A non-intrusive load monitoring method and system based on SResNet-SCNN-BiLSTM network

Abstract

The present invention discloses a non-intrusive load monitoring method and system based on SResNet-SCNN-BiLSTM network, which relates to the field of power load monitoring, including: selecting the load data of each electrical appliance from a household power consumption data set to obtain a total power sequence, and pre-processing the total power sequence to obtain an initial sample data set; using a CGAN network to expand the initial sample data set to obtain an expanded sample data set; combining an improved residual network and an SCNN-BiLSTM network to construct a load decomposition network model, and using the expanded sample data set for training; using the trained load decomposition network model to identify and decompose the total household power load to be identified, and obtain the loads of different electrical appliances. The present invention can effectively improve the accuracy of non-intrusive load monitoring in identifying power loads of different electrical equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of power load monitoring, and more specifically to a non-intrusive load monitoring method and system based on an SResNet-SCNN-BiLSTM network. Background Art

[0002] With the continuous deterioration of the global climate and the increasing shortage of renewable energy, energy conservation and emission reduction have gradually become an important issue that cannot be delayed. Residential and commercial buildings account for about 40% of global energy consumption, and carbon dioxide emissions generated by building electricity account for more than 1 / 3 of the world. Therefore, it is very important to monitor the power consumption of each appliance in the home to improve energy efficiency, and load monitoring has become an important topic in academia and industry.

[0003] Traditional load monitoring adopts an intrusive method, that is, installing sensors on the user's electrical appliances and then recording the power consumption of each appliance. Although this method ensures that the monitored data is reliable and has low noise, it is difficult to implement. Another method is non-intrusive load monitoring (NILM), which was first proposed by Hart in 1992. NILM installs sensors at the electricity meter in the home, records the total meter reading, and estimates the power consumption of each appliance through a series of algorithms, thereby obtaining the overall pattern of electricity consumption. However, the use of traditional CNN neural network models to perform NILM tasks often fails to take into account the higher-level semantics contained in energy consumption behavior; and the existing deep learning lacks sufficient labeled data, making the model prone to overfitting and poor generalization ability. The commonly used method of generating data through interpolation technology and enhancing the data set may lack the diversity of the original data and cannot capture the global dependency and long-term trends in the data, resulting in the inability of the existing NILM method to accurately identify and decompose the power load of different electrical equipment.

[0004] Therefore, how to improve the accuracy of non-intrusive load monitoring in identifying power loads of different electrical equipment is an urgent problem that technicians in this field need to solve. Summary of the invention

[0005] In view of this, the present invention provides a non-intrusive load monitoring method and system based on the SResNet-SCNN-BiLSTM network, uses CGAN for data enhancement, and uses a model combining SCNN-BiLSTM and residual network to identify and decompose the power load of different electrical equipment.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention discloses a non-intrusive load monitoring method based on SResNet-SCNN-BiLSTM network, and the specific steps are as follows:

[0008] Selecting the load data of each electrical appliance from the household electricity consumption data set to obtain a total power sequence, and preprocessing the total power sequence to obtain an initial sample data set;

[0009] Expanding the initial sample data set using a CGAN network to obtain an expanded sample data set;

[0010] Combining the improved residual network and the SCNN-BiLSTM network to build a load decomposition network model, using the improved residual network to extract features and output a high-dimensional feature map, reshaping the high-dimensional feature map, and using SCNN to extract spatial features from the reshaped high-dimensional feature map, and adjusting it to a feature matrix suitable for BiLSTM; using the expanded sample data set to train the load decomposition network model;

[0011] The trained load decomposition network model is used to identify and decompose the total household electricity load to obtain the load of different electrical appliances.

[0012] Further, the preprocessing includes:

[0013] Performing data alignment on the load data of each electrical appliance according to the timestamp information of the load data of each electrical appliance;

[0014] Use the backward filling method to fill the missing values ​​in the load data of each electrical appliance after data alignment;

[0015] According to the types of electrical appliances in the total power sequence, a category label is assigned to the sample data.

