A non-intrusive load decomposition method based on attention

Abstract

The present invention belongs to the technical field of non-intrusive load decomposition, and specifically is a non-intrusive load decomposition method based on attention. The present invention adds an encoder and decoder module with a special attention unit to the regression subnetwork, so that the special attention unit is used in the regression subtask to enable the regression subnetwork to implicitly pay more attention to certain events (such as turning on or off a device) and specific signal parts (such as high power consumption) and make predictions; at the same time, a classification subnetwork is combined, and the execution result of the classification subnetwork assists the regression subnetwork in load decomposition by clarifying the on / off state of the device. The attention-based deep neural network improves the generalization ability of the network, and improves the network's ability to correctly detect the on and off state of the device and accurately locate the signal part with high power consumption in the non-intrusive load decomposition task.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-intrusive load decomposition, and in particular relates to a non-intrusive load decomposition method based on attention. Background Art

[0002] Electricity meters can only report the total power load data of a household. Therefore, how to infer the power load of a specific device and perform power load decomposition is the main research goal at this stage. How to use deep learning-related methods to improve the accuracy of load decomposition is a hot topic in current research. The non-intrusive load decomposition task can be regarded as a sequence-to-sequence (seq2seq) task, which inputs a string of aggregated signal data sequences and outputs a load signal sequence of a specific device (it is not ruled out that in some cases, the sequence-to-point method is used, using the midpoint of the sequence as a feature representative point). The traditional seq2seq network is usually a model composed of an encoder-decoder. Such a model usually has two challenges. First, the processing effect of long sequences is poor. The encoder discards the intermediate state and only uses the final hidden state vector to initialize the decoder; second, it cannot meet the alignment requirements between the input and output sequences. Summary of the invention

[0003] The purpose of the present invention is to provide an attention-based non-intrusive load decomposition method to improve the performance of the encoder-decoder network and enhance the load decomposition effect of deep neural networks in non-intrusive load decomposition tasks.

[0004] The attention-based non-intrusive load decomposition method provided by the present invention adds an encoder-decoder module with a special attention unit to the regression sub-network, so that the special attention unit is used in the regression sub-task to enable the regression sub-network to implicitly pay more attention to certain events (such as turning on or off a device) and specific signal parts (such as high power consumption) and make predictions; at the same time, a classification sub-network is combined, and the execution result of the classification sub-network assists the regression sub-network in load decomposition by clarifying the on / off state of the device. The attention-based deep neural network improves the generalization ability of the network, improves the performance of the network in the non-intrusive load decomposition task, can better align data, and the attention unit can achieve better utilization of intermediate states.

[0005] The non-intrusive load decomposition method based on attention provided by the present invention comprises the following specific steps:

[0006] Step 1: Data preprocessing, determine different sliding window sizes according to the device sampling frequency;

[0007] Step 2: Add attention units and build a regression subnetwork;

[0008] Step 3: Build a classification sub-network and determine the overall network structure;

[0009] Step 4: Determine the evaluation indicators and train the network model;

[0010] Step 5: Process the data generated by the overlapping windows;

[0011] Step 6: The final signal sequence is used as the appliance fingerprint for aggregated data load decomposition, thereby obtaining information related to the operation of the appliance.

[0012] Each step is further described below.

[0013] Step 1: Data preprocessing, determine different sliding window sizes based on the device sampling frequency

[0014] During the training phase of the network model, the sliding window technology is used to read the aggregate signal data. The input and output windows need to have appropriate window sizes. If they are too small, the device activation data of a certain device (the power consumption of the device during the entire cycle) cannot be fully captured. If they are too large, the device activation data of other devices will be taken into account, which will affect the final decomposition result.

[0015] Therefore, when performing load decomposition on aggregated data in a data set, it is necessary to process the data in advance and obtain the sampling rate of the aggregated load data in the data set, so as to obtain the window length L corresponding to each device.

[0016] Step 2: Add attention units and build a regression subnetwork

[0017] The model of the regression subnetwork is expressed as The mapping relationship is as follows:

[0018]

[0019]

[0020] Where i represents the device serial number. For example, if there are three types of electrical appliances that need to be identified, a, b, and c, and they are numbered in this order, then the device serial numbers i corresponding to the three electrical appliances are 1, 2, and 3, and a mapping relationship is established; x t,L Represents a data sequence of length L starting at time t, is the predicted output of the regression subnet, expressed as the device activation of length L for device i starting at time t.

