Electric energy meter load decomposition method based on data enhancement and deep learning

Abstract

The invention discloses an electric energy meter load decomposition method based on data enhancement and deep learning. The method comprises the steps of collecting power load data and processing the collected data; constructing a training set based on the processed power load data, performing data enhancement on the training set, dividing samples of the training set into a plurality of class clusters through clustering, and then generating an average sequence as a synthetic sample for each class cluster through a DBA algorithm, thereby improving the number and quality of the samples in the training set; the training set after data enhancement is used for training a load decomposition network model, the model comprises a CNN layer and a GRU layer, and then the trained model is used for load decomposition. By means of data filling, synthetic sample generation based on the training set and the like, the scale and quality of the training set are improved, so that the accuracy of load decomposition is improved.

Description

Technical Field

[0001] The invention belongs to the field of power system load monitoring, and in particular relates to a load decomposition method. Background Art

[0002] With the growth of energy demand and the expansion of power system scale, the demand for power load monitoring and management has become increasingly urgent. Traditional power load decomposition methods usually rely on expensive sensor equipment or invasive monitoring methods, which are costly and difficult to promote and apply. Therefore, non-invasive load decomposition technology has become the key to solving this problem. This technology analyzes the high-frequency aggregated power load signals collected by the energy meter, identifies and decomposes the power load information of each individual appliance or equipment, and realizes intelligent monitoring and management of the power system.

[0003] In recent years, machine learning technology has been widely used in the field of non-intrusive load decomposition. By learning and building models from a large amount of data, machine learning technology can achieve accurate decomposition and identification of power load signals. Current research mainly focuses on deep neural network methods. Specific neural network methods used for load decomposition include forward neural networks, denoising autoencoders, convolutional networks, recurrent networks, generative networks, and attention mechanisms. In 2011, Kolter et al. constructed the first publicly available REDD dataset. In 2019, Davies et al. used 1D-CNN to extract features from raw power signals. Jia et al. proposed a tree CNN to obtain event features from high-dimensional load data. Although deep learning methods have achieved high accuracy in load decomposition, such technologies require a large amount of high-quality load data.

[0004] Currently, the data sets collected for load decomposition are mainly divided into two categories: high-frequency data and low-frequency data. The advantage of using high-frequency data is very obvious, because these data sets retain the entire signal, so that the richest information can be extracted. However, the cost of collecting high-frequency data is a major obstacle, because dedicated infrastructure is required, and the hardware itself and additional installation work are expensive. Therefore, most of the high-frequency data sets currently used have problems such as insufficient data volume and low data quality. Due to the demand for high-quality data sets for deep learning, the accuracy of load decomposition is relatively low. Summary of the invention

[0005] The present invention proposes an electric energy meter load decomposition method based on data enhancement and deep learning, the purpose of which is to solve the problem of low load decomposition accuracy caused by the small amount and poor quality of data sets.

[0006] The technical solution of the present invention is as follows:

[0007] A method for decomposing electric energy meter load based on data enhancement and deep learning, comprising the following steps:

[0008] Step 1: Collect power load data and process the collected data;

[0009] Step 2: construct a training set based on the processed power load data, and perform data enhancement on the training set; the data enhancement refers to dividing the samples of the training set into multiple clusters by clustering, and then generating average sequences as synthetic samples for each cluster by the DBA algorithm, thereby improving the number and quality of samples in the training set;

[0010] Step 3: Use the data-enhanced training set to train a load decomposition network model, wherein the load decomposition network model includes a CNN layer and a GRU layer, wherein the CNN layer is used to extract features and the GRU layer is used to perform load identification;

[0011] Step 4: Input the actually collected power load data into the trained load decomposition network model to obtain the load decomposition result.

