Non-intrusive load monitoring method and system based on knowledge distillation algorithm

Abstract

The invention relates to a non-intrusive load monitoring method and system based on a knowledge distillation algorithm. The non-intrusive load monitoring method comprises the following steps: step 1, constructing a simulation data set; 2, normalizing data in the simulation data set to obtain input data for model training, verification and testing; step 3, obtaining a teacher load decomposition model used for decomposing the consumed power of the load equipment from the total power signal; 4, the obtained network model is a student load decomposition model used for decomposing the consumed power of the load equipment from the total power signal; step 5, obtaining a required final student load decomposition model; and step 6, inputting test total power signal data in the simulation data set into the final student load decomposition model trained by the knowledge distillation algorithm in the step 5 to obtain target equipment power data. According to the invention, the problem of deployment and application of the NILM technology on the user side edge equipment can be solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-invasive power load monitoring, and relates to a non-invasive user load monitoring method and system, in particular to a non-invasive load monitoring method and system based on a knowledge distillation algorithm. Background Art

[0002] With the continuous development of smart grids and the continuous promotion of smart sensor technology, power big data analysis has gradually become a hot topic for experts and scholars in the power industry. Non-intrusive load monitoring (NILM) technology refers to installing only one sensor at the user's entrance, monitoring the power consumption and working status of each or each type of electrical appliance in the house by collecting and analyzing the user's total entrance data, so as to understand the power consumption and power usage patterns of each electrical appliance in the residents' homes. NILM technology is of great significance for optimizing grid planning, operation and management, and saving electricity and electricity bills for users. It is also one of the key technologies to promote refined management on the demand side and energy conservation and emission reduction.

[0003] NILM technology has received widespread attention since it was proposed. Currently, many methods have been applied to non-invasive load decomposition. Among them, the popular deep learning theory in recent years has become an important tool for realizing NILM technology. Among them, convolutional neural networks can effectively mine the internal characteristics of various source and load devices, and the Seq2point model proposed based on convolutional neural networks converts the NILM problem into a sequence-to-point model training process. The midpoint power value of a single source and load device window is obtained based on the continuously sliding total power window sequence, showing good results in the field of load decomposition. Since the model has more than millions of parameters that need to be fine-tuned and requires a large amount of labeled data for training, it can only be trained and verified in the cloud.

[0004] Therefore, existing non-intrusive load monitoring methods require transmitting massive amounts of data from the edge to the cloud for training. However, large amounts of data transmission between the cloud and the edge increase communication costs and also pose certain risks to user privacy and security.

[0005] Therefore, in response to the above problems, the present invention proposes a non-intrusive load monitoring method and system based on a knowledge distillation algorithm, which establishes a suitable lightweight neural network model so that training, verification and testing can be completed only on edge devices.

[0006] After searching, no public documents of the prior art that are identical or similar to the present invention were found. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a non-intrusive load monitoring method and system based on data generation and knowledge distillation algorithm, which can solve the problem of deploying and applying NILM technology on user-side edge devices.

[0008] The present invention solves the practical problem by adopting the following technical solutions:

[0009] A non-intrusive load monitoring method based on a knowledge distillation algorithm comprises the following steps:

[0010] Step 1: construct a simulation data set based on background load and target electrical appliance working curve;

[0011] Step 2: Divide the simulation data set based on the background load and the target electrical appliance working curve constructed in step 1 into a training set, a validation set, and a test set in a ratio of 3:1:1, and normalize the data in the simulation data set to obtain input data for model training, validation, and testing;

[0012] Step 3: construct a teacher neural network structure based on the denoising LSTM autoencoder, and perform supervised training and verification on the teacher neural network structure through the data set normalized in step 2 to obtain a teacher load decomposition model for decomposing the power consumption of the load equipment from the total power signal;

[0013] Step 4: construct a student neural network structure, and conduct supervised training and verification on the student neural network structure through the data set processed in step 2. The obtained network model is the student load decomposition model used to decompose the power consumed by the load equipment from the total power signal;

[0014] Step 5: Set the hyperparameters of the knowledge distillation algorithm, calculate the difference loss between the teacher load decomposition model and the student load decomposition model based on the feature output of the teacher load decomposition model in steps 3 and 4 and the feature output of the student load decomposition model, substitute the difference loss into the student load decomposition model in step 4 and retrain to obtain the desired final student load decomposition model;

[0015] Step 6: Input the test total power signal data in the simulation data set into the final student load decomposition model trained by the knowledge distillation algorithm in step 5. The output is the target device power data decomposed from the total power signal sequence.