[0016] Furthermore, the CGAN network includes a generator and a discriminator; the generator generates sample data similar to the real data distribution by adding potential random noise to the total power sequence according to the total power sequence and the category label; the discriminator judges the authenticity of the input sample data according to the input sample data and the corresponding category label, and outputs a binary judgment result; the generator and the discriminator are alternately trained, and after the training is completed, the trained generator is used to generate new sample data to expand the initial sample data set.

[0017] Furthermore, the load decomposition network model uses an improved residual network to extract features from the sample data in the expanded sample data set to obtain a high-dimensional feature map of the sample data, and obtains a first feature map after adjustment by a first reshaping layer; uses an SCNN network to extract spatial features from the first feature map to obtain a second feature map, and obtains a feature matrix after adjustment by a second reshaping layer; uses a BiLSTM network to convert the feature matrix into multiple time series tensors, and then extracts context features of the time series tensors from two directions, and the fully connected layer outputs the load power of each electrical appliance based on the context features.

[0018] Furthermore, the improved residual network includes a starting stage, four main stages, and an ending stage; the first residual module of each stage reduces the size of the feature map by half through a convolution layer with a stride of 2, and increases the number of channels. The size of the feature map represents the spatial resolution of the feature map, and the number of channels of the feature map is the number of feature maps composed of the extracted feature information; each of the main stages uses the convolution layer in the residual module to extract features, a batch normalization layer is connected after the convolution layer, and a ReLU activation function is connected after the batch normalization layer.

[0019] Furthermore, the SCNN network slices the first feature map vertically and horizontally, performs convolution stacking operations on the vertically sliced ​​slices in a top-to-bottom and bottom-to-top order, and performs convolution stacking operations on the horizontally sliced ​​slices in a left-to-right and right-to-left order; the convolution stacking operation is to perform a small convolution on the first slice and superimpose the convolution result on the next slice until the superposition reaches the last slice.

[0020] Furthermore, the BiLSTM network uses a recurrent connection to learn the global dependency between the time series tensors, and the formula is:

[0021]

[0022]

[0023] in, and represents the input gate and output gate, σ represents the sigmoid function, and represents the weight matrix, represents the bias; represents the input vector of the lth layer at the i-th time step, and v represents the input i t The weight vector, v * Represents The weight vector of , t represents the time step, i t represents the input gate activation value at time step t, ht+1 represents the hidden state of the reverse LSTM at the next time step t+1.

[0024] Furthermore, the output of the fully connected layer is y=[y1,y2,…,y n ], where y is an n-dimensional vector, n is the total number of electrical appliances, y i Represents the power of the i-th appliance, and the fully connected layer is decomposed by the following formula:

[0025] y = σ(WF + b);

[0026] Among them, σ is the activation function, b is the bias vector, and W is the weight matrix of the fully connected layer.

[0027] Furthermore, in the training using the expanded sample data set:

[0028] Using MSE as the loss function, the formula is as follows:

[0029]

[0030] Among them, a i represents the eigenvector, g(a i ) represents the output vector, u i is the target label vector, and D represents the number of samples in the dataset;

[0031] The Adam optimizer is used to train the model, and the mean absolute error and signal polymer difference are used as evaluation indicators. The formula is as follows:

[0032]

[0033] Among them, MAE is the mean absolute error, SAE is the signal polymer difference, is the estimated power of appliance j at time t, is the actual value; E are the actual total energy consumption and predicted total energy consumption of electrical equipment,

[0034] The present invention also discloses a non-intrusive load monitoring system based on SResNet-SCNN-BiLSTM network, comprising:

[0035] Data acquisition module: selects the load data of each appliance from the household electricity consumption data set to obtain a total power sequence, and preprocesses the total power sequence to obtain an initial sample data set;

[0036] Data expansion module: using the CGAN network to expand the initial sample data set to obtain an expanded sample data set;

[0037] Model training module: combining the improved residual network and the SCNN-BiLSTM network to build a load decomposition network model, using the improved residual network to extract features and output a high-dimensional feature map, reshaping the high-dimensional feature map, and using SCNN to extract spatial features from the reshaped high-dimensional feature map, and adjusting it to a feature matrix suitable for BiLSTM; using the expanded sample data set to train the load decomposition network model;

[0038] Identification module: Use the trained load decomposition network model to identify and decompose the total household electricity load to obtain the load of different electrical appliances.

