similar data points are simple and easy to use, but are limited by local similarity and lack global information about the dataset. Imputation methods based on global models can utilize the features and distribution information of the entire dataset, but have higher computational complexity and are more affected by extreme data points. By combining the two, we first use GCN to mine local similarity in the data for local imputation, and then use GAN adversarial training based on the results of local imputation for global imputation. Experiments have verified that the imputation method proposed in this invention is more accurate and stable than existing imputation algorithms. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the local interpolation model based on GCN of the present invention.

[0046] Figure 2 This is a schematic diagram of the GAN-based global interpolation model of the present invention.

[0047] Figure 3 This is a schematic diagram of the algorithm flow of the present invention.

[0048] Figure 4 This is a schematic diagram of the random missing interpolation error results of the present invention.

[0049] Figure 5 This is a schematic diagram of the missing interpolation error results of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0051] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0052] A method for repairing user load data that considers both local and global factors includes the following steps:

[0053] S1. First, GCN is used to mine the local similarity of the data and perform local imputation; that is, the local imputation model uses the local similarity between data points to imput missing values ​​based on the local similarity.

[0054] S2. Then, based on the results of local imputation, GAN adversarial training is used to perform global imputation; that is, the global imputation model uses the features and distribution information of the entire dataset to imput missing values ​​based on global information.

[0055] S3. Finally, conduct experimental simulations to verify the effectiveness of combining local interpolation with global interpolation in improving interpolation performance.

[0056] Graph convolutional networks (GCNs) are a type of neural network model proposed in recent years for learning graph-structured data. They are capable of extracting node information and correlations between nodes within graph-structured data. In local interpolation, the original data is transformed into graph-structured data, and GCNs utilize the relationships and similarities between nodes to perform data interpolation and prediction. The principle of the GCN-based local interpolation model is as follows:

[0057] 1) Calculate the similarity matrix

[0058] Each feature vector in the original dataset is represented as a node in the graph. The similarity between these nodes is then calculated to form a similarity matrix. This similarity matrix is ​​processed to obtain the adjacency matrix, from which the graph structure data can be derived. The similarity calculation formula based on Euclidean distance is as follows:

[0059]

[0060] In the formula: d represents the Euclidean distance; Let M represent the Hadamard product; M is a binary mask matrix used to indicate whether data is missing. If M = 0, it indicates that data is missing; otherwise, it indicates that data is not missing. i It is the i-th column of matrix M.

[0061] However, dataset X contains missing data points. The similarity of these missing data points cannot be directly calculated using equation (1). Therefore, it is necessary to process the similarity of the missing data points by considering the similarity of their K nearest non-missing data points. Let the missing data point be x. m Its nearest K non-missing points are x ik The similarity expression for missing points calculated using the Gaussian kernel function is as follows:

[0062]

[0063] In the formula: σ is the bandwidth parameter of the Gaussian kernel function.

[0064] This yields a relatively complete similarity matrix S. ij However, this matrix is ​​dense, computationally expensive, and difficult to apply directly to graph convolution. A thresholding operation is needed to obtain a sparse matrix. In this paper, the similarity matrix S... ij Sort each row from largest to smallest, specifying a quantile p, and retain only the top p% of values ​​in each row. As shown in the following formula:

[0065]

[0066] 2) Constructing a GCN autoencoder

[0067] An autoencoder consists of an encoder and a decoder. The encoder maps the original input x to a low-dimensional space h = encode(x) for intermediate representation, while the decoder maps the encoded input back to the original dimensional space to reconstruct the input. By continuously reducing x through training The error between the two is reduced to improve the accuracy of the reconstruction. However, since the data is missing and unknown during the training phase, it cannot be directly used to train the autoencoder. Therefore, a denoising autoencoder (DAE) is needed. The denoising autoencoder takes noisy original samples (corrupted original samples) as input and reconstructs the original samples as output. In the local interpolation model, 50% of the input is randomly removed using a dropout layer before being used as input to the DAE to complete the reconstruction of the original missing input.