[0021] The regression subnetwork is mainly composed of an encoder and a decoder. In the present invention, the attention unit is added between the encoder and the decoder, and the regression subnetwork is built in the order of data input to output.

[0022] Step 2.1: The encoder consists of a convolutional neural network (CNN) with four one-dimensional convolutional layers and a recurrent neural network (RNN) consisting of a bidirectional LSTM. The four one-dimensional convolutional layers have the same structure, with F filters, K convolution kernels, a stride of 1, and a ReLU activation function; the number of input and output units of the bidirectional LSTM is H, and a tangent hyperbolic activation function is used.

[0023] CNN processes the input aggregate load signal and extracts device-specific signatures as a set of feature maps. RNN takes a set of feature maps as input. Bidirectional LSTM obtains the hidden state of all information from two directions and generates a sequence containing the hidden state of all aggregate load information. Therefore, the final sequence h output by the encoder is t By connecting the two directions and The hidden state vector is obtained as:

[0024]

[0025] Step 2.2: The attention unit between the encoder and decoder consists of a single-layer feedforward neural network with H input and output units and a tangent hyperbolic activation function. The network calculates the attention weights and outputs the context vector C as the weighted average of the encoder output over time. Not all feature maps generated by CNN have the same contribution in identifying the device activation of the target device. The attention mechanism captures the significant activation feature map of the device by setting different weights. The mathematical expression of the attention unit is as follows:

[0026]

[0027] α t =softmax(e t ),

[0028]

[0029] Among them, V a , W a and b a It is an attention parameter learned jointly with other components of the model, and outputs a context vector C as the input of the decoder.

[0030] Step 2.3: The decoder consists of two fully connected layers. The first fully connected layer has H input units and L output units, and uses the ReLU activation function; the second fully connected layer has L input and output units and uses a linear activation function.

[0031] Step 3: Build a classification subnetwork and determine the overall network structure

[0032] Step 3.1: The model of the classification subnetwork is represented as The mapping relationship is as follows:

[0033]

[0034]

[0035] A typical network consists of eight layers:

[0036] The first layer is a one-dimensional convolution with a kernel size of 10, a filter number of 30, a stride of 1, and a ReLU activation function.

[0037] The second layer is a one-dimensional convolution with a kernel size of 8, a filter number of 30, a step size of 1, and a ReLU activation function.

[0038] The third layer is a one-dimensional convolution with a kernel size of 6, 40 filters, a step size of 1, and a ReLU activation function.

[0039] The fourth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a step size of 1, and a ReLU activation function.

[0040] The fifth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a stride of 1, and a ReLU activation function.

[0041] The sixth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a step size of 1, and a ReLU activation function.

[0042] The seventh layer is a fully convolutional layer with 1024 input neuron nodes and 1024 output neuron nodes, followed by a ReLU activation function.

[0043] The eighth layer is the full convolution layer, with 1024 input neuron nodes and L output neuron nodes, followed by a Sigmoid activation function;

[0044] Step 3.2: The overall structure of the network is as follows Figure 2 As shown, the final estimated output of the network Output for the regression subnetwork Probability distribution of the classification subnetwork output The result of multiplication, where ⊙ is component multiplication:

[0045]

[0046] Step 4: Determine the evaluation indicators and train the network model

[0047] Step 4.1: To measure the accuracy of the classification subnetwork in judging the on / off status of the device, three indicators are used for evaluation, namely F1 score, precision (P) and recall (R). The larger the values ​​of the three, the better the classification performance. TP is the number of correctly predicted positive samples, FP is the number of negative samples predicted as positive, and FN is the number of positive samples predicted as negative.

[0048]

[0049]

[0050] Step 4.2: To measure the output of the entire network, two indicators are used for evaluation, namely Mean Absolute Error (MAE) and Signal Aggregate Error (SAE). The smaller the two values, the better the model decomposition effect, as shown below:

[0051]

[0052]

[0053] Among them, y i (t) and They represent the actual load value and estimated load value of device i in time t, N is the number of disjoint time periods of length K, and r i (τ) and They represent the sum of the actual load and the sum of the predicted load in τ time periods respectively.

