A non-intrusive load identification method and device

Through deep neural network logistic regression Stacking modeling combined with LSTM and DAE models, the electrical load is decomposed, which solves the problem of insufficient universality and universality of load decomposition in the existing technology, and achieves higher precision load recognition and decomposition effects .

CN115758249BActive Publication Date: 2024-11-19INFORMATION & COMM BRANCH OF STATE GRID JIANGSU ELECTRIC POWER
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211482460.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-11-24
Publication Date
2024-11-19
Estimated Expiration
2042-11-24

AI Technical Summary

Technical Problem

The existing non-invasive load monitoring systems are not universal and versatile in electrical load decomposition. The single model has limited decomposition effect on different electrical appliances, especially for electrical appliances with faster variable frequency.

Method used

The non-invasive load identification algorithm modeled by deep neural network (DNN) logistic regression Stacking modeling combines long and short-term memory network (LSTM) and denoising autoencoder network (DAE) to initially decompose the target load, and passes through a fully connected neural network The decomposition results are polymerized to form the final load decomposition results.

Benefits of technology

It improves the accuracy of load decomposition, can more accurately identify and decompose load information of different electrical appliances, and enhances the universality and versatility of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115758249B_ABST
    Figure CN115758249B_ABST
Patent Text Reader

Abstract

The present invention discloses a non-intrusive load identification method and device, which adopts a stacking integrated algorithm to optimize a non-intrusive load monitoring model, performs a preliminary decomposition of the load by a base model, and realizes accurate load decomposition by a meta-model. The method belongs to the field of power monitoring technology. It includes the following steps: (1) data acquisition, collecting residential electricity consumption information; (2) data preprocessing, constructing an original data set and dividing the data subsets; (3) constructing a long short-term memory network (LSTM) load decomposition model as the first base learner of the load identification system; (4) constructing a denoising autoencoder network (DAE) as the second base learner of the system; (5) constructing an aggregation model to output the final load identification result. The method can quickly and accurately classify the voltage signals generated by different working states of different electrical appliances, so as to facilitate subsequent processing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a non-intrusive load identification method and device, belonging to the technical field of non-intrusive load monitoring. Background Art

[0002] As the core of the interactive service system between the power grid and the user side, smart electricity consumption is an important step to strengthen the national smart grid. One of its key technologies is the advanced metering infrastructure (AMI) standard, system and terminal technology. The load monitoring system is an important part of AMI and the first step to realize the smart grid. At present, the electric meter can only read the power consumption information of the user's bus, but cannot obtain the load power consumption information on the user side. The load monitoring system is the key to breaking through this bottleneck problem.

[0003] At present, load monitoring systems are divided into two categories according to technical forms: intrusive load monotoring (ILM) and non-intrusive load monotoring (NILM). NILM was first proposed by Professor Hart in the 1980s. NILM is classified as a combinatorial optimization problem. It uses the time series data of active power and reactive power obtained at a sampling frequency of 1Hz, and uses clustering algorithms to decompose the load. In recent years, scholars at home and abroad have conducted corresponding research on NILM. The literature [Non-intrusive household load decomposition method based on genetic optimization] uses genetic optimization algorithms to combine the current waveforms, power and current frequency domain characteristics of various types of electrical appliances to realize the recognition of the status of different electrical appliances. It only studies the identification of changes in the working status of a single electrical appliance, and lacks research on the decomposition of power data of a single electrical appliance. The paper [Denoising autoencoders for Non-Intrusive Load Monitoring: Improvements and comparative evaluation] proposed a model of deep neural network based on denoising autoencoder, and compared it with the maximum a posteriori algorithm model in the field of hidden Markov model, showing the advanced application of deep learning in the field of NILM and its robustness to noise interference. The paper [Non-Intrusive Load Monitoring Based on Attention and Residual Network] introduced the attention mechanism and residual network for load decomposition under the framework of seq2seq. The sampling interval of the selected data set is short, and the length of the sampling input of different target electrical appliances varies greatly, so the actual application is more complicated. The paper [Improving Non-Intrusive Load Disaggregation through an Attention-Based Deep Neural Network] adopts a method combining regression subclass model and classification subclass model, and adds attention mechanism to the regression submodel to improve the effect of load decomposition, but the decomposition effect is poor for electrical appliances with fast changing frequency such as capacitors and inductors.