[0039] It can be seen from the above technical solutions that compared with the prior art, the present invention discloses a non-intrusive load monitoring method and system based on the SResNet-SCNN-BiLSTM network, which uses a conditional generative adversarial network to solve the problem that the model may not be able to understand the important characteristics of the time series when there is a lack of sufficient amount of time series data; the use of an improved residual network can increase the network depth while providing a better path for information propagation in the network layer, thereby promoting the learning process; the use of an improved SCNN-BiLSTM model can further mine the potential features in the data, thereby improving the decomposition accuracy, so that the present invention can effectively improve the accuracy of non-intrusive load monitoring in identifying power loads of different electrical equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0041] Figure 1 It is a schematic diagram of the overall process of an embodiment of the present invention.

[0042] Figure 2 It is a schematic diagram of the overall structure of the load decomposition network model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] The embodiment of the present invention discloses a non-intrusive load monitoring method based on SResNet-SCNN-BiLSTM network, such as Figure 1-2 As shown, the specific steps are as follows:

[0045] The load data of each appliance is selected from the household electricity consumption data set to obtain a total power sequence, and the total power sequence is preprocessed to obtain an initial sample data set;

[0046] The initial sample data set is expanded using the CGAN network to obtain an expanded sample data set;

[0047] Combining the improved residual network and the SCNN-BiLSTM network to build a load decomposition network model, using the improved residual network to extract features and output a high-dimensional feature map, reshaping the high-dimensional feature map, and using SCNN to extract spatial features from the reshaped high-dimensional feature map, and adjusting it to a feature matrix suitable for BiLSTM; using the expanded sample data set to train the load decomposition network model;

[0048] The trained load decomposition network model is used to identify and decompose the total household electricity load to obtain the load of different electrical appliances.

[0049] In a specific embodiment, the preprocessing includes:

[0050] According to the timestamp information of the load data of each electrical appliance, the load data of each electrical appliance is aligned;

[0051] Use the backward filling method to fill the missing values ​​in the load data of each electrical appliance after data alignment;

[0052] Assign category labels to the sample data based on the type of appliance in the total power series.

[0053] Specifically, the present invention uses the UKDALE dataset, which is a public household electricity data set, which contains detailed energy consumption data of household electricity in the UK. The load data set records the total active power consumption data of the whole house and the active power consumption data of more than ten types of electrical appliances of five households in the UK from November 2012 to January 2015 at a frequency of 1 / 6Hz. The load data of the corresponding electrical appliances are selected from the UKDALE dataset as the total power sequence; for example, the power load data of four electrical appliances, namely, a microwave oven, a refrigerator, a dishwasher and a washing machine, within a set time of a certain household are selected to obtain the total power sequence of electricity consumption. The load data of each electrical appliance are aligned according to the timestamp, and the missing values ​​in the data are supplemented by the backward filling method, that is, the missing value data is filled with the next valid value; however, there may still be some missing data at some time points, and these missing values ​​are deleted to ensure the integrity of the data to obtain a sample data. Then, according to the selected type of electrical appliances, the sample data category label is assigned.

[0054] In a specific embodiment, the CGAN network includes a generator and a discriminator; the generator generates sample data similar to the real data distribution by adding potential random noise to the total power sequence according to the total power sequence and the category label; the discriminator judges the authenticity of the input sample data according to the input sample data and the corresponding category label, and outputs a binary judgment result; the generator and the discriminator are alternately trained, and after the training is completed, the trained generator is used to generate new sample data to expand the initial sample data set.