[0012] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, in step 1, the outliers in the collected power load data are first deleted, and then the missing values ​​in the power load data are filled by interpolation. The filling method is: let the missing value be P k , k represents the position of the missing value, and its preceding and following values ​​are P k-1 and P k+1 ,but

[0013] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, step 2 specifically includes:

[0014] Step 2-1, using a sliding window to extract a number of subsequences as samples from the processed power load data, wherein the subsequences are power time series, using the extracted samples to construct a data set, and then dividing the data set into a training set and a test set;

[0015] Step 2-2, clustering the training samples in the training set to obtain multiple clusters;

[0016] Step 2-3, generate several average sequences of each type of cluster distribution of the training set through the DBA algorithm as synthetic samples of this type of cluster;

[0017] Step 2-4: Add all synthetic samples to the training set to obtain a data-enhanced training set.

[0018] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, the clustering process in step 2-2 is:

[0019] Step 2-2-1, define the number of clusters K;

[0020] Step 2-2-2, create a cluster center set, randomly select a training sample from the training set as the first cluster center and add it to the cluster center set;

[0021] Step 2-2-3, calculate the average sequence C of all cluster centers in the current cluster center set by using the DBA algorithm;

[0022] Step 2-2-4: All training samples in the training data set that are not selected as cluster centers form a temporary sample set X, calculate the DTW distance between all training samples in the temporary sample set X and the average sequence C, and take the i-th training sample x i The DTW distance between the average sequence C is recorded as D(x i ,C);

[0023] Step 2-2-5: Calculate the probability of each training sample being the next cluster center based on the DTW distance. i The probability P(x i )for:

[0024]

[0025] Step 2-2-6, select one of the training samples in the temporary sample set X as the next cluster center according to the corresponding probability and add it to the cluster center set;

[0026] Step 2-2-7, if the number of cluster centers in the cluster center set is K, execute step 2-2-8 to start iterative updating of cluster centers, otherwise go to step 2-2-3; Step 2-2-8, all training samples in the training data set that are not currently selected as cluster centers constitute a temporary sample set Y, calculate the DTW distance between all training samples in the temporary sample set Y and all cluster centers, and assign the training samples to the cluster of the cluster center corresponding to the minimum DTW distance, thereby obtaining K clusters.

[0027] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, the clustering process further includes:

[0028] Step 2-2-9, calculate the average sequence of each cluster by DBA algorithm, and use the obtained average sequence as the cluster center of each cluster;

[0029] Step 2-2-10: Determine whether the current cluster centers of each cluster meet the preset convergence conditions relative to the previous cluster centers, or whether the number of iterative updates of the cluster centers reaches the preset value. If so, end step 2-2, otherwise jump to step 2-2-8.

[0030] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, in step 2-3, for each cluster, the following method is repeated M times to generate M synthetic samples:

[0031] Step 2-3-1: Randomly select a training sample S from the current cluster * ;

[0032] Step 2-3-2: Calculate other training samples and training sample S in the current cluster * The DTW distance between them is used to obtain N-1 closest training samples, and then the N-1 training samples are compared with the training sample S through the DTW algorithm. * Align, and then compare the aligned N-1 training samples with the training sample S * The sequence set T = {T1, T2, T3, ..., T N}, let the kth training sample T k For S * ;

[0033] Step 2-3-3, randomly sample to obtain N weights;

[0034] Step 2-3-4: Calculate the number of all training samples in the sequence set T to T k DTW distance, and then distribute N weights to each training sample in the sequence set T according to the principle that the closer the DTW distance, the greater the weight, and the farther the DTW distance, the smaller the weight;

[0035] Step 2-3-5: Calculate the current average sequence

[0036]

[0037] Where, T i,l Represents the i-th training sample T in the current sequence set T i The lth sampling point, w i ′ T i The current weight, L is the sampling length of the training samples in the current sequence set T;

[0038] Step 2-3-6: Determine the previous average sequence T k and the current average series Whether the following convergence conditions are met:

[0039]

[0040] In the formula, DTW(T k ,T i ) represents the i-th training sample T in the current sequence set T i With the previous average sequence T k The DTW distance, Represents the i-th training sample T in the current sequence set T i With the current average series DTW distance, ξ is the preset threshold;

[0041] If the convergence condition is met, the current average sequence As the synthetic sample generated this time, end step 2-3; if the convergence condition is not met, set the current sequence set T Then the other N-1 training samples in the sequence set T are compared with T through the DTW algorithm. k Align, and then compare the aligned N-1 training samples with the training sample T k Reorganize the current sequence set T = {T1, T2, T3, ..., T N}, then skip to step 2-3-4.