[0004] Research institutions at home and abroad have achieved good research results. The above NILM models have their own advantages in different electrical appliance decomposition, but their corresponding disadvantages in the decomposition of other electrical appliances are also obvious. In terms of non-intrusive load identification algorithms, the current non-intrusive power load intelligent identification systems are generally based on a single model. Commonly used models include hidden Markov models, graph models, and deep learning models. A single model often only has a good identification effect on certain categories of electrical loads, and its universality and versatility need to be improved. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a non-intrusive load identification method and device, a non-intrusive load identification algorithm based on deep neural network (DNN) logistic regression Stacking modeling, using the base model DAE model and LSTM model to perform preliminary decomposition of the target load, and then using a fully connected neural network to aggregate the decomposition results of the base model to obtain the final decomposition result. This method realizes non-intrusive load decomposition.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a non-intrusive load identification method, comprising the following steps:

[0008] Acquire data to be identified;

[0009] Preprocessing the data to be identified, constructing an original data set and dividing it into several data subsets;

[0010] Inputting the data subset into a Stacking ensemble learning model to obtain a final power decomposition sequence of the target load;

[0011] The method for constructing the Stacking ensemble learning model includes:

[0012] Construct a long short-term memory network (LSTM) load decomposition model as the first base learner of the first layer of the Stacking ensemble learning model;

[0013] Construct a denoising autoencoder network (DAE) as the second base learner in the first layer of the Stacking ensemble learning model;

[0014] Construct a fully connected layer as the meta-learner of the second layer of the Stacking ensemble learning model.

[0015] Furthermore, the data to be identified is obtained, including:

[0016] Collect high-frequency loads and low-frequency loads.

[0017] Further, preprocessing the data to be identified includes:

[0018] Median filtering and data downsampling, wherein the data downsampling is used to reduce the sampling frequency, and the median filtering is used to eliminate interference in the sampled data.

[0019] Furthermore, the data is divided into subsets, including:

[0020] For the data set S = {(y n ,x n),n=1,...,N}, where x n is the feature vector of the nth sample, y n is the predicted value corresponding to the nth sample, p is the number of features contained, that is, each feature vector is (x 1 ,x 2 ,...,x p );

[0021] Randomly divide the data into K subsets S of roughly equal size 1 ,S 2 ,...,S K ;

[0022] Where S -k =SS k , respectively define S k and S -k is the k-fold test set and training set in K-fold cross validation;

[0023] For the first-layer prediction algorithm, there are K base learners, and for the training set S -k Use the kth algorithm to train the base model L k , k=1,...,K.

[0024] Furthermore, the input of the long short-term memory network (LSTM) load decomposition model is a data subset, and the output is a LSTM target load decomposition sequence;

[0025] The LSTM load decomposition model consists of a stack of input layer, output layer and hidden layer. Each unit has a cell tuple, and its state at time t is recorded as c t Through the input gate i t 、Forget gate t and output gate o t Receive the current state x t and the state h of the tuple at the previous moment t-1 ; The state of the memory cell c t-1 It will be input into each gate as internal information.

[0026] After receiving input information, the input gate, forget gate, and output gate will perform internal operations to determine whether to activate the cell tuple.

[0027] The signal of the input gate is transformed by a nonlinear function and then superimposed with the memory cell state processed by the forget gate to form a new memory cell state c. t .

[0028] Memory cell state c t The output h of the unit is formed by the operation of nonlinear functions and the dynamic control of the output gate. t .

[0029] Further, the input of the denoising autoencoder network is a data subset, and the output is a DAE target load decomposition sequence;

[0030] The denoising autoencoder network is a symmetrical structure divided into two parts: an encoder and a decoder. When constructing the network, it is only necessary to set the number of neurons in the hidden layers of the encoder and decoder networks. During the training process, the network will adjust its own parameters.

[0031] In the coding layer, the original data is mapped to the hidden feature space after the operation of the convolution layer, and then mapped to the label space of the target load through the fully connected layer to generate a denoised hidden layer. The decoder then reverses this process to obtain the power curve of the target load.

[0032] The one-dimensional convolution layer essentially creates multiple filters to extract data features.

[0033]

[0034] Where: is the jth feature map of the l-1th layer, is the convolution kernel function of the j-th feature map and the l-1-th feature map of the l-th layer, The jth bias parameter of the lth layer, f(x) is the nonlinear activation function of the convolutional neural network

[10] The nonlinear activation function used in this convolutional layer is the rectified linear unit (ReLU), which can be obtained

[0035]

[0036] Where: is the output value of the convolution operation, for The activation value of .

