A deep learning-based non-intrusive load decomposition method and system

Abstract

The application discloses a kind of non-invasive load decomposition method and system based on deep learning, it is related to electric power load management analysis technical field, including the following steps: obtaining total table data and subtable data, based on total table data, the missing value of subtable data is filled, and the sampling frequency of power data of total table and subtable is unified, wherein the total table data is responsible for sampling total power, and the subtable data is responsible for sampling each load power;Limited state class load event is detected based on the method of sliding window and step determination, and the refinement of total power to limited state class load is based on the preset neural network;Limited state class load is separated from total power signal, and the remaining power signal is obtained, the operating mode of power continuous change class load in remaining power is learned based on the preset neural network, and the decomposition of power consumption is realized.

Description

Technical Field

[0001] This invention relates to the field of power load management and analysis technology, specifically a non-intrusive load decomposition method and system based on deep learning. Background Technology

[0002] With the popularization of IoT, integration, and communication technologies, the market share of devices with continuously varying power is constantly increasing. Especially in recent years, with the development of clean energy and the transformation of the energy structure, photovoltaics, as a typical representative of continuously varying power loads, is gradually entering homes. This trend presents new challenges to non-intrusive load monitoring and identification. Decomposing finite-state loads under continuous power superposition and modeling the characteristics of continuously varying power loads are of great significance for optimizing the operation of power distribution systems. Real-time monitoring of photovoltaic power generation can provide power companies with data support for optimizing energy structure and improving power quality. It also allows users to understand the power generation of their residential photovoltaic systems, thereby optimizing their electricity consumption behavior to maximize the utilization of solar energy.

[0003] Patent CN118606670A discloses a load identification method and load monitoring device based on a genetic algorithm. This method uses a genetic algorithm to perform crossover or mutation operations on individuals in the initial identification population of electrical appliances to obtain the electricity consumption behavior identification results. However, it does not explicitly consider normally open loads with continuously changing power, resulting in insufficient model generalization ability. Patent CN117996838B discloses a distributed photovoltaic identification device based on non-intrusive load monitoring. This device employs non-intrusive load monitoring technology for distributed photovoltaic power generation systems, utilizing low-complexity sliding window and support vector machine techniques to decompose distributed photovoltaic power generation from the main feeder of the load. However, it requires processing a large amount of electrical and non-electrical parameter data, placing high demands on data processing and storage. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a non-intrusive load decomposition method and system based on deep learning. This method can separate finite-state loads and continuously variable power loads, refine the power consumption of each load, and intuitively compare the predicted power consumption with the actual power consumption of each load. The resulting power decomposition results have high accuracy and can provide more comprehensive data support for optimizing energy management and energy conservation and emission reduction.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a non-intrusive load decomposition method based on deep learning, the method comprising the following steps:

[0006] Obtain the master table data and sub-table data, fill in the missing values ​​of the sub-table data based on the master table data, and unify the sampling frequency of the power data of the master table and the sub-table. The master table data is responsible for sampling the total power, and the sub-table data is responsible for sampling the power of each load.

[0007] The method based on sliding window and step decision is used to detect finite state load events, and the total power is refined into finite state loads based on a preset neural network.

[0008] The finite-state loads are separated from the total power signal to obtain the remaining power signal. Based on a preset neural network, the operating modes of the continuously changing loads in the remaining power are learned and separated to achieve power decomposition.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: preprocessing the original total power sequence and each load power sequence in the total table data and the sub-table data respectively, filling in missing values, and unifying the sampling frequency of the sub-table at each load with the sampling frequency of the total table at the household entrance.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of detecting finite-state load events using the sliding window and step decision-making method:

[0011] The trend of the total power sequence is fitted within a sliding window, and a sudden increase is detected at the end. Once an input event is detected, the sampled value is replaced with a predicted value that conforms to the trend of the sequence value within the sliding window while marking the sampling period.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the formula for determining the occurrence of the finite state class load event is:

[0013]

[0014] Where t is the sampling time, s represents the s-th sampling period, and α + For pre-assigned thresholds;

[0015] The fitted expression for the values ​​within the sliding window before the input event occurs is calculated using the least squares method:

[0016] P = k × t + b

[0017]

[0018] Where k is the weight of the expression, b is the bias of the expression, h is the sliding window length, and t is the mean of the sampling time t before the input event occurs;

[0019] Replace the sampled value with the predicted value P.n :

[0020] P n =k×t n +b

[0021] For loads that have multiple state transitions and intermittent operation during a single switching event, a time delay t is set. delay The time interval will be less than t delay Multiple marking events are merged into a single load switching event;

[0022] Different threshold ranges are set according to the rated power of the equipment to filter out events from different equipment. If there are overlapping load events, multiple event detections are performed to separate events from different loads.

[0023] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the preset neural network uses a convolutional neural network with an attention mechanism and a deep learning neural network with a bidirectional long short-term memory network to establish a unique mapping relationship between the total load and each target load, defining the relationship between the total power load power sequence Y of a household at time t and the power sequence X of a single load as follows:

[0024]

[0025] Where a is a state variable consisting of 0-1 variables, 0 indicates that the device is disconnected, 1 indicates that the device is running, N indicates that there are N loads, and the function f represents the NILM mapping method.

[0026] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the decomposition process of the finite state class load is as follows:

[0027] Using the load active power slice sequence as input and the actual load active power as output, the time window is cut before the one-dimensional data is input into the neural network, thus cutting the long time series into short segments.

[0028] The data is divided into training and test sets according to a certain ratio. The training set is used to learn the pattern, weights and biases. After training, the generalization ability of the model and system is verified on the test set.