[0016] Moreover, the specific method of step 1 is:

[0017] According to the load data of sub-item electrical appliances in the public data set, the typical working curves of sub-item electrical appliances are constructed and randomly superimposed on the original total power data. A simulation data set based on the background load and the working curve of the target electrical appliances is constructed. The simulation data set is processed using the median filtering method to eliminate the noise in the total power data.

[0018] Moreover, the simulation data set of step 2 includes the denoised labeled total power data, and the training, verification and test data in the data set are obtained, and the sequence data used are normalized.

[0019] Moreover, the calculation formula for normalizing the sequence data used in step 2 is as follows:

[0020]

[0021] In formula (1), x t is the actual measured value of the power sequence, is the average value of the power sequence, σ is the standard deviation of the power sequence, and the power sequence includes the total power sequence and the load power sequence.

[0022] Moreover, the specific method of constructing the teacher neural network structure based on the denoising LSTM autoencoder in step 3 is:

[0023] First, an encoder structure is constructed to mine the implicit features of the target device in the total power data, and then the power consumption value of the target appliance is restored through the constructed decoder structure. The two are combined to obtain the required teacher neural network structure based on the denoising LSTM autoencoder.

[0024] Moreover, the teacher neural network structure of step 3 is composed of an encoder structure and a decoder structure. First, some convolution features are extracted based on one or two layers of standard convolution network structures and two or one layers of downsampling structures, and then the LSTM network structure and the one-layer downsampling structure are input to obtain the timing features. Finally, the downsampled timing features are input to the one-layer standard convolution network structure and the one-layer downsampling structure to obtain the encoder network structure, and then the implicit features output by the encoder are upsampled by two layers and two or one layers of standard convolution network structures to obtain the convolution features, and then the LSTM network structure and the one-layer upsampling structure are input to obtain the timing features, and finally the upsampled timing features are input to the upsampling structure and the one-layer standard convolution network structure to obtain the decoder network structure, and finally the encoder and decoder structures are combined to obtain the required teacher neural network structure based on the denoising LSTM autoencoder.

[0025] Moreover, the step 3 performs supervised training and verification on the teacher neural network structure through the data set normalized in step 2, and the specific method of obtaining the teacher load decomposition model for decomposing the power consumption of the load equipment from the total power signal is:

[0026] During the supervised training and verification process, the network parameters are updated in a supervised manner; the input of the lightweight sequence-to-point convolutional neural network is the total power signal data window sequence of the target generalized load node, and the output of the network is the power value corresponding to the midpoint of the window sequence; the total power signal data window sequence and the load device power value of the window sequence are input into the teacher neural network structure, and as the input window of the network model continues to slide, a teacher load decomposition model for decomposing the power consumption of the load device from the total power signal is obtained.

[0027] Moreover, the student neural network structure of step 4 is a lightweight convolutional neural network structure, which consists of two convolutional layer structures, two fully connected layer structures and one gated recurrent unit structure;

[0028] The specific method of the step 4 of performing supervised training and verification on the student neural network structure through the data set normalized in step 2 to obtain the student load decomposition model for decomposing the power consumed by the load equipment from the total power signal is:

[0029] During the supervised training and verification process, the network parameters are updated in a supervised manner; the input of the lightweight sequence-to-point convolutional neural network is the total power signal data window sequence of the target generalized load node, and the output of the network is the power value corresponding to the midpoint time of the window sequence; the total power signal data window sequence and the load device power value of the window sequence are input into the student neural network structure, and as the input window of the network model continues to slide, a student load decomposition model for decomposing the power consumption of the load device from the total power signal is obtained.

[0030] Moreover, the difference loss between the teacher load decomposition model and the student load decomposition model in step 5 is calculated by the feature output of the teacher load decomposition model and the feature output of the student load decomposition model. During the process of re-supervised training of the student neural network model, the loss function is readjusted to:

[0031]

[0032] In formula (4), the part of the loss function that is added is the difference divergence between the feature output of the teacher model and the feature output of the student model, f T is the characteristic output of the teacher network model, f S is the feature output of the student network model, and n is the number of features output by the model.