[0037] Furthermore, the data subset is input into the Stacking ensemble learning model to obtain the final power decomposition sequence of the target load, including:

[0038] The original data set is divided into several sub-data sets, which are input into each base learner of the first-layer prediction model. Each base learner outputs its own prediction result to obtain the output of the first layer.

[0039] The output of the first layer is used as the input of the second layer, and the preliminary decomposition results of the two base learners are aggregated to obtain the final decomposition result of the target load, that is, the LSTM target load power decomposition sequence and the target load power decomposition sequence are aggregated, and the decomposition results of the base model are aggregated using a fully connected neural network to obtain the final power decomposition sequence of the target load.

[0040] Furthermore, the training method of the Stacking ensemble learning model includes:

[0041] For the data set S = {(y n ,x n ),n=1,...,N}, where x n is the feature vector of the nth sample, y n is the predicted value corresponding to the nth sample, p is the number of features contained, that is, each feature vector is (x 1 ,x 2 ,...,x p ).

[0042] Randomly divide the data into K subsets S of roughly equal size 1 ,S 2 ,...,S K . Where S -k =SS k , respectively define S k and S -k is the k-th fold test set and training set in K-fold cross validation.

[0043] For the first-layer prediction algorithm, there are K base learners, and for the training set S -k Use the kth algorithm to train the base model L k , k=1,...,K.

[0044] For the k-th fold test set S in K-fold cross validation k For each sample x in n , base learner L k Its prediction is denoted as z kn .

[0045] After completing the cross-validation process, the output data of the K base learners constitute a new data sample, namely: S new ={(y n ,z 1n ,...,z kn )}, n=1,...,N. The newly generated data set is the second layer input data of Stacking. The meta-learner L is obtained by summarizing these data using the second layer prediction algorithm. new .

[0046] In a second aspect, the present invention provides a non-intrusive load identification device, the device comprising:

[0047] Input module: used to obtain the data to be identified;

[0048] Preprocessing module: used to preprocess the data to be identified, construct the original data set and divide it into several data subsets;

[0049] Decomposition module: used for inputting the data subset into the Stacking ensemble learning model to obtain the final power decomposition sequence of the target load;

[0050] The method for constructing the Stacking ensemble learning model includes:

[0051] Construct a long short-term memory network (LSTM) load decomposition model as the first base learner of the first layer of the Stacking ensemble learning model;

[0052] Construct a denoising autoencoder network (DAE) as the second base learner in the first layer of the Stacking ensemble learning model;

[0053] Construct a fully connected layer as the meta-learner of the second layer of the Stacking ensemble learning model.

[0054] In a third aspect, the present invention provides a non-intrusive load identification device, characterized in that it includes a processor and a storage medium;

[0055] The storage medium is used to store instructions;

[0056] The processor is used to operate according to the instructions to execute the steps of the method according to the first aspect.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. Based on the idea of ​​deep neural network (DNN) logistic regression Stacking modeling, the present invention uses DAE model and LSTM model to perform preliminary load decomposition on the original bus load power sequence, and obtains the preliminary power sequence of the target load. The fully connected layer is introduced to aggregate the preliminary decomposition results of the two models to obtain the final decomposition result of the target load, which greatly improves the accuracy of load decomposition.

[0059] 2. The idea of ​​the denoising autoencoder model (DAE model) is to regard the NILM problem as a denoising problem. The idea comes from natural language processing, that is, the power curve of the target load is regarded as the target sound of speech recognition, and the power curve of the bus is the audio mixed with noise. The purpose of the algorithm is to remove the noise to obtain the power curve of the target load, providing a new idea for non-invasive load decomposition. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a system block diagram of the present invention;

[0061] Figure 2 It is a network framework diagram of the denoising autoencoder in the present invention;

[0062] Figure 3 It is a diagram of the Stacking ensemble learning optimization framework in the present invention;

[0063] Figure 4 is an original signal diagram of the first embodiment of the present invention;

[0064] Figure 5 This is the load decomposition example 1 of the first embodiment of the present invention;

[0065] Figure 6 This is load decomposition example 2 of embodiment 1 of the present invention. DETAILED DESCRIPTION

[0066] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0067] Embodiment 1:

[0068] This embodiment provides a non-intrusive load identification method based on deep neural network (DNN) logistic regression stacking modeling. Figure 1 As shown, the following steps are included:

[0069] 1) Data collection: collecting residents’ electricity consumption information;

[0070] 2) Data preprocessing: constructing the original data set and dividing the data into subsets;

[0071] 3) Construct a long short-term memory network (LSTM) load decomposition model as the first base learner of the load identification system;

[0072] 4) Construct a denoising autoencoder network (DAE) as the second base learner of the system;

[0073] 5) Build an aggregation model and output the final load identification result, that is, the final power decomposition sequence of the target load.