[0042] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, the load decomposition network model includes a CNN layer, a Flatten layer, a GRU layer and a fully connected layer connected in sequence.

[0043] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, the CNN layer uses multiple convolutional layers, each of which is followed by a pooling layer;

[0044] The convolution calculation formula in the CNN layer is: is the output of the jth neuron in the kth layer; * represents convolution; f(x) is the activation function; c j is the set of all neurons connected to the jth neuron in the kth layer in the k-1th layer; is the output of the i-th neuron in the k-1-th layer; The weight connecting the i-th neuron in the k-1th layer and the j-th neuron in the k-th layer is automatically learned through the training data; is the bias term of the jth neuron in the kth layer;

[0045] The Flatten layer is used to expand the multi-dimensional feature map into a one-dimensional vector.

[0046] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, the calculation process of the GRU layer in a single time step includes:

[0047]

[0048] Among them, x t Enter information for the current moment, h t-1 is the hidden state of the previous moment, h t is the hidden state passed to the next moment, is the candidate hidden state; r t is the reset gate, z t is the update gate, W z , W r , W are weight matrices, [·,·] represents the concatenation of matrices, σ is the sigmoid activation function, and f(·) is the activation function.

[0049] As a further improvement of the electric energy meter load decomposition method based on data enhancement and deep learning, the formula of the activation function f(·) is:

[0050]

[0051] Compared with the prior art, the present invention has the following positive effects:

[0052] 1. The present invention improves the scale and quality of the training set by means of data filling and generating synthetic samples based on the training set, thereby improving the accuracy of load decomposition.

[0053] 2. The load decomposition network model is based on CNN and GRU, has fewer parameters, and has better feature extraction capabilities and efficiency.

[0054] 3. The present invention further improves the activation function of GRU, increases the gradient in the central area, avoids the gradient from saturating too quickly, and thus enables the network to converge faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is an overall flow chart of the method of the present invention;

[0056] Figure 2 It is the flow chart of clustering algorithm;

[0057] Figure 3 Flowchart for generating synthetic sequences;

[0058] Figure 4 This is the architecture diagram of the load decomposition network model based on CNN and GRU. DETAILED DESCRIPTION

[0059] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0060] like Figure 1 , a method for decomposing electric energy meter load based on data enhancement and deep learning, comprising the steps of:

[0061] Step 1: Collect power load data and process the collected data.

[0062] In the process of power load data collection, some force majeure factors such as sensor damage or sudden load failure may cause power load data anomalies, such as outliers that are much higher than the average operating load, or negative values. If such outliers are not processed, the loss function may be abnormal during model training, thereby increasing the decomposition error. Therefore, before training the model, the outliers in the collected power load data must be deleted, and then the missing values ​​in the power load data must be filled by interpolation to avoid the subsequent DTW distance calculation from failing due to missing values.

[0063] The filling method is: set the missing value to P k , k represents the position of the missing value, and its preceding and following values ​​are P k-1 and P k+1 ,but

[0064] Step 2: construct a training set based on the processed power load data, and perform data enhancement on the training set. The data enhancement refers to dividing the samples of the training set into multiple clusters by clustering, and then generating average sequences as synthetic samples for each cluster by the DBA algorithm, thereby improving the number and quality of samples in the training set.

[0065] The DBA algorithm (Dynamic Time Warping Barycenter Averaging) is an algorithm for calculating the average series of multiple time series. It can handle time series of different lengths and take into account the local shapes and global trends between them.