[0054] Step 5: Process the data generated by overlapping windows

[0055] The number of filters F in the regression subnetwork, the size K of each kernel, and the number of neurons H in the recurrent layer are optimized by grid search in the parameter space of F = {16, 32, 64}, K = {4, 8, 16}, and H = {256, 512, 1024}. Each device has its own parameter combination. Because the sliding window takes a sample length as the sliding distance, overlapping window data will be generated. The decomposed signal is reconstructed by using the median filter of the overlapping part to obtain the final output.

[0056] Step 6: The signal sequence finally obtained is used as the appliance fingerprint, thereby obtaining information related to the operation of the appliance.

[0057] The network model obtained according to the above steps and training methods can decompose the aggregated load signal of electrical appliances, extract the fingerprint of specific electrical appliances, and determine the current operating status of the appliance. When applied to household electric meter data, the operating status of known household appliances can be determined, and the appliances that are running under the current load can be identified.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] The present invention adds an encoder-decoder module with a special attention unit to the regression sub-network, so that the special attention unit is used in the regression sub-task to enable the regression sub-network to implicitly pay more attention to certain events (such as turning on or off a device) and specific signal parts (such as high power consumption) and make predictions; at the same time, a classification sub-network is combined, and the execution result of the classification sub-network assists the regression sub-network in load decomposition by clarifying the on / off state of the device. The attention-based deep neural network improves the generalization ability of the network, improves the network's ability to correctly detect the on and off state of the device and accurately locate the signal part with high power consumption in the non-intrusive load decomposition task, can better align data, and the attention unit can achieve better utilization of the intermediate state. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The flowchart of the attention-based non-intrusive load decomposition method of the present invention.

[0061] Figure 2 This is a structural diagram of the network model of the present invention. DETAILED DESCRIPTION

[0062] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0063] Example:

[0064] The present invention proposes a non-intrusive load decomposition method based on attention, and its flow chart is as follows: Figure 1 As shown, it can be divided into the following steps:

[0065] Step 1: Data preprocessing, determine different sliding window sizes according to the device sampling frequency;

[0066] Step 2: Add attention units and build a regression subnetwork;

[0067] Step 3: Build a classification sub-network and determine the overall network structure;

[0068] Step 4: Determine the evaluation indicators and train the network model;

[0069] Step 5: Process the data generated by the overlapping windows;

[0070] Step 6: The final signal sequence is used as the appliance fingerprint for aggregated data load decomposition, thereby obtaining information related to the operation of the appliance.

[0071] Each step is further described below.

[0072] 1. Data preprocessing, determine different sliding window sizes according to the device sampling frequency

[0073] During the training phase of the network model, the sliding window technology is used to read the aggregate signal data. The input and output windows need to have appropriate window sizes. If they are too small, the device activation data of a certain device (the power consumption of the device during the entire cycle) cannot be fully captured. If they are too large, the device activation data of other devices will be taken into account, which will affect the final decomposition result.

[0074] Therefore, when performing load decomposition on aggregated data in a data set, it is necessary to process the data in advance and obtain the sampling rate of the aggregated load data in the data set, so as to obtain the window length L corresponding to each device. In this embodiment, data from two data sets are processed. One is the Reference Energy Disaggregation Data Set (REDD) (JZKolter and MJJohnson, "Redd: A public data set for energydisaggregation research," in Workshop on data mining applicationsinsustainability (SIGKDD), San Diego, CA, vol. 25, 2011, pp. 59-62.) dataset. Considering the three devices with the highest power consumption: dishwasher (DW), refrigerator (FR) and microwave oven (MW), because the sampling frequency of REDD data is 3s (K = 3), the window lengths L corresponding to the three devices are 2304, 496 and 128 respectively; the other is the Domestic Appliance Level Electricity dataset (UK-DALE) (J.Kelly and W.Knotten belt, "The uk-dale dataset, domestic appliance level electricity demand and whole-house demandfrom five uk homes," Scientific data, vol.2, p.150007, 2015.) dataset, considering the five devices with the highest power consumption: dishwasher (DW), refrigerator (FR), kettle (KE), microwave oven (MW) and washing machine (WM). Since the sampling frequency of UK-DALE data is 6s (K=6), the window lengths L corresponding to the three devices are 1536, 512, 128, 288 and 1024 respectively.