[0055] The formula of CGAN is:

[0056]

[0057] Among them, S represents the generator, N represents the discriminator, z is random noise, l * is the objective function when training the generator and discriminator models, c represents the conditional information, E represents the expected value, S′, N′ represents the optimized generator S and discriminator N, and b represents the true value.

[0058] In a specific embodiment, the load decomposition network model uses an improved residual network to extract features from sample data in an expanded sample data set to obtain a high-dimensional feature map of the sample data, which is adjusted by a first reshaping layer to obtain a first feature map; uses an SCNN network to extract spatial features from the first feature map to obtain a second feature map, which is adjusted by a second reshaping layer to obtain a feature matrix; uses a BiLSTM network to convert the feature matrix into multiple time series tensors, and then extracts context features of the time series tensors from two directions, and the fully connected layer outputs the load power of each electrical appliance based on the context features.

[0059] In a specific embodiment, the improved residual network includes a starting stage, four main stages, and an ending stage; the first residual module of each stage reduces the size of the feature map by half through a convolution layer with a stride of 2, and increases the number of channels. The size of the feature map represents the spatial resolution (height and width) of the feature map, and the number of channels of the feature map is the number of feature maps composed of the extracted feature information. By reducing the size of the feature map, the amount of calculation and the amount of parameters can be reduced, and the feature information can be gradually aggregated to obtain a higher-level representation; each main stage uses the convolution layer in the residual module for feature extraction, and a batch normalization layer is connected after the convolution layer, and a ReLU activation function is connected after the batch normalization layer.

[0060] Specifically, the initial stage is used to adjust the dimension of the input features and perform preliminary feature extraction. The total power sequence is used as input data, and its features include time step information, such as total power values ​​per minute, hour, or day, and periodicity, such as daily peaks and valleys, or seasonal power changes. Through convolution, batch normalization operations increase the batch dimension, reshape it into a two-dimensional shape, and increase the channel dimension, the total power sequence can be adjusted to the feature dimension that adapts to the input requirements of iResNet.

[0061] The main stage is mainly used to perform feature extraction operations of the residual module to gradually refine the feature representation; it usually includes multiple convolutional layers, each of which is used for further feature extraction. The feature map after each convolution operation passes through a batch normalization layer to stabilize the training process and reduce internal covariate shift. The feature map after batch normalization passes through the ReLU activation function to increase the nonlinear representation ability of the network; the four main stages contain different residual modules, among which stages 1, 2, 3, and 4 contain 3, 4, 6, and 3 residual modules respectively. The end stage is used for the final processing and final adjustment of the feature map. The formula of the ReLU activation function is: G(x) = max(0, x); where G represents the activation function itself, x represents the input value, max represents the maximum value operation, and 0 represents the threshold.

[0062] At the end stage, the output feature map size of the improved residual network is (none, 2, 2, 512). Since the output of the residual network is a high-dimensional feature map, its shape may not be the shape expected by the subsequent convolutional layer, which will lead to model construction errors. Therefore, the output of the residual network is reshaped into a feature map with a larger spatial dimension and fewer channels through the first reshaping layer. The output feature map is sent to SCNN for spatial feature extraction. The reshaping operation here needs to ensure that the total number of elements remains unchanged, 2×2×512=32×32×2, so the adjusted new shape is (none, 32, 32, 2).

[0063] In a specific embodiment, the SCNN network slices the first feature map vertically and horizontally, performs convolution stacking operations on the slices after vertical slicing in a top-to-bottom and bottom-to-top order, and performs convolution stacking operations on the slices after horizontal slicing in a left-to-right and right-to-left order; the convolution stacking operation is to perform a small convolution on the first slice and superimpose the convolution result on the next slice until the superposition reaches the last slice.