[0033] A non-intrusive load monitoring system based on a knowledge distillation algorithm, comprising a neural network model training and verification module, and a neural network model testing module;

[0034] The neural network model training and verification module first trains and verifies the teacher load decomposition model and the student load decomposition model through the total power data and the target load device label power data, and then trains and verifies the final student load decomposition model;

[0035] The neural network model testing module is only used to test the decomposition effect of the final student load decomposition model, and does not test the teacher load decomposition model. Specifically, the power series data of the load equipment is obtained by substituting the sliding total power window in the test set into the established final student load decomposition model.

[0036] Advantages and beneficial effects of the present invention:

[0037] 1. The present invention proposes a non-intrusive load monitoring method and system based on a knowledge distillation algorithm, constructs a data set containing more labeled data based on a data generation algorithm, and combines the knowledge distillation algorithm with the NILM problem, aiming to solve the problem of deploying and applying NILM technology on user-side edge devices. The method of the present invention mainly includes: randomly superimposing the sub-item electrical load data of the public data set with the total power data to construct a simulation data set, using the median filtering method to process the simulation data set to eliminate the noise in the total power data, and then constructing a student neural network structure model, introducing the feature output knowledge of the teacher neural network model into the training process of the student neural network model after distillation, and obtaining a lightweight student sequence-to-point convolutional neural network model through supervised training and verification of the simulation data set, and decomposing the power consumption sequence of the target load equipment obtained from the total power signal sequence. Aiming at the problem of complex parameters and calculation amount of the existing NILM sequence-to-point convolutional neural network model, the present invention designs a lightweight sequence-to-point load decomposition method based on data generation and knowledge distillation algorithm, which greatly reduces the parameters and calculation amount of model training, and provides effective support for the application of lightweight models in edge devices.

[0038] 2. The present invention takes into account that the knowledge distillation algorithm trains the student neural network model through the teacher neural network model, which greatly reduces the parameters and computational complexity of the model. Therefore, the present invention combines the knowledge distillation algorithm with the sequence-to-point neural network to propose a non-intrusive load monitoring method and system based on the knowledge distillation algorithm, which can solve the problem of deploying and applying NILM technology on user-side edge devices, and can quickly realize the monitoring and identification of electrical equipment at the edge. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of a non-intrusive load monitoring method based on a knowledge distillation algorithm of the present invention;

[0040] Figure 2 It is a neural network structure diagram of the teacher denoising autoencoder designed by the present invention;

[0041] Figure 3 This is a structural diagram of the lightweight student sequence-to-point convolutional neural network designed by the present invention. DETAILED DESCRIPTION

[0042] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0043] The design idea of ​​a non-invasive load monitoring method based on a knowledge distillation algorithm proposed in the present invention is: the sub-item electrical load data of a public data set is randomly superimposed on the total power data to construct a simulation data set, and then a student neural network structure model is constructed. The feature output knowledge of the teacher neural network model is distilled and introduced into the training process of the student neural network model, which can greatly reduce the number of model parameters and the computational complexity. A lightweight student sequence-to-point convolutional neural network model is obtained through supervised training and verification of the simulation data set. The input of the total power signal sequence can realize the power decomposition of the target load equipment.

[0044] A non-intrusive load monitoring method based on knowledge distillation algorithm, such as Figure 1 As shown, the following steps are included:

[0045] Step 1: construct a simulation data set based on background load and target electrical appliance working curve;

[0046] The specific method of step 1 is:

[0047] According to the load data of sub-item electrical appliances in the public data set, the typical working curves of sub-item electrical appliances are constructed and randomly superimposed on the original total power data. A simulation data set based on the background load and the working curve of the target electrical appliances is constructed. The simulation data set is processed using the median filtering method to eliminate the noise in the total power data.

[0048] Step 2: Divide the simulation data set based on the background load and the target electrical appliance working curve constructed in step 1 into a training set, a validation set, and a test set in a ratio of 3:1:1, and normalize the data in the simulation data set to obtain input data for model training, validation, and testing;

[0049] The simulation data set of step 2 includes the denoised labeled total power data, obtains the training, verification and test data in the data set, and normalizes the sequence data used;

[0050] The calculation formula for normalizing the sequence data used in step 2 is as follows:

[0051]

[0052] In formula (1), x tis the actual measured value of the power sequence, is the average value of the power sequence, σ is the standard deviation of the power sequence, and the power sequence includes the total power sequence and the load power sequence.