[0066] Step 2 specifically includes:

[0067] Step 2-1: Use a sliding window to extract several subsequences from the processed power load data as samples, wherein the subsequences are power time series. Use the extracted samples to construct a data set, and then divide the data set into a training set and a test set.

[0068] Step 2-2: Cluster the training samples in the training set to obtain multiple clusters.

[0069] like Figure 2 , the clustering process is:

[0070] Step 2-2-1, define the number of clusters K. K is specified by the user and can be determined based on factors such as the number of electrical appliances.

[0071] Step 2-2-2: Create a cluster center set, randomly select a training sample from the training set as the first cluster center and add it to the cluster center set.

[0072] Step 2-2-3: Calculate the average sequence C of all cluster centers in the current cluster center set through the DBA algorithm.

[0073] Step 2-2-4: All training samples in the training data set that are not selected as cluster centers form a temporary sample set X, calculate the DTW distance between all training samples in the temporary sample set X and the average sequence C, and take the i-th training sample x i The DTW distance between the average sequence C is recorded as D(x i ,C).

[0074] DTW (Dynamic Time Warping) is a method for calculating the distance between two time series, and is particularly suitable for processing time series data. The DTW algorithm includes steps such as building a distance matrix, finding the best path, and calculating the total distance. It aligns two time series (different lengths are allowed) according to the optimal path and finds a "curved" path so that the sum of the distances between all points on the path is minimized, thereby reflecting the similarity of the two sequences.

[0075] Step 2-2-5: Calculate the probability of each training sample being the next cluster center based on the DTW distance. i The probability P(x i )for:

[0076]

[0077] Step 2-2-6: Select one of the training samples in the temporary sample set X as the next cluster center according to the corresponding probability and add it to the cluster center set. This step can use the roulette method or the random number generation method.

[0078] Step 2-2-7: If the number of cluster centers in the cluster center set is K, execute step 2-2-8 to start iterative updating of cluster centers, otherwise go to step 2-2-3.

[0079] Step 2-2-8, all the training samples in the training data set that are not currently selected as cluster centers constitute a temporary sample set Y, calculate the DTW distances between all the training samples in the temporary sample set Y and all the cluster centers, and assign the training samples to the cluster of the cluster center corresponding to the minimum DTW distance, thereby obtaining K clusters.

[0080] Step 2-2-9, calculate the average sequence of each cluster through the DBA algorithm, and use the obtained average sequence as the cluster center of each cluster. At this time, the total DTW distance from the training samples in the cluster to the cluster center is the smallest.

[0081] Step 2-2-10: Determine whether the current cluster center of each cluster meets the preset convergence condition (the difference is lower than the threshold) relative to the previous cluster center, or whether the number of iterative updates of the cluster center reaches the preset value. If so, end step 2-2, otherwise jump to step 2-2-8.

[0082] Step 2-3: For each type of cluster distribution in the training set, a number of average sequences are generated through the DBA algorithm as synthetic samples of this type of cluster.

[0083] Specifically, Figure 3 , for each cluster, repeat the following method M times to generate M synthetic samples:

[0084] Step 2-3-1: Randomly select a training sample S from the current cluster * .

[0085] Step 2-3-2: Calculate other training samples and training sample S in the current cluster * The DTW distance between them is used to obtain N-1 closest training samples, and then the N-1 training samples are compared with the training sample S through the DTW algorithm. * Align, and then compare the aligned N-1 training samples with the training sample S * The sequence set T = {T1, T2, T3, ..., T N}, let the kth training sample T k For S * .

[0086] Step 2-3-3: Obtain N weights by random sampling using the Mersenne Twister algorithm.

[0087] Step 2-3-4: Calculate the number of all training samples in the sequence set T to T k Then, N weights are assigned to each training sample in the sequence set T according to the principle that the closer the DTW distance, the greater the weight, and the farther the DTW distance, the smaller the weight.