[0064] The formula of the SCNN network is:

[0065]

[0066] Among them, X i,j,k is the input three-dimensional tensor; i, j, k represent the index of channel, row, and column, respectively, representing different electrical equipment, time, and corresponding feature information; H represents the maximum value of the feature map; f represents the nonlinear activation function ReLU; K m,u,n Represents the weight of the convolution kernel.

[0067] In K m,u,n Here m represents the row of the convolution kernel, u represents the depth of the input feature map, and n represents the column of the convolution kernel. The convolution kernel is a set of learnable weights used to detect features in the input feature map. Each element in the convolution kernel represents the weight applied to the corresponding element in the input feature map. The weights are adjusted during training to learn the most relevant features.

[0068] Since the output of the SCNN network is a feature map, and BiLSTM requires a two-dimensional feature matrix, the shape of which is composed of time steps and corresponding feature vectors, it needs to be adjusted to a feature matrix suitable for BiLSTM through a reshaping layer to be converted into a tensor about the time series, where each time step corresponds to a one-dimensional vector. The size of the feature map output by the previous layer of the second reshaping layer is (none, 4, 4, 5). After the reshaping operation, it is necessary to ensure that the total number of elements remains unchanged, 4×4×5=4×20, so the adjusted new shape is (none, 4, 20), where batch_size is the batch size.

[0069] In a specific embodiment, the BiLSTM network uses a recurrent connection to learn the global dependency between time series tensors. The formula is:

[0070]

[0071] in, and represents the input gate and output gate, σ represents the sigmoid function, and represents the weight matrix, represents the bias; represents the input vector of the lth layer at the i-th time step, and v represents the input i t The weight vector, v * Represents The weight vector of , t represents the time step, i t represents the input gate activation value at time step t, h t+1 represents the hidden state of the reverse LSTM at the next time step t+1.

[0072] In a specific embodiment, the output of the fully connected layer is y = [y1, y2, ..., y n ], where y is an n-dimensional vector, n is the total number of electrical appliances, y i Represents the power of the i-th appliance, and the fully connected layer is decomposed by the following formula:

[0073] y = σ(WF + b);

[0074] Among them, σ is the activation function, b is the bias vector, F is the feature input to the fully connected layer, and W is the weight matrix of the fully connected layer.

[0075] In a specific embodiment, the network parameters of the SResNet-SCNN-BiLSTM network are shown in Table 1.

[0076] Table 1 Network parameter table

[0077]

[0078]

[0079] In a specific embodiment, during training using an expanded sample data set:

[0080] Using MSE as the loss function, the formula is as follows:

[0081]

[0082] Among them, a i represents the eigenvector, g(a i ) represents the output vector, u i is the target label vector, and D represents the number of samples in the dataset;

[0083] The Adam optimizer is used to train the model, and the mean absolute error and signal polymer difference are used as evaluation indicators. The formula is as follows:

[0084]

[0085] Among them, MAE is the mean absolute error, SAE is the signal polymer difference, is the estimated power of appliance j at time t, is the actual value, T is the length of the research period; E are the actual total energy consumption and predicted total energy consumption of electrical equipment,

[0086] The embodiment of the present invention also discloses a non-intrusive load monitoring system based on the SResNet-SCNN-BiLSTM network, including:

[0087] Data acquisition module: selects the load data of each appliance from the household electricity consumption data set to obtain the total power sequence, and preprocesses the total power sequence to obtain the initial sample data set;

[0088] Data expansion module: Use the CGAN network to expand the initial sample data set to obtain an expanded sample data set;

[0089] Model training module: Combine the improved residual network and SCNN-BiLSTM network to build a load decomposition network model, use the improved residual network to extract features and output high-dimensional feature maps, reshape the high-dimensional feature maps, and use SCNN to extract spatial features from the reshaped high-dimensional feature maps and adjust them to feature matrices suitable for BiLSTM; use the expanded sample data set to train the load decomposition network model;

[0090] Identification module: Use the trained load decomposition network model to identify and decompose the total household electricity load to obtain the load of different electrical appliances.