[0053] In this embodiment, step 2 constructs the training, verification and test data sets required by this method based on the simulation data set. For example, five generalized load devices are selected: electric kettle, microwave oven, refrigerator, dishwasher and washing machine as the objects of load data collection. The public data set of Building No. 1 in a certain community is determined as the total data set, and the load data set of a certain resident user in Building No. 1 is divided into a training set, a verification set and a test set in a ratio of 3:1:1. The total power data of the user and the power data of the load device are normalized to obtain the input data for model training, verification and testing.

[0054] The residential user load data set includes labeled total power data, sliding to obtain sequence windows of training, verification and test data, and normalizing the sequence window power data used according to the following formula, that is, the power minus its mean value and then divided by its standard deviation.

[0055]

[0056] In formula (1), x t is the real measured value of the power data in the sequence window. The power sequence includes the total power sequence and the load power sequence. The sequence value is obtained by measuring the equipment. is the mean value of the power data, and σ is the standard deviation of the power data.

[0057] Step 3: construct a teacher neural network structure based on the denoising LSTM autoencoder, and perform supervised training and verification on the teacher neural network structure through the data set normalized in step 2 to obtain a teacher load decomposition model for decomposing the power consumption of the load equipment from the total power signal;

[0058] The specific method of constructing the teacher neural network structure based on the denoising LSTM autoencoder in step 3 is:

[0059] First, an encoder structure is constructed to mine the implicit features of the target device in the total power data, and then the power consumption value of the target appliance is restored through the constructed decoder structure. The two are combined to obtain the required teacher neural network structure based on the denoising LSTM autoencoder.

[0060] The teacher neural network structure of step 3 is composed of an encoder structure and a decoder structure. First, some convolution features are extracted based on one or two layers of standard convolution network structures and two or one layers of downsampling structures. Then, the LSTM network structure and the one layer of downsampling structure are input to obtain the timing features. Finally, the downsampled timing features are input to the one layer of standard convolution network structure and the one layer of downsampling structure to obtain the encoder network structure. Then, the implicit features output by the encoder are upsampled by two layers and the two layers of standard convolution network structures are input to obtain the convolution features. Then, the upsampled timing features are input to the upsampling structure and the one layer of standard convolution network structure to obtain the decoder network structure. Finally, the encoder and decoder structures are combined to obtain the required teacher neural network structure based on the denoising LSTM autoencoder.

[0061] The specific method of step 3 of performing supervised training and verification on the teacher neural network structure through the data set normalized in step 2 to obtain the teacher load decomposition model for decomposing the power consumption of the load equipment from the total power signal is:

[0062] During the supervised training and verification process, the network parameters are updated in a supervised manner; the input of the lightweight sequence-to-point convolutional neural network is the total power signal data window sequence of the target generalized load node, and the output of the network is the power value corresponding to the midpoint of the window sequence; the total power signal data window sequence and the load device power value of the window sequence are input into the teacher neural network structure, and as the input window of the network model continues to slide, a teacher load decomposition model for decomposing the power consumption of the load device from the total power signal is obtained.

[0063] In this embodiment, the teacher neural network structure is constructed based on the denoising LSTM autoencoder in step 3. Figure 2 As shown, the network input is represented by the total power sequence sliding window Y t:t+W-1 , t is the time, W represents the length of the total power sequence sliding window, W is an odd number, and the output of the teacher neural network is the midpoint value of the power sequence window of the mth load device contained The teacher neural network model in the supervised training and verification process is shown in formula (2):

[0064]

[0065] In formula (2), ε is W-dimensional Gaussian random noise, and F is the overall function of the neural network parameters;

[0066] The loss function in the supervised training process is specifically expressed as:

[0067]

[0068] In formula (3), x t is the power measurement value of the target appliance, x' t is the power estimate of the target appliance, T is the total length of the total power sequence input, and the total power sequence value Y t:t+W-1 When input into the teacher neural network, zero values ​​with a length of [(W-1) / 2] are added to both ends of the total power sequence.