[0075] 2. Add attention units and build regression subnetwork

[0076] Model representation of regression subnetwork And the mapping relationship is as follows:

[0077]

[0078]

[0079] Where i represents the device serial number. For example, if there are three types of electrical appliances that need to be identified, a, b, and c, and they are numbered in this order, then the device serial numbers i corresponding to the three electrical appliances are 1, 2, and 3, and a mapping relationship is established; x t,L Represents a data sequence of length L starting at time t, is the predicted output of the regression subnet, expressed as the device activation of length L for device i starting at time t.

[0080] The regression subnetwork is mainly composed of an encoder and a decoder. In the present invention, the attention unit is added between the encoder and the decoder, and the regression subnetwork is built in the order of data input to output.

[0081] Step 2.1: The encoder consists of a convolutional neural network (CNN) with four one-dimensional convolutional layers and a recurrent neural network (RNN) consisting of a bidirectional LSTM. The four one-dimensional convolutional layers have the same structure, with F filters, K convolution kernels, a stride of 1, and a ReLU activation function; the number of input and output units of the bidirectional LSTM is H, and a tangent hyperbolic activation function is used.

[0082] CNN processes the input aggregate load signal and extracts device-specific signatures as a set of feature maps. RNN takes a set of feature maps as input. Bidirectional LSTM obtains the hidden state of all information from two directions and generates a sequence containing the hidden state of all aggregate load information. Therefore, the final sequence h output by the encoder is t By connecting the two directions and The hidden state vector is obtained as:

[0083]

[0084] Step 2.2: The attention unit between the encoder and decoder consists of a single-layer feedforward neural network with H input and output units and a tangent hyperbolic activation function. The network calculates the attention weights and outputs the context vector C as the weighted average of the encoder output over time. Not all feature maps generated by CNN have the same contribution in identifying the device activation of the target device. The attention mechanism captures the significant activation feature map of the device by setting different weights. The mathematical expression of the attention unit is as follows:

[0085]

[0086] α t =softma(e t ),

[0087]

[0088] Where Va , W a and b a It is an attention parameter learned jointly with other components of the model, and outputs a context vector C as the input of the decoder.

[0089] Step 2.3: The decoder consists of two fully connected layers. The first fully connected layer has H input units and L output units, and uses the ReLU activation function; the second fully connected layer has L input and output units and uses a linear activation function.

[0090] 3. Build a classification sub-network and determine the overall network structure

[0091] Step 3.1: Model representation of the classification subnetwork And the mapping relationship is as follows:

[0092]

[0093]

[0094] The network consists of eight layers:

[0095] The first layer is a one-dimensional convolution with a kernel size of 10, a filter number of 30, a stride of 1, and a ReLU activation function.

[0096] The second layer is a one-dimensional convolution with a kernel size of 8, a filter number of 30, a step size of 1, and a ReLU activation function.

[0097] The third layer is a one-dimensional convolution with a kernel size of 6, 40 filters, a step size of 1, and a ReLU activation function.

[0098] The fourth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a step size of 1, and a ReLU activation function.

[0099] The fifth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a stride of 1, and a ReLU activation function.

[0100] The sixth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a step size of 1, and a ReLU activation function.

[0101] The seventh layer is a fully convolutional layer with 1024 input neuron nodes and 1024 output neuron nodes, followed by a ReLU activation function.

[0102] The eighth layer is the fully convolutional layer. The number of input neuron nodes is 1024, the number of output neuron nodes is L, and it is followed by a Sigmoid activation function.

[0103] Step 3.2: The overall structure of the network is as follows Figure 2As shown, the final estimated output of the network Output for the regression subnetwork Probability distribution of the classification subnetwork output The result of multiplication, where ⊙ is component multiplication:

[0104]

[0105] 4. Determine evaluation indicators and train network models

[0106] Step 4.1: Measure the accuracy of the classification sub-network in judging the on / off status of the device. Three indicators are used for evaluation, namely F1 score, precision (P) and recall (R). The larger the values ​​of the three, the better the classification performance. TP is the number of correctly predicted positive samples, FP is the number of negative samples predicted as positive, and FN is the number of positive samples predicted as negative.