[0091] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0092] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A non-intrusive load monitoring method based on SResNet-SCNN-BiLSTM network, characterized in that: The specific steps are as follows: Selecting the load data of each electrical appliance from the household electricity consumption data set to obtain a total power sequence, and preprocessing the total power sequence to obtain an initial sample data set; Expanding the initial sample data set using a CGAN network to obtain an expanded sample data set; Combining the improved residual network and the SCNN-BiLSTM network to build a load decomposition network model, using the improved residual network to extract features and output a high-dimensional feature map, reshaping the high-dimensional feature map, and using SCNN to extract spatial features from the reshaped high-dimensional feature map, and adjusting it to a feature matrix suitable for BiLSTM; using the expanded sample data set to train the load decomposition network model; Using the trained load decomposition network model, the total household electricity load is identified and decomposed to obtain the load of different electrical appliances; The load decomposition network model uses an improved residual network to extract features from the sample data in the expanded sample data set to obtain a high-dimensional feature map of the sample data, and obtains a first feature map after adjustment by a first reshaping layer; uses an SCNN network to extract spatial features from the first feature map to obtain a second feature map, and obtains a feature matrix after adjustment by a second reshaping layer; The feature matrix is ​​converted into multiple time series tensors using a BiLSTM network, and then context features of the time series tensors are extracted from two directions. The fully connected layer outputs the load power of each electrical appliance according to the context features. The improved residual network includes a starting stage, four main stages, and an ending stage; the first residual module of each stage reduces the size of the feature map by half through a convolution layer with a stride of 2, and increases the number of channels, the size of the feature map represents the spatial resolution of the feature map, and the number of channels of the feature map is the number of feature maps composed of the extracted feature information; each of the main stages uses the convolution layer in the residual module to extract features, a batch normalization layer is connected after the convolution layer, and a ReLU activation function is connected after the batch normalization layer; The SCNN network slices the first feature map vertically and horizontally, performs convolution stacking operations on the vertically sliced ​​slices in a top-to-bottom and bottom-to-top order, and performs convolution stacking operations on the horizontally sliced ​​slices in a left-to-right and right-to-left order; the convolution stacking operation is to perform a small convolution on the first slice and stack the convolution result on the next slice until the stacking reaches the last slice; The BiLSTM network uses a recurrent connection to learn the global dependency between the time series tensors. The formula is: in, and represents the input gate and output gate, σ represents the sigmoid function, and represents the weight matrix, represents the bias; represents the input vector of the lth layer at the i-th time step, and v represents the input i t The weight vector, v * Represents The weight vector of , t represents the time step, i t represents the input gate activation value at time step t, h t+1 represents the hidden state of the reverse LSTM at the next time step t+1; The output of the fully connected layer is y=[y1,y2,…,y n ], where y is an n-dimensional vector, n is the total number of electrical appliances, y i Represents the power of the i-th appliance, and the fully connected layer is decomposed by the following formula: y = σ(WF + b); Among them, σ is the activation function, b is the bias vector, and W is the weight matrix of the fully connected layer.

2. According to claim 1, a non-intrusive load monitoring method based on SResNet-SCNN-BiLSTM network is characterized in that: The pre-processing comprises: Performing data alignment on the load data of each electrical appliance according to the timestamp information of the load data of each electrical appliance; Use the backward filling method to fill the missing values ​​in the load data of each electrical appliance after data alignment; According to the types of electrical appliances in the total power sequence, a category label is assigned to the sample data.

3. According to claim 2, a non-intrusive load monitoring method based on SResNet-SCNN-BiLSTM network is characterized in that: The CGAN network includes a generator and a discriminator; the generator generates sample data similar to the real data distribution by adding potential random noise to the total power sequence according to the total power sequence and the category label; the discriminator judges the authenticity of the input sample data according to the input sample data and the corresponding category label, and outputs a binary judgment result; the generator and the discriminator are alternately trained, and after the training is completed, the trained generator is used to generate new sample data to expand the initial sample data set.