[0069] Step 4: construct a student neural network structure, and conduct supervised training and verification on the student neural network structure through the data set processed in step 2. The obtained network model is the student load decomposition model used to decompose the power consumed by the load equipment from the total power signal;

[0070] The student neural network structure of step 4 is a lightweight convolutional neural network structure, which consists of two convolutional layer structures, two fully connected layer structures and one gated recurrent unit structure.

[0071] The specific method of the step 4 of performing supervised training and verification on the student neural network structure through the data set normalized in step 2 to obtain the student load decomposition model for decomposing the power consumed by the load equipment from the total power signal is:

[0072] During the supervised training and verification process, the network parameters are updated in a supervised manner; the input of the lightweight sequence-to-point convolutional neural network is the total power signal data window sequence of the target generalized load node, and the output of the network is the power value corresponding to the midpoint time of the window sequence; the total power signal data window sequence and the load device power value of the window sequence are input into the student neural network structure, and as the input window of the network model continues to slide, a student load decomposition model for decomposing the power consumption of the load device from the total power signal is obtained.

[0073] In this embodiment, the student neural network structure of step 4 is as follows: Figure 3 As shown in the figure, the training and verification process is consistent with step 3, so it will not be repeated here.

[0074] Step 5: Set the hyperparameters of the knowledge distillation algorithm, calculate the difference loss between the teacher load decomposition model and the student load decomposition model based on the feature output of the teacher load decomposition model in steps 3 and 4 and the feature output of the student load decomposition model, substitute the difference loss into the student load decomposition model in step 4 and retrain to obtain the desired final student load decomposition model;

[0075] The difference loss between the teacher load decomposition model and the student load decomposition model in step 5 is calculated by the feature output of the teacher load decomposition model and the feature output of the student load decomposition model. During the process of re-supervised training of the student neural network model, the loss function is readjusted to:

[0076]

[0077] In formula (4), the part of the loss function that is added is the difference divergence between the feature output of the teacher model and the feature output of the student model, f T is the characteristic output of the teacher network model, f S is the feature output of the student network model, and n is the number of features output by the model.

[0078] In this embodiment, the hyperparameters of the knowledge distillation algorithm are set, and the learning rates of the teacher load decomposition model and the student load decomposition model are 0.001 and 0.0005, respectively.

[0079] Step 6: Input the test total power signal data in the simulation data set into the final student load decomposition model trained by the knowledge distillation algorithm in step 5. The output is the target device power data decomposed from the total power signal sequence.

[0080] A non-intrusive load monitoring system based on a knowledge distillation algorithm, comprising a neural network model training and verification module, and a neural network model testing module;

[0081] The neural network model training and verification module first trains and verifies the teacher load decomposition model and the student load decomposition model through the total power data and the target load device label power data, and then trains and verifies the final student load decomposition model;

[0082] The neural network model testing module is only used to test the decomposition effect of the final student load decomposition model, and does not test the teacher load decomposition model. Specifically, the power series data of the load equipment is obtained by substituting the sliding total power window in the test set into the established final student load decomposition model.

[0083] The above two functional modules are used to implement a non-intrusive load monitoring method and system based on knowledge distillation algorithm.

[0084] The present invention is further described below by means of specific examples:

[0085] Taking the actual application scenario as an example, the specific steps of realizing non-invasive lightweight load decomposition using the method of the present invention are as follows:

[0086] 1) Based on the load data of sub-items of electrical appliances in the public data set, a typical working curve of each sub-item of electrical appliances is constructed, which is randomly superimposed on the original total power data to construct a user simulation data set based on the background load and the working curve of the target electrical appliances;

[0087] Taking the UK-DALE public dataset in the UK as an example, the data of building No. 1 in the UK-DALE dataset was selected as the research object, and the typical working curve of the target electrical appliance was extracted. It was randomly superimposed on the original total power data to obtain a simulation dataset based on the background load and the working curve of the target electrical appliance.

[0088] 2) Obtain training, verification and test data from the user simulation data set and normalize the data;

[0089] The simulation data set obtained in the previous step is segmented, and the ratio of the training, validation and test data sets is set to 3:1:1. Five generalized load devices are selected, namely: electric kettle, microwave oven, refrigerator, dishwasher and washing machine, and load decomposition is performed under the teacher and student neural network structure designed by the present invention.

[0090] 3) First, the network parameters of the teacher denoising LSTM autoencoder are initialized, the normalized power data is input into the network structure for training and the network model is obtained, and the model feature output is calculated.