[0074] Specifically, the residential electricity consumption data information collected is:

[0075] Collect high-frequency loads and low-frequency loads.

[0076] Specifically, the steps of data preprocessing are:

[0077] Median filtering and data downsampling, wherein the data downsampling is used to reduce the sampling frequency, and the median filtering is used to eliminate interference in the sampled data.

[0078] Specifically, the steps of dividing the data subsets into K subsets of substantially equal size are:

[0079] For the data set S = {(y n ,x n ),n=1,...,N}, where x n is the feature vector of the nth sample, y n is the predicted value corresponding to the nth sample, p is the number of features contained, that is, each feature vector is (x 1 ,x 2 ,...,x p ). Randomly divide the data into K subsets S of roughly equal size 1 ,S 2 ,...,S K . Where S -k =SS k , respectively define S k and S -k is the k-fold test set and training set in K-fold cross validation. For the first-layer prediction algorithm, there are K base learners. -k Use the kth algorithm to train the base model L k , k=1,...,K.

[0080] Specifically, a long short-term memory network (LSTM) load decomposition model is constructed:

[0081] The network is generally composed of a stack of input layer, output layer and hidden layer. Each unit has a cell tuple, which is a unit with memory function. Its state at time t is recorded as c t Through the input gate i t 、Forget gate t and output gate o t Receive the current state x t and the state h of the tuple at the previous moment t-1 The state of the memory cell c t-1 As internal information, it will be input into each gate. After receiving the input information, the input gate, forget gate and output gate will perform internal operations to determine whether to activate the cell tuple. After the signal of the input gate is transformed by a nonlinear function, it is superimposed with the memory cell state processed by the forget gate to form a new memory cell state c t . Memory cell state c t The output h of the unit is formed by the operation of nonlinear functions and the dynamic control of the output gate. t .

[0082] Specifically, construct a denoising autoencoder network (DAE):

[0083] The idea of ​​the denoising autoencoder model (DAE model) is to regard the NILM problem as a denoising problem. The idea comes from natural language processing, that is, the power curve of the target load is regarded as the target sound of speech recognition, and the power curve of the bus is the audio mixed with noise. The purpose of the algorithm is to remove the noise and obtain the power curve of the target load.

[0084] The DAE applied in NILM is essentially an autoencoder, whose goal is to reconstruct the power curve of the target load from the bus power curve containing noise. Figure 2 As shown in the figure, in terms of network structure, DAE is a symmetrical structure divided into two parts: encoder and decoder. When building the network, you only need to set the number of neurons in the hidden layers of the encoder and decoder networks. During the training process, the network will adjust its own parameters.

[0085] In the coding layer, the original data is mapped to the hidden feature space after the operation of the convolution layer, and then mapped to the label space of the target load through the fully connected layer to generate a denoised hidden layer. The decoder then reverses this process to obtain the power curve of the target load.

[0086] The one-dimensional convolution layer essentially creates multiple filters to extract data features.

[0087]

[0088] Where: is the jth feature map of the l-1th layer, is the convolution kernel function of the j-th feature map and the l-1-th feature map of the l-th layer, The jth bias parameter of the lth layer, f(x) is the nonlinear activation function of the convolutional neural network

[10] The nonlinear activation function used in this convolutional layer is the rectified linear unit (ReLU), which can be obtained

[0089]

[0090] Where: is the output value of the convolution operation, for The activation value of .

[0091] Specifically, the model aggregation method:

[0092] According to the generation method of individual learners, the current ensemble learning methods can be roughly divided into two categories: the first category is the serialization method in which there is a strong dependency between individual learners and they must be generated serially, represented by Boosting; the second category is the parallelization method in which there is no strong dependency between individual learners and they can be generated simultaneously, represented by Bagging and Stacking.