[0068] like Figure 1As shown, a local interpolation model is constructed using a GCN as both encoder and decoder. In the encoding stage, graph convolutions only involve first-order neighbor nodes. However, in the decoding stage, to acquire more information from neighbor nodes and improve reconstruction accuracy, graph convolutions are extended to second-order neighbor nodes. Furthermore, to further improve the quality of data interpolation, a skip layer and global information are added to the decoder. The graph convolution operations performed in the skip layer are similar to those in the encoder, except that the graph convolutions in the skip layer do not involve the node itself. Therefore, the output of the skip layer only contains information about neighbor nodes and not about the current node itself. Thus, the skip layer helps force the model to learn the similarity between adjacent nodes, preventing the autoencoder from learning the identity function, thereby improving the model's understanding and reconstruction capabilities. To increase the contribution of the most similar node, only first-order neighbor nodes are considered in the skip layer. Since GCNs typically focus more on node and edge-level information, adding global information is beneficial for improving the expressive power of graph neural networks. Global information usually refers to the statistical information of the entire dataset, such as the mean and mode. This paper uses the mean as global information and adds a global information vector g to the decoder. The global information is combined with each node in a weighted manner to enhance the expressive power of the node representation.

[0069] The definitions of GCN-based encoders and decoders are as follows:

[0070]

[0071]

[0072] In the formula: is a symmetric normalized Laplacian matrix that does not consider self-connection; g is a global information vector that combines global information with each node in a weighted manner.

[0073] Using MSE as the loss function for the autoencoder:

[0074]

[0075] Due to the outstanding performance of GANs in data reconstruction, we consider introducing adversarial training on top of using GCNs for local imputation to create a global imputation model. A GAN consists of a generator and a discriminator. The generator is responsible for generating fake data, while the discriminator is responsible for distinguishing between the generated fake data and real data. The two are trained adversarially, ultimately forcing the probability distribution of the generated fake data to approximate the probability distribution of the real data. In the global imputation model, the generator uses a GCN-based autoencoder, and the discriminator uses a multi-layer perceptron (MLP).

[0076] The structure of the global interpolation model is as follows: Figure 2 As shown, the model first performs local imputation through the generator to obtain the filled data. Then, X and are input into the discriminator, which outputs the probability values ​​of their authenticity. The discriminator and the generator are updated through training feedback. After multiple rounds of adversarial training, the filled data is very close to the probability distribution of the real incomplete data X. Then, the corresponding values ​​are used to fill the missing parts of X to complete the global imputation of the data.

[0077] Traditional GANs suffer from mode collapse and gradient vanishing problems during training. Later, researchers proposed WGAN using Wasserstein distance to address these issues. However, WGAN uses weight clipping to limit the discriminator's network parameters, leading to uneven parameter distribution and training instability. WGAN-GP, building on WGAN, uses gradient penalty instead of weight clipping, improving training stability and making it easier to generate high-quality samples.

[27] Therefore, the GAN model used in this paper is the WGAN-GP form. The loss function used in this paper for WGAN-GP is:

[0078]

[0079] In the formula: E(·) is the expected value; P x (x) and The probability distributions of the real data x and the generated data are respectively. The probability distribution of ; D(·) is the discriminator function; λ is the weight coefficient of the gradient penalty term; where ε is a random number.

[0080] Furthermore, in order to improve the generator's ability to fool the discriminator while minimizing the reconstruction error, the generator's loss function needs to be modified from the L1 function above, becoming:

[0081]

[0082] During training, the discriminator's weights are updated every 5 times, while the generator is updated once. Training the discriminator aims to enable it to more accurately distinguish between generated and real data, and also to avoid excessive noise from training the generator too many times, thus making the training process more stable and improving the quality of the generated data.

[0083] like Figure 3 As shown, the algorithm flow is as follows:

[0084] The original data is represented as graph-structured data, and local interpolation is performed using a GCN autoencoder.

[0085] We utilize the GCN autoencoder as the generator in GAN.

[0086] GCN updates based on the training feedback of GAN to improve the interpolation effect and form a global interpolation model.