[0107]

[0108]

[0109] Step 4.2: Measure the output of the entire network. Two indicators are used for evaluation, namely Mean Absolute Error (MAE) and Signal Aggregate Error (SAE). The smaller the two values, the better the model decomposition effect, as shown below, where y i (t) and They represent the actual load value and estimated load value of device i in time t, N is the number of disjoint time periods of length K, and r i (τ) and They represent the sum of the actual load and the sum of the predicted load in τ time periods respectively.

[0110]

[0111]

[0112] 5. Processing data generated by overlapping windows

[0113] The number of filters F in the regression subnetwork, the size K of each kernel, and the number of neurons H in the recurrent layer are optimized using grid search in the parameter space of F = {16, 32, 64}, K = {4, 8, 16}, and H = {256, 512, 1024}, and each device has its own parameter combination. For the above devices, the encoder parameters of DW are set to F = 32, K = 16; the encoder parameters of FR are set to F = 32, K = 8; the encoder parameters of MW are set to F = 16, K = 8, etc., which are the best values ​​obtained during the hyperparameter optimization process. Because the sliding window takes a sample length as the sliding distance, overlapping window data will be generated. The final output is obtained by reconstructing the decomposed signal using the median filter of the overlapping part.

[0114] 6. The final signal sequence is used as the appliance fingerprint to obtain information related to the operation of the appliance

[0115] The network model obtained according to the above steps and training methods can decompose the aggregated load signal of electrical appliances, extract the fingerprint of specific electrical appliances, and determine the current operating status of the appliance. When applied to household electric meter data, the operating status of known household appliances can be determined, and the appliances that are running under the current load can be identified.

[0116] The network model extracts and identifies the appliance data of dishwashers, refrigerators, and microwave ovens in the two data sets, and verifies the output results of the classification subnetwork and the output results of the entire network according to the evaluation indicators in step 4, and obtains the results in Table 1, where the overall evaluation is the average of all values ​​in the current row. The data processing results for refrigerators are relatively poor, which may be due to the large data changes caused by the electrical characteristics of refrigerators and the changeable working state. The model is not perfect enough to learn its operating data. However, for appliances such as dishwashers and microwave ovens that only have two state changes, namely, turning on to wash dishes and turning off to stop working, the processing results are relatively good. Table 1 also shows the comparison between this method and some other existing non-intrusive load decomposition methods. Overall, the network model in this paper has a relatively good experimental result. Among them, FHMM (N.Batra, J.Kelly, O.Parson, H.Dutta, W.Knottenbelt, A.Rogers, A.Singh, and M.Srivastava, "Nilmtk: an open source toolkit for non-intrusive load monitoring," in Proceedings of the 5th international conference on Futureenergy systems, 2014, pp.265–276.) is a traditional non-deep learning method, and SCSNet is a method proposed by (K.Chen, Y.Zhang, Q.Wang, J.Hu, H.Fan, and J.He, "Scale and context-aware convolutional non-intrusive load monitoring," IEEE Transactions on Power Systems, 2019).

[0117] Table 1 Evaluation of load decomposition and extraction results

[0118]

Claims

1. A non-intrusive load decomposition method based on attention, characterized in that: The specific steps are: Step 1: Data preprocessing, determining different sliding window sizes according to the device sampling frequency; including: obtaining the sampling rate of the aggregated load data in the data set, and obtaining the window length L corresponding to each device; Step 2: Add attention units and build a regression subnetwork; specifically: The model of the regression subnetwork is expressed as The mapping relationship is as follows: Where i represents the device serial number; x t,L Represents a data sequence of length L starting at time t, is the predicted output of the regression subnet, represented as the device activation of length L starting from device i at time t; The regression subnetwork is mainly composed of an encoder and a decoder. The attention unit is added between the encoder and the decoder, and the regression subnetwork is built in the order from data input to output. Step 3: Build a classification sub-network and determine the overall network structure; specifically include: The model of the classification subnetwork is expressed as The mapping relationship is as follows: The regression subnetwork and the classification subnetwork form the overall network structure. The estimated output of the network Output for the regression subnetwork Probability distribution of the classification subnetwork output The result of multiplication, where ⊙ is component multiplication: Step 4: Determine the evaluation indicators and train the network model; specifically include: In order to measure the accuracy of the classification sub-network in judging the on / off status of the device, three indicators are used for evaluation, namely F1 score, precision P and recall R. The larger the values ​​of the three, the better the classification performance. Let TP be the number of correctly predicted positive samples, FP be the number of negative samples predicted as positive, and FN be the number of positive samples predicted as negative, then: In order to measure the output results of the entire network, two indicators are used for evaluation: mean absolute error MAE and signal aggregation error SAE. The smaller the two values, the better the model decomposition effect. The specific expression is as follows: Among them, y i (t) and They represent the actual load value and estimated load value of device i in time t, N is the number of disjoint time periods of length K, and r i (τ) and They represent the sum of the actual load and the sum of the predicted load in τ time periods respectively; Step 5: Process the data generated by the overlapping windows; specifically including: Assume the number of filters F in the regression subnetwork, the size of each kernel K, and the number of neurons H in the recurrent layer. Use grid search to perform hyperparameter optimization in the parameter space of F = {16, 32, 64}, K = {4, 8, 16}, and H = {256, 512, 1024}. Each device has its own parameter combination. Since the sliding window takes a sample length as the sliding distance, overlapping window data will be generated. The decomposed signal is reconstructed by using the median filter of the overlapping part to obtain the final output. Step 6: The final signal sequence is used as the appliance fingerprint for aggregated data load decomposition, thereby obtaining information related to the operation of the appliance.