[0093] like Figure 3 As shown in the figure, the Stacking ensemble learning model first divides the original data set into several sub-data sets and inputs them into the base learners of the first-layer prediction model. Each base learner outputs its own prediction result. Then, the output of the first layer is used as the input of the second layer to train the meta-learner of the second-layer prediction model, and the model in the second layer outputs the final prediction result. The Stacking learning framework generalizes the output results of multiple models to improve the overall prediction accuracy.

[0094] The training method of stacking ensemble learning is: for the data set S = {(y n ,x n ),n=1,...,N}, where x n is the feature vector of the nth sample, y n is the predicted value corresponding to the nth sample, p is the number of features contained, that is, each feature vector is (x 1 ,x 2 ,...,x p ). Randomly divide the data into K subsets S of roughly equal size 1 ,S 2 ,...,S K . Where S -k =SS k , respectively define S k and S -k is the k-fold test set and training set in K-fold cross validation. For the first-layer prediction algorithm, there are K base learners. -k Use the kth algorithm to train the base model L k , k = 1, ..., K. For the k-th fold test set S in K-fold cross validation k For each sample x in n , base learner L k Its prediction is denoted as z kn After completing the cross-validation process, the output data of the K base learners constitute a new data sample, namely: S new ={(y n ,z 1n ,...,z kn)}, n=1,...,N. The newly generated data set is the second layer input data of Stacking. The meta-learner L is obtained by summarizing these data using the second layer prediction algorithm. new The stacking configuration allows the training results of the first-layer algorithm to be fully used in the induction process of the second-layer algorithm. The second-layer algorithm can discover and correct the prediction errors in the first-layer learning algorithm to improve the accuracy of the model.

[0095] In step 5), the final model is built and the results are output:

[0096] Based on the original bus load power sequence, the DAE model and LSTM model are used to perform preliminary load decomposition to obtain the preliminary power sequence of the target load. The fully connected layer is introduced to aggregate the preliminary decomposition results of the two models to obtain the final decomposition result of the target load, that is, the corresponding output results of the two models are convolved using a 1×1×2 convolution kernel, and the final aggregation result is output after the activation function.

[0097] In specific implementation, Figure 4 Taking the load decomposition of the original signal graph as an example, the specific results are:

[0098] Figure 5 Example 1 and 2 of the load decomposition shown Figure 6 This is the second load decomposition example of the first embodiment of the present invention. In the figure, the dark triangle line is the actual value of the load, and the light five-pointed star line is the load prediction value (the final power decomposition sequence of the target load) given by the method of the present invention. It can be seen that the method of the present invention is accurate and reliable and can perform load decomposition more accurately.

[0099] Embodiment 2:

[0100] This embodiment provides a non-intrusive load identification device, the device comprising:

[0101] Input module: used to obtain the data to be identified;

[0102] Preprocessing module: used to preprocess the data to be identified, construct the original data set and divide it into several data subsets;

[0103] Decomposition module: used for inputting the data subset into the Stacking ensemble learning model to obtain the final power decomposition sequence of the target load;

[0104] The method for constructing the Stacking ensemble learning model includes:

[0105] Construct a long short-term memory network (LSTM) load decomposition model as the first base learner of the first layer of the Stacking ensemble learning model;

[0106] Construct a denoising autoencoder network (DAE) as the second base learner in the first layer of the Stacking ensemble learning model;

[0107] Construct a fully connected layer as the meta-learner of the second layer of the Stacking ensemble learning model.

[0108] The device of this embodiment can be used to implement the method described in the first embodiment.

[0109] Embodiment three:

[0110] This embodiment provides a non-intrusive load identification device, including a processor and a storage medium;

[0111] The storage medium is used to store instructions;

[0112] The processor is used to operate according to the instructions to execute the steps of the method according to the first embodiment.