[0087] Case Study:

[0088] User data from 31 enterprises in a certain region of a province / city in China, collected from May 1, 2020 to March 31, 2022, was selected. Only historical load characteristics were included, with a sampling interval of 1 hour. The training, validation, and test sets were divided in a 7:1:2 ratio. The root mean square error (RMSE) was used as the evaluation metric for the experimental results. RMSE measures the closeness between predicted and actual values; a smaller value indicates better prediction performance. In the imputation experiment, two common missing data scenarios were manually set: random missing data and fragment missing data. In the random missing data scenario, six different random missing rates ranging from 10% to 60% were set; in the fragment missing data scenario, nine different fragment missing rates ranging from 1 to 9 days were set. In the interpolation experiments, to compare the interpolation results, several baseline models were used for comparison, including GCIN (a local interpolation model without adversarial training), KNN (an interpolation algorithm based on k-nearest neighbors), MICE (a multi-interpolation algorithm based on chain rules), Mean (mean interpolation), and MF (an interpolation algorithm based on matrix factorization).

[0089] Imputation results for random missing scenes:

[0090] From Table 1 and Figure 4 It can be seen that, in the random missing data scenario, the AGCIN model (the model proposed in this paper) exhibits the best imputation performance under different missing rates, and the performance advantage of the AGCIN model becomes more and more obvious as the missing rate increases. The GCIN model, which only performs local imputation without adversarial training, performs slightly worse than the AGCIN model, but is still better than other imputation methods. This shows that graph convolutional neural networks can effectively learn the features and correlations of data in local imputation tasks, and that adding GANs for global imputation can effectively improve imputation accuracy, verifying the effectiveness of combining local and global imputation in improving imputation performance.

[0091] Table 1 Comparison of Random Missing Imputation Errors

[0092]

[0093]

[0094] Imputation results for scenes with missing fragments:

[0095] Table 2 Comparison of Fragment Missing Imputation Errors

[0096]

[0097] From Table 2 and Figure 5 It can be seen that, in fragment missing scenarios, AGCIN and GCIN models still exhibit optimal and suboptimal imputation performance for different numbers of missing days. However, compared to random missing scenarios, the performance difference between AGCIN and GCIN models decreases. This is because random missing data is more random and discontinuous, while fragment missing data is continuously missing, and the correlation and trends between data are more obvious. The standalone local imputation (GCIN) model can already effectively utilize the inherent relationships between data to achieve good imputation performance. Although in continuous missing scenarios, the standalone local imputation model can already effectively utilize the local correlations between data, adding global imputation via GAN can still slightly improve imputation performance.

[0098] Performance Analysis:

[0099] Experiments conducted in two scenarios—random missing data and fragment missing data—validated the effectiveness of combining local and global imputation in improving imputation accuracy. Experimental results show that in the fragment missing data scenario, the correlation and trends between data points are more pronounced, and local imputation can effectively utilize the inherent correlations within the data. In the random missing data scenario, the randomness and discontinuity of missing data are stronger, and the performance of the local imputation model is relatively limited. Introducing a GAN for global imputation, and combining local and global imputation, can better capture the overall data distribution and characteristics, further improving the accuracy of the imputation results. Therefore, combining local and global imputation is effective in improving imputation accuracy, fully utilizing the local correlations and global distribution characteristics of the data to obtain more accurate missing value estimates.