[0088] Step 2-3-5: Calculate the current average sequence

[0089]

[0090] Where, T i,l Represents the i-th training sample T in the current sequence set T i The lth sampling point, w i ′ T i The current weight, L is the sampling length of the training samples in the current sequence set T.

[0091] Step 2-3-6: Determine the previous average sequence T k and the current average series Whether the following convergence conditions are met:

[0092]

[0093] In the formula, DTW(T k ,T i ) represents the i-th training sample T in the current sequence set T i With the previous average sequence T k The DTW distance, Represents the i-th training sample T in the current sequence set T i With the current average series DTW distance, ξ is the preset threshold.

[0094] If the convergence condition is met, the current average sequence As the synthetic sample generated this time, end step 2-3; if the convergence condition is not met, set the current sequence set T Then the other N-1 training samples in the sequence set T are compared with T through the DTW algorithm. k Align, and then compare the aligned N-1 training samples with the training sample T k Reorganize the current sequence set T = {T1, T2, T3, ..., T N}, then skip to step 2-3-4.

[0095] Step 2-4: Add all synthetic samples to the training set to obtain a data-enhanced training set.

[0096] Step 3: Use the data-enhanced training set to train the load decomposition network model, wherein the load decomposition network model includes a CNN layer and a GRU layer, wherein the CNN layer is used to extract features, and the GRU layer is used to perform load identification.

[0097] Specifically, Figure 4,The load decomposition network model includes a CNN layer, a Flatten layer, a GRU layer and a fully connected layer ,in sequence.

[0098] CNN layers are good at extracting features from local time windows and capturing local patterns in the payload data through convolution operations. CNN layers use multiple convolutional layers, each followed by a pooling layer to gradually reduce the dimension of the features while extracting high-level features to retain important information.

[0099] The convolution calculation formula in the CNN layer is: is the output of the jth neuron in the kth layer; * represents convolution; f(x) is the activation function. The introduction of the activation function can enhance the nonlinear expression ability of the model. In this embodiment, the activation function is the RELU function; c j is the set of all neurons connected to the jth neuron in the kth layer in the k-1th layer; is the output of the i-th neuron in the k-1-th layer; The weight connecting the i-th neuron in the k-1th layer and the j-th neuron in the k-th layer is automatically learned through the training data; is the bias term of the jth neuron in the kth layer.

[0100] After the CNN layer, a Flatten layer is used to expand the multi-dimensional feature map into a one-dimensional vector, making it suitable for passing to the GRU layer.

[0101] The GRU layer is a gated recurrent unit, a variant of the long short-term memory network. It simplifies the internal structure of the LSTM and contains only two gates: the update gate and the reset gate. It can capture the temporal dependencies in the input information. The reset gate determines how to combine the new input information with the previous memory. The update gate is used to control the extent to which the state information of the previous moment is brought into the current state.

[0102] The GRU layer processes the expanded feature vectors and models the long-term and short-term dependencies of the time series. Compared with ordinary RNN, GRU can better solve the problems of gradient vanishing and gradient explosion when processing long data sequences, and can also better mine the more valuable information hidden in the previous and next associations in the data. The input and output structure of GRU is similar to that of ordinary RNN, and its internal ideas are similar to those of LSTM. Compared with LSTM, GRU has one less gate inside and fewer parameters than LSTM, but it can achieve the same functions as LSTM.

[0103] GRU first obtains two gated states through the last transmitted state and the input of the current node, then uses the reset gate to obtain the input data after reset, concatenates it with the current input data and inputs it into the activation function, then enters the update memory stage, using the update gate for forgetting and selective memory.

[0104] In the present invention, the calculation process of the GRU layer in a single time step includes:

[0105]

[0106] Among them, x t Enter information for the current moment, h t-1 is the hidden state of the previous moment. The hidden state acts as the memory of the neural network, which contains the information of the data that the node has seen before. t is the hidden state passed to the next moment, is the candidate hidden state; r t is the reset gate, z t is the update gate, W z , W r , W are weight matrices, [·,·] represents the concatenation of matrices, σ is the sigmoid activation function, through which data can be converted into values ​​in the range of 0-1, f(·) is the activation function, and the formula is:

[0107]

[0108] f(x) increases the gradient in the central area without saturating the gradient too quickly, allowing the network to converge faster.