4. According to claim 1, a non-intrusive load monitoring method based on SResNet-SCNN-BiLSTM network is characterized in that: In the training using the expanded sample data set: Using MSE as the loss function, the formula is as follows: Among them, a i represents the eigenvector, g(a i ) represents the output vector, u i is the target label vector, and D represents the number of samples in the dataset; The Adam optimizer is used to train the model, and the mean absolute error and signal polymer difference are used as evaluation indicators. The formula is as follows: Among them, MAE is the mean absolute error, SAE is the signal polymer difference, is the estimated power of appliance j at time t, is the actual value, T is the length of the research period; E are the actual total energy consumption and predicted total energy consumption of electrical equipment, 5. A non-intrusive load monitoring system based on SResNet-SCNN-BiLSTM network, characterized in that: include: Data acquisition module: selects the load data of each appliance from the household electricity consumption data set to obtain a total power sequence, and preprocesses the total power sequence to obtain an initial sample data set; Data expansion module: using the CGAN network to expand the initial sample data set to obtain an expanded sample data set; Model training module: combining the improved residual network and the SCNN-BiLSTM network to build a load decomposition network model, using the improved residual network to extract features and output a high-dimensional feature map, reshaping the high-dimensional feature map, and using SCNN to extract spatial features from the reshaped high-dimensional feature map, and adjusting it to a feature matrix suitable for BiLSTM; using the expanded sample data set to train the load decomposition network model; Identification module: Use the trained load decomposition network model to identify and decompose the total household power load to obtain the load of different electrical appliances; The load decomposition network model uses an improved residual network to extract features from the sample data in the expanded sample data set to obtain a high-dimensional feature map of the sample data, and obtains a first feature map after adjustment by a first reshaping layer; uses an SCNN network to extract spatial features from the first feature map to obtain a second feature map, and obtains a feature matrix after adjustment by a second reshaping layer; The feature matrix is ​​converted into multiple time series tensors using a BiLSTM network, and then context features of the time series tensors are extracted from two directions. The fully connected layer outputs the load power of each electrical appliance according to the context features. The improved residual network includes a starting stage, four main stages, and an ending stage; the first residual module of each stage reduces the size of the feature map by half through a convolution layer with a stride of 2, and increases the number of channels, the size of the feature map represents the spatial resolution of the feature map, and the number of channels of the feature map is the number of feature maps composed of the extracted feature information; each of the main stages uses the convolution layer in the residual module to extract features, a batch normalization layer is connected after the convolution layer, and a ReLU activation function is connected after the batch normalization layer; The SCNN network slices the first feature map vertically and horizontally, performs convolution stacking operations on the vertically sliced ​​slices in a top-to-bottom and bottom-to-top order, and performs convolution stacking operations on the horizontally sliced ​​slices in a left-to-right and right-to-left order; the convolution stacking operation is to perform a small convolution on the first slice and stack the convolution result on the next slice until the stacking reaches the last slice; The BiLSTM network uses a recurrent connection to learn the global dependency between the time series tensors. The formula is: in, and represents the input gate and output gate, σ represents the sigmoid function, and represents the weight matrix, represents the bias; represents the input vector of the lth layer at the i-th time step, and v represents the input i t The weight vector, v * Represents The weight vector of , t represents the time step, i t represents the input gate activation value at time step t, h t+1 represents the hidden state of the reverse LSTM at the next time step t+1; The output of the fully connected layer is y=[y1,y2,…,y n ], where y is an n-dimensional vector, n is the total number of electrical appliances, y i Represents the power of the i-th appliance, and the fully connected layer is decomposed by the following formula: y = σ(WF + b); Among them, σ is the activation function, b is the bias vector, and W is the weight matrix of the fully connected layer.

Images