[0100] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for repairing user load data considering both local and global, characterized in that, Comprise the following steps: S1, first using GCN to mine the local similarity of data, local interpolation; that is, the local interpolation model uses the local similarity between data points, and the missing values are supplemented based on the local similarity; S2, based on the results of local interpolation, the global interpolation is carried out using the adversarial training of GAN; that is, the global interpolation model uses the feature and distribution information of the whole data set, and the missing values are supplemented based on the global information; S3, finally, the effectiveness of combining local interpolation and global interpolation to improve the interpolation performance is verified through experiment simulation; Step S1 is as follows: (1) Calculate the similarity matrix; Each feature vector of the original data set is represented as a node in the graph, and then the similarity between each node is calculated to form a similarity matrix. The adjacency matrix is obtained after processing the similarity matrix, and the data of the graph structure is obtained according to the adjacency matrix. The similarity calculation formula based on Euclidean distance is as follows: S ij = d(x i ☉(M i ☉M j ),x j ☉(M i ☉M j )) In the formula, d represents the Euclidean distance; represents Hadamard product; M is a binary mask matrix, used to represent whether the data is missing, if M = 0, it represents that the data is missing, otherwise it represents that the data is not missing, M i is the i-th column of the matrix M; The similarity of the missing point is obtained by processing the similarity of its K nearest non-missing points; assuming that the data missing point is x m The K nearest non-missing points of x are x ik The similarity expression of the missing point is calculated according to the Gaussian kernel function In the formula: σ is the bandwidth parameter of the Gaussian kernel function; Thus, a relatively complete similarity matrix S is obtained ij A thresholding operation is performed on S to obtain a sparse matrix; each row of S is sorted in descending order, and a quantile p is specified, and only the top p% values are retained for each row; as follows: ij A thresholding operation is performed on S to obtain a sparse matrix; each row of S is sorted in descending order, and a quantile p is specified, and only the top p% values are retained for each row; as follows: (2) Construct GCN autoencoder; The autoencoder consists of an encoder and a decoder, the encoder is used to map the original input x into a low-dimensional space h = encode(x) for intermediate representation, and the decoder is used to map the encoded input into the original dimensional space for input reconstruction: By training to reduce the error between x and ; using a denoising autoencoder; the denoising autoencoder receives a noisy original sample as input and reconstructs the original sample as output; in the local interpolation model, by using a dropout layer to randomly delete 50% of the input as the input of the denoising autoencoder, the reconstruction of the original input with missing values is completed; The GCN is used as the encoder and decoder to construct the local interpolation model. The graph convolution in the encoding stage only involves 1-order neighbor nodes, and the graph convolution in the decoding stage is extended to 2-order neighbor nodes. A skip layer and global information are added in the decoder part. Only 1-order neighbor nodes are considered in the skip layer part. A global information vector g is added in the decoder to combine the global information with each node through weighting, enhancing the expression ability of node representation. The definitions of the GCN-based encoder and decoder are as follows: wherein: is the symmetric normalized Laplacian matrix without self-connections; g is the global information vector, which combines the global information with each node in a weighted manner; MSE is used as the loss function of the autoencoder: Step S2 is as follows: The adversarial training is introduced based on the local interpolation using GCN to form a global interpolation model. GAN consists of a generator and a discriminator. The generator is responsible for generating fake data, while the discriminator is responsible for distinguishing between the generated fake data and the real data. Through adversarial training, the probability distribution of the generated fake data is finally forced to approach the probability distribution of the real data. In the global interpolation model, the generator uses the GCN-based autoencoder, and the discriminator uses the multilayer perceptron. The global interpolation model first performs local interpolation through the generator to obtain the filled data, and then inputs X and into the discriminator. The discriminator outputs the probability value of their authenticity. The discriminator and the generator are updated through training feedback, and then the corresponding value is used to fill the missing part of X to complete the global interpolation of the data. The GAN model is in the form of WGAN-GP; the loss function of WGAN-GP is: where E(·) is the mathematical expectation; P x (x) and are the probability distribution of real data x and the probability distribution of generated data respectively; D(·) is the function of the discriminator; λ is the weight coefficient of the gradient penalty term; where ε is a random number; The loss function of the generator needs to be changed based on L1 to: During the training process, the weight of the discriminator is set to update every 5 times, and the generator is updated once.

2. The method of claim 1, wherein the method is characterized by, Step S3 is as follows: The user data of a plurality of enterprises is selected, only historical load characteristics are contained, and a sampling interval is 1h; a training set, a verification set and a test set are divided in a ratio of 7:1:2; a root mean square error (RMSE) is used as an evaluation index of an experimental result; the RMSE is used for measuring the closeness between a predicted value and an actual value; wherein two kinds of missing scenarios of load data are manually set, and are respectively a random missing scenario and a segment missing scenario; in the random missing scenario, six different random missing rates from 10% to 60% are set; in the segment missing scenario, nine different segment missing from 1 day to 9 days are set; GCIN, KNN, MICE, Mean and MF baseline models are used as comparison.

Images