[0091] 4) Then the student lightweight sequence-to-point neural network parameters are initialized, the normalized power data is input into the network structure for training and the network model is obtained, and the model feature output is calculated. The teacher neural network model is subjected to knowledge distillation and retraining to obtain the required student neural network model; the student lightweight sequence-to-point neural network model occupies 0.11 megabytes of memory and 0.266 mega floating-point operations, which are 125 times and 24 times lower than the classic sequence-to-point Seq2point model that occupies 13.84 megabytes of memory and 6.42 mega floating-point operations, respectively.

[0092] 5) Finally, the trained model is tested and the test set total power sequence of building No. 1 is input into the trained model to decompose the target load of building No. 1.

[0093] In the example of this research material, the evaluation indicators for the load decomposition results include mean absolute error (MAE) and signal aggregate error (SAE), and their expressions are as follows:

[0094]

[0095] Here, MAE is used to calculate the power prediction value (that is, the power decomposition value obtained by the method of the present invention) and the actual measured value x t The average absolute error between , where T represents the total duration of the test period, and SAE is used to calculate the power prediction value during the test period. The error ratio between the actual value r.

[0096] The MAE and SAE results of load decomposition using the method of the present invention with a small amount of labeled data are shown in Table 1.

[0097] Table 1 MAE and SAE indicators of decomposition results

[0098]

[0099] In order to verify the improvement effect of the median filtering method on the present method, the present invention verifies its effectiveness through ablation experiments. The specific decomposition results are shown in Table 2.

[0100] Table 2 Decomposition results of ablation experiments

[0101]

[0102] As shown in Table 2, if the experimental data is not processed by median filtering, the decomposition accuracy will be reduced. Therefore, the median filtering method plays an important role in improving the decomposition accuracy.

[0103] It can be seen from the above research materials that the method of the present invention can greatly reduce the number of parameters and calculations of the neural network model, and has a good decomposition effect while realizing NILM quickly at the edge.

[0104] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation manner. Any other implementation manners derived by those skilled in the art based on the technical solution of the present invention also fall within the scope of protection of the present invention.

Claims

1. A non-intrusive load monitoring method based on knowledge distillation algorithm, characterized in that: The following steps are involved: Step 1: construct a simulation data set based on background load and target electrical appliance working curve; Step 2: Divide the simulation data set based on the background load and the target electrical appliance working curve constructed in step 1 into a training set, a validation set, and a test set in a ratio of 3:1:1, and normalize the data in the simulation data set to obtain input data for model training, validation, and testing; Step 3: construct a teacher neural network structure based on the denoising LSTM autoencoder, and perform supervised training and verification on the teacher neural network structure through the data set normalized in step 2 to obtain a teacher load decomposition model for decomposing the power consumption of the load equipment from the total power signal; Step 4: construct a student neural network structure, and conduct supervised training and verification on the student neural network structure through the data set processed in step 2. The obtained network model is the student load decomposition model used to decompose the power consumed by the load equipment from the total power signal; Step 5: Set the hyperparameters of the knowledge distillation algorithm, calculate the difference loss between the teacher load decomposition model and the student load decomposition model based on the feature output of the teacher load decomposition model in steps 3 and 4 and the feature output of the student load decomposition model, substitute the difference loss into the student load decomposition model in step 4 and retrain to obtain the desired final student load decomposition model; Step 6: Input the test total power signal data in the simulation data set into the final student load decomposition model trained by the knowledge distillation algorithm in step 5. The output is the target device power data decomposed from the total power signal sequence.

2. According to claim 1, a non-intrusive load monitoring method based on knowledge distillation algorithm is characterized in that: The specific method of step 1 is: According to the load data of sub-item electrical appliances in the public data set, the typical working curves of sub-item electrical appliances are constructed and randomly superimposed on the original total power data. A simulation data set based on the background load and the working curve of the target electrical appliances is constructed. The simulation data set is processed using the median filtering method to eliminate the noise in the total power data.

3. The non-intrusive load monitoring method based on knowledge distillation algorithm according to claim 1 is characterized in that: The simulation data set of step 2 includes the denoised labeled total power data, and the training, verification and test data in the data set are obtained, and the sequence data used are normalized.