[0109] The final fully connected layer maps the output of the GRU layer to the load decomposition values.

[0110] The load decomposition network model is trained using the data-augmented training set until the loss function converges to a stable level (no longer changes) and the model's decomposition performance indicator recall>0.5. The training is then terminated, and the test set is then sent to the trained model. The performance of the model is evaluated by comparing the results with the actual annotations.

[0111] Step 4: Input the actually collected power load data into the trained load decomposition network model to obtain the load decomposition result.

[0112] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. A method for decomposing electric energy meter load based on data enhancement and deep learning, characterized in that The steps include: Step 1: Collect power load data and process the collected data; Step 2: construct a training set based on the processed power load data, and perform data enhancement on the training set; the data enhancement refers to dividing the samples of the training set into multiple clusters by clustering, and then generating average sequences as synthetic samples for each cluster by the DBA algorithm, thereby improving the number and quality of samples in the training set; Step 3: Use the data-enhanced training set to train a load decomposition network model, wherein the load decomposition network model includes a CNN layer and a GRU layer, wherein the CNN layer is used to extract features and the GRU layer is used to perform load identification; Step 4: Input the actually collected power load data into the trained load decomposition network model to obtain the load decomposition result.

2. The method for decomposing electric energy meter load based on data enhancement and deep learning according to claim 1, characterized in that: In step 1, the outliers in the collected power load data are first deleted, and then the missing values ​​in the power load data are filled by interpolation. The filling method is: let the missing value be P k , k represents the position of the missing value, and its preceding and following values ​​are P k -1 and P k+1 ,but 3. The method for decomposing electric energy meter load based on data enhancement and deep learning according to claim 1, characterized in that: Step 2 specifically includes: Step 2-1, using a sliding window to extract a number of subsequences as samples from the processed power load data, wherein the subsequences are power time series, using the extracted samples to construct a data set, and then dividing the data set into a training set and a test set; Step 2-2, clustering the training samples in the training set to obtain multiple clusters; Step 2-3, generate several average sequences of each type of cluster distribution of the training set through the DBA algorithm as synthetic samples of this type of cluster; Step 2-4: Add all synthetic samples to the training set to obtain a data-enhanced training set.

4. The method for decomposing electric energy meter load based on data enhancement and deep learning according to claim 3, characterized in that: The clustering process in step 2_2 is: Step 2-2-1, define the number of clusters K; Step 2-2-2, create a cluster center set, randomly select a training sample from the training set as the first cluster center and add it to the cluster center set; Step 2-2-3, calculate the average sequence C of all cluster centers in the current cluster center set by using the DBA algorithm; Step 2-2-4: All training samples in the training data set that are not selected as cluster centers form a temporary sample set X, calculate the DTW distance between all training samples in the temporary sample set X and the average sequence C, and take the i-th training sample x i The DTW distance between the average sequence C is recorded as D(x i , C); Step 2-2-5: Calculate the probability of each training sample as the next cluster center based on the DTW distance. i The probability P(x i )for: Step 2-2-6, select one of the training samples in the temporary sample set X as the next cluster center according to the corresponding probability and add it to the cluster center set; Step 2-2-7, if the number of cluster centers in the cluster center set is K, execute step 2-2-8 to start iterative updating of cluster centers, otherwise go to step 2-2-3; Step 2-2-8, all training samples in the training data set that are not currently selected as cluster centers constitute a temporary sample set Y, calculate the DTW distance between all training samples in the temporary sample set Y and all cluster centers, and assign the training samples to the cluster of the cluster center corresponding to the minimum DTW distance, thereby obtaining K clusters.