[0113] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0114] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0115] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0116] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0117] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A non-intrusive load identification method, It is characterized in that The following steps are involved: Acquire data to be identified; Preprocessing the data to be identified, constructing an original data set and dividing it into several data subsets; Inputting the data subset into a Stacking ensemble learning model to obtain a final power decomposition sequence of the target load; The method for constructing the Stacking ensemble learning model includes: Construct a long short-term memory network load decomposition model as the first base learner of the first layer of the Stacking ensemble learning model; Build a denoising autoencoder network as the second base learner in the first layer of the Stacking ensemble learning model; Construct a fully connected layer as the meta-learner of the second layer of the Stacking ensemble learning model; Obtain the data to be identified, including: Collect high-frequency loads and low-frequency loads; The input of the denoising autoencoder network is a data subset, and the output is a DAE target load decomposition sequence; The denoising autoencoder network is a symmetrical structure divided into an encoder and a decoder. When the encoding layer works, the original data is mapped to the hidden feature space after the operation of the convolution layer, and then mapped to the label space of the target load through the fully connected layer to generate a denoised hidden layer. The decoder then reverses this process to obtain the power curve of the target load. The one-dimensional convolution layer essentially creates multiple filters to extract data features. Where: is the jth feature map of the l-1th layer, is the convolution kernel function of the i-th index of the j-th feature map of the l-th layer, The jth bias parameter of the lth layer, f(x) is the nonlinear activation function of the convolutional neural network; we can get: Where: is the output value of the convolution operation, for The activation value of .

2. The non-intrusive load identification method according to claim 1, It is characterized in that Preprocessing the data to be identified includes: Median filtering and data downsampling, wherein the data downsampling is used to reduce the sampling frequency, and the median filtering is used to eliminate interference in the sampled data.

3. The non-intrusive load identification method according to claim 1, It is characterized in that Divide the data into subsets, including: For the data set S = {(y n ,x n ),n=1,...,N}, where x n is the feature vector of the nth sample, y n is the predicted value corresponding to the nth sample, p is the number of features contained, that is, each feature vector is (x 1 ,x 2 ,...,x p ); Randomly divide the data into K subsets S of equal size 1 ,S 2 ,...,S K ; Where S -k =SS k , respectively define S k and S -k is the k-fold test set and training set in K-fold cross validation; For the first-layer prediction algorithm, there are K base learners, and for the training set S -k Use the kth algorithm to train the base model L k , k=1,...,K.

4. The non-intrusive load identification method according to claim 1, It is characterized in that The input of the long short-term memory network load decomposition model is a data subset, and the output is an LSTM target load decomposition sequence.

5. The non-intrusive load identification method according to claim 1, It is characterized in that The data subset is input into the Stacking ensemble learning model to obtain the final power decomposition sequence of the target load, including: The original data set is divided into several sub-data sets, which are input into each base learner of the first-layer prediction model. Each base learner outputs its own prediction result to obtain the output of the first layer. The output of the first layer is used as the input of the second layer, and the preliminary decomposition results of the two base learners are aggregated to obtain the final decomposition result of the target load, that is, the LSTM target load power decomposition sequence and the target load power decomposition sequence are aggregated to obtain the final power decomposition sequence of the target load.

6. The non-intrusive load identification method according to claim 5, It is characterized in that The training method of the Stacking ensemble learning model includes: For the data set S = {(y n ,x n ),n=1,...,N}, where x n is the feature vector of the nth sample, y n is the predicted value corresponding to the nth sample, p is the number of features contained, that is, each feature vector is (x 1 ,x 2 ,...,x p ); Randomly divide the data into K subsets S of equal size 1 ,S 2 ,...,S K ; where S -k =SS k , respectively define S k and S -k is the k-fold test set and training set in K-fold cross validation; For the first-layer prediction algorithm, there are K base learners, and for the training set S -k Use the kth algorithm to train the base model L k , k=1,...,K; For the k-th fold test set S in K-fold cross validation k For each sample x in n , base learner L k Its prediction is denoted as z kn ; After completing the cross-validation process, the output data of the K base learners constitute a new data sample, namely: S new ={(y n ,z 1n ,...,z kn )}, n = 1, ..., N; the newly generated data set is the second layer input data of Stacking; the meta-learner L obtained by summarizing these data new .

7. A non-intrusive load identification device for performing the method of claim 1, It is characterized in that The device comprises: Input module: used to obtain data to be identified; Preprocessing module: used to preprocess the data to be identified, construct the original data set and divide it into several data subsets; Decomposition module: used for inputting the data subset into the Stacking ensemble learning model to obtain the final power decomposition sequence of the target load; The method for constructing the Stacking ensemble learning model includes: Construct a long short-term memory network load decomposition model as the first base learner of the first layer of the Stacking ensemble learning model; Build a denoising autoencoder network as the second base learner in the first layer of the Stacking ensemble learning model; Construct a fully connected layer as the meta-learner of the second layer of the Stacking ensemble learning model.

8. A non-intrusive load identification device, It is characterized in that including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-6.