4. The non-intrusive load monitoring method based on knowledge distillation algorithm according to claim 3 is characterized in that: The calculation formula for normalizing the sequence data used in step 2 is as follows: In formula (1), x t is the actual measured value of the power sequence, is the average value of the power sequence, σ is the standard deviation of the power sequence, and the power sequence includes the total power sequence and the load power sequence.

5. The non-intrusive load monitoring method based on knowledge distillation algorithm according to claim 1 is characterized in that: The specific method of constructing the teacher neural network structure based on the denoising LSTM autoencoder in step 3 is: First, an encoder structure is constructed to mine the implicit features of the target device in the total power data, and then the power consumption value of the target appliance is restored through the constructed decoder structure. The two are combined to obtain the required teacher neural network structure based on the denoising LSTM autoencoder.

6. The non-intrusive load monitoring method based on knowledge distillation algorithm according to claim 1 is characterized in that: The teacher neural network structure of step 3 is composed of an encoder structure and a decoder structure. First, some convolution features are extracted based on one or two layers of standard convolution network structures and two or one layers of downsampling structures. Then, the LSTM network structure and the one layer of downsampling structure are input to obtain the timing features. Finally, the downsampled timing features are input to the one layer of standard convolution network structure and the one layer of downsampling structure to obtain the encoder network structure. Then, the implicit features output by the encoder are upsampled by two layers and the two layers of standard convolution network structures are input to obtain the convolution features. Then, the upsampled timing features are input to the upsampling structure and the one layer of standard convolution network structure to obtain the decoder network structure. Finally, the encoder and decoder structures are combined to obtain the required teacher neural network structure based on the denoising LSTM autoencoder.

7. The non-intrusive load monitoring method based on knowledge distillation algorithm according to claim 1 is characterized in that: The specific method of step 3 of performing supervised training and verification on the teacher neural network structure through the data set normalized in step 2 to obtain the teacher load decomposition model for decomposing the power consumption of the load equipment from the total power signal is: During the supervised training and verification process, the network parameters are updated in a supervised manner; the input of the lightweight sequence-to-point convolutional neural network is the total power signal data window sequence of the target generalized load node, and the output of the network is the power value corresponding to the midpoint of the window sequence; the total power signal data window sequence and the load device power value of the window sequence are input into the teacher neural network structure, and as the input window of the network model continues to slide, a teacher load decomposition model for decomposing the power consumption of the load device from the total power signal is obtained.

8. The non-intrusive load monitoring method based on knowledge distillation algorithm according to claim 1 is characterized in that: The student neural network structure of step 4 is a lightweight convolutional neural network structure, which consists of two convolutional layer structures, two fully connected layer structures and one gated recurrent unit structure; The specific method of the step 4 of performing supervised training and verification on the student neural network structure through the data set normalized in step 2 to obtain the student load decomposition model for decomposing the power consumed by the load equipment from the total power signal is: During the supervised training and verification process, the network parameters are updated in a supervised manner; the input of the lightweight sequence-to-point convolutional neural network is the total power signal data window sequence of the target generalized load node, and the output of the network is the power value corresponding to the midpoint time of the window sequence; the total power signal data window sequence and the load device power value of the window sequence are input into the student neural network structure, and as the input window of the network model continues to slide, a student load decomposition model for decomposing the power consumption of the load device from the total power signal is obtained.

9. The non-intrusive load monitoring method based on knowledge distillation algorithm according to claim 1 is characterized in that: The difference loss between the teacher load decomposition model and the student load decomposition model in step 5 is calculated by the feature output of the teacher load decomposition model and the feature output of the student load decomposition model. During the process of re-supervised training of the student neural network model, the loss function is readjusted to: In formula (4), the part of the loss function that is added is the difference divergence between the feature output of the teacher model and the feature output of the student model, f T is the characteristic output of the teacher network model, f S is the feature output of the student network model, and n is the number of features output by the model.

10. A non-intrusive load monitoring system based on knowledge distillation algorithm, characterized in that: Including neural network model training and verification module, and neural network model testing module; The neural network model training and verification module first trains and verifies the teacher load decomposition model and the student load decomposition model through the total power data and the target load device label power data, and then trains and verifies the final student load decomposition model; The neural network model testing module is only used to test the decomposition effect of the final student load decomposition model, and does not test the teacher load decomposition model. Specifically, the power series data of the load equipment is obtained by substituting the sliding total power window in the test set into the established final student load decomposition model.

Images