5. The method for decomposing electric energy meter load based on data enhancement and deep learning according to claim 4, characterized in that: The clustering process also includes: Step 2-2-9, calculate the average sequence of each cluster by DBA algorithm, and use the obtained average sequence as the cluster center of each cluster; Step 2-2-10: Determine whether the current cluster centers of each cluster meet the preset convergence conditions relative to the previous cluster centers, or whether the number of iterative updates of the cluster centers reaches the preset value. If so, end step 2-2, otherwise jump to step 2-2-8.

6. The method for decomposing electric energy meter load based on data enhancement and deep learning according to claim 4, characterized in that: In steps 2-3, for each cluster, repeat the following method M times to generate M synthetic samples: Step 2-3-1: Randomly select a training sample S from the current cluster * ; Step 2-3-2: Calculate other training samples and training sample S in the current cluster * The DTW distance between them is used to obtain N-1 closest training samples, and then the N-1 training samples are compared with the training sample S through the DTW algorithm. * Align, and then compare the aligned N-1 training samples with the training sample S * The sequence set T = {T1, T2, T3, ..., T N }, let the kth training sample T k For S * ; Step 2-3-3, randomly sample to obtain N weights; Step 2-3-4: Calculate all training samples in sequence set T to T k DTW distance, and then distribute N weights to each training sample in the sequence set T according to the principle that the closer the DTW distance, the greater the weight, and the farther the DTW distance, the smaller the weight; Step 2-3-5: Calculate the current average sequence Where, T i，l Represents the i-th training sample T in the current sequence set T i The lth sampling point, w i ′ is T i The current weight, L is the sampling length of the training samples in the current sequence set T; Step 2-3-6: Determine the previous average sequence T k and the current average series Whether the following convergence conditions are met: In the formula, DTW(T k , T i ) represents the i-th training sample T in the current sequence set T i With the previous average sequence T k The DTW distance, Represents the i-th training sample T in the current sequence set T i With the current average series DTW distance, ξ is the preset threshold; If the convergence condition is met, the current average sequence As the synthetic sample generated this time, end step 2_3; if the convergence condition is not met, set the current sequence set T Then the other N-1 training samples in the sequence set T are compared with T through the DTW algorithm. k Align, and then compare the aligned N-1 training samples with the training sample T k Reorganize the current sequence set T = {T1, T2, T3, ..., T N }, then skip to step 2-3-4.

7. The method for decomposing electric energy meter load based on data enhancement and deep learning according to any one of claims 1 to 6, characterized in that: The load decomposition network model includes CNN layers, Flatten layers, GRU layers and fully connected layers connected in sequence.

8. The method for decomposing electric energy meter load based on data enhancement and deep learning according to any one of claims 1 to 6, characterized in that: The CNN layer uses multiple convolutional layers, each followed by a pooling layer; The convolution calculation formula in the CNN layer is: is the output of the jth neuron in the kth layer; * represents convolution; f(x) is the activation function; c j is the set of all neurons connected to the ,th neuron in the k-1th layer; is the output of the i-th neuron in the k-1-th layer; The weight connecting the i-th neuron in the k-1th layer and the i-th neuron in the k-th layer is automatically learned through training data; is the bias term of the jth neuron in the kth layer; The Flatten layer is used to expand the multi-dimensional feature map into a one-dimensional vector.

9. The method for decomposing electric energy meter load based on data enhancement and deep learning according to any one of claims 1 to 6, characterized in that: The calculation process of the GRU layer in a single time step includes: Among them, x t Enter information for the current moment, h t-1 is the hidden state of the previous moment, ht is the hidden state transferred to the next moment, is the candidate hidden state; r t is the reset gate, z t is the update gate, W z , W r , W are weight matrices, [·, ·] represents the concatenation of matrices, σ is the sigmoid activation function, and f(·) is the activation function.

10. The method for decomposing electric energy meter load based on data enhancement and deep learning as claimed in claim 9, characterized in that The formula of the activation function f(·) is:

Images