2. The attention-based non-intrusive load decomposition method according to claim 1, characterized in that: In step 2, the encoder is composed of a convolutional neural network (CNN) having four one-dimensional convolutional layers and a recurrent neural network (RNN) consisting of a bidirectional LSTM; the four one-dimensional convolutional layers have the same structure, all of which are F filters, K convolution kernels, a step size of 1, and all use a ReLU activation function; the number of input and output units of the bidirectional LSTM is H, and a tangent hyperbolic activation function is used; The CNN processes the input aggregate load signal and extracts the device-specific signature as a set of feature maps. The RNN takes a set of feature maps as input, and the bidirectional LSTM obtains the hidden state of all information from two directions to generate a sequence containing the hidden state of all aggregate load information. Therefore, the final sequence h output by the encoder is t By connecting the two directions and The hidden state vector is obtained as: The attention unit between the encoder and the decoder consists of a single-layer feedforward neural network with H input and output units and a tangent hyperbolic activation function. The network calculates the attention weights and outputs a context vector C as the weighted average of the encoder output over time. Since all feature maps generated by the CNN do not have the same contribution in identifying the device activation of the target device, the attention mechanism captures the significant activation feature maps of the device by setting different weights. The mathematical expression of the attention unit is as follows: a t =softmax(e t ); Among them, V a , W a and b a It is the attention parameter learned together with other components of the model, outputting the context vector C as the input of the decoder; The decoder consists of two fully connected layers. The number of input units and the number of output units of the first fully connected layer are H, and the ReLU activation function is adopted. The number of input and output units of the second fully connected layer are both L, and the linear activation function is adopted.

3. The attention-based non-intrusive load decomposition method according to claim 2, characterized in that: In step 3, the classification subnetwork has an eight-layer structure: The first layer is a one-dimensional convolution with a kernel size of 10, a filter number of 30, a stride of 1, and a ReLU activation function. The second layer is a one-dimensional convolution with a kernel size of 8, a filter number of 30, a step size of 1, and a ReLU activation function. The third layer is a one-dimensional convolution with a kernel size of 6, 40 filters, a step size of 1, and a ReLU activation function. The fourth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a step size of 1, and a ReLU activation function. The fifth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a stride of 1, and a ReLU activation function. The sixth layer is a one-dimensional convolution with a kernel size of 5, a filter number of 50, a step size of 1, and a ReLU activation function. The seventh layer is a fully convolutional layer with 1024 input neuron nodes and 1024 output neuron nodes, followed by a ReLU activation function. The eighth layer is the fully convolutional layer. The number of input neuron nodes is 1024, the number of output neuron nodes is L, and it is followed by a Sigmoid activation function.

4. The attention-based non-intrusive load decomposition method according to claim 3, characterized in that: The signal sequence finally obtained in step 6 is used as the appliance fingerprint of the aggregated data load decomposition, so as to obtain the information related to the operation of the appliance, specifically including: using the trained network model to load decompose the aggregated load signal of the appliance, extracting the fingerprint of the specific appliance, judging the current operating status of the appliance, using the household meter data to judge the operating status of known household appliances, and identifying the appliances running under the current load.

Images