[0029] Before inputting the data into the model and system, normalization is performed to convert it to the same scale, mapping the data to (-1,1):

[0030]

[0031] Where x t and x t 'These are the data before and after normalization, x' and 'x' respectively. minand x max These are the minimum and maximum values, respectively.

[0032] Setting the CNN as the first layer of the neural network, local features are extracted from the input data through layer-by-layer filtering to learn the unique power consumption pattern of each load. The calculation formulas for the convolutional layer and pooling layer are as follows:

[0033]

[0034] Where x represents the input data, which in the convolutional layer formula is the load power data after slicing the total power, and W is the convolutional kernel matrix. For convolution operations, b is the bias of the convolutional layer. α is the activation function, β is the pooling layer weight, β is the pooling layer bias, and pool(·) is the pooling function.

[0035] By setting BiLSTM as the second layer of the neural network, both local and global information are considered in the temporal data. The output h is obtained by fitting the hidden layer computation results of the forward and backward iterations. t :

[0036]

[0037] Among them, w t and v t For the weight vector, The hidden layer states during forward iteration. For the hidden layer states in the backward iteration, n t This represents the offset of the hidden layer state at that moment;

[0038] In the attention mechanism, different levels of attention are assigned to different parts of the input data. A linear mapping is performed on the input sequence X to obtain matrices Q, K, and V, which are linearly mapped from the input matrix. The formula is as follows:

[0039] Q = W q X

[0040] K = W k X

[0041] V = W v X

[0042] Output:

[0043]

[0044] Where D K The matrix used to stabilize the training gradients is softmax(·), which is the normalization function.

[0045] The input x will be in the fully connected layer. l-1 With weight matrix wl Multiply, and add the bias term b l Then, it is mapped to the output value y through an activation function. l Integrate the features extracted from the aforementioned networks:

[0046] y l =w l x l-1 +b l

[0047] Data denormalization is performed to restore the (-1,1) interval data processed in the model to the original data range, and the final output is the active power decomposition sequence of the load.

[0048] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: when separating the finite-state class load from the total power signal to obtain the residual power signal:

[0049] Noise is divided into random noise and non-random noise. It is assumed that when no load is applied, the total active power data collected by the main meter is non-random noise, with a set duration t. noise The detected single occurrence time is less than t noise The marked event is determined to be random noise caused by glitch pulses during sampling. The formula for separating the remaining power is:

[0050]

[0051] Where P j The total power is denoted by , j represents the j-th sampling period, n represents the number of loads that are on and consuming power during the j-th sampling period, and N represents noise.

[0052] Secondly, in order to achieve the above objectives, this invention discloses a non-intrusive load decomposition system based on deep learning, comprising:

[0053] The data processing module is used to acquire master table data and sub-table data, fill missing values ​​in sub-table data based on master table data, and unify the sampling frequency of power data in master table and sub-table. The master table data is responsible for sampling total power, and the sub-table data is responsible for sampling the power of each load.

[0054] The load detection module is used to detect finite-state load events based on the sliding window and step decision method, and refines the total power to finite-state loads based on a preset neural network.

[0055] The load decomposition module is used to separate finite-state loads from the total power signal to obtain the remaining power signal. Based on a preset neural network, it learns and separates the operating modes of continuously changing loads in the remaining power to achieve power decomposition.

[0056] In another aspect of the present invention, in order to achieve the above-mentioned objective, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor, and when the processor loads and executes the computer program, it employs a non-intrusive load decomposition method based on deep learning as described above.

[0057] In another aspect of the present invention, in order to achieve the above-mentioned objective, a computer-readable storage medium is disclosed, wherein a computer program is stored in the computer program, and when the computer program is loaded and executed by a processor, a non-intrusive load decomposition method based on deep learning as described above is employed.

[0058] The beneficial effects of this invention are:

[0059] This invention can detect finite-state load switching events through real-time power mutations under continuous power superposition interference. It considers the diversity of load characteristics, accurately locating the timing of different load switching events while reducing data storage requirements, providing more precise input data for subsequent load decomposition models and systems. This invention considers the actual composition of total power data. Using only total active power data as source data, it leverages the mapping capabilities of neural networks to eliminate errors in event detection, other load decomposition errors, and interference from unknown noise to a certain extent. This achieves accurate decomposition of finite-state loads and loads with continuously changing power, significantly reducing the need for data sampling. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0062] Figure 2 This is a schematic diagram of the algorithm flow according to an embodiment of the present invention;

[0063] Figure 3 This is a diagram illustrating the replacement of predicted event detection values.

[0064] Figure 4 This is a diagram showing the steps of load decomposition for finite-state classes.

[0065] Figure 5 This is a diagram showing the decomposition effect of the load's active power.

[0066] Figure 6 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Example 1:

[0069] like Figure 1 As shown, a non-intrusive load decomposition method based on deep learning includes the following steps:

[0070] S101: Obtain the total table data and sub-table data, fill in the missing values ​​of the sub-table data based on the total table data, and unify the sampling frequency of the power data of the total table and the sub-table. The total table data is responsible for sampling the total power, and the sub-table data is responsible for sampling the power of each load.

[0071] The total table data and sub-table data are preprocessed to fill missing values ​​for the original total power sequence and each load power sequence, and the sampling frequency of each load sub-table is unified with the sampling frequency of the total table at the household entrance.

[0072] S102: The method based on sliding window and step decision is used to detect finite state load events, and the total power is refined into finite state loads based on a preset neural network.

[0073] The process of detecting finite-state load events using the sliding window and step decision method:

[0074] The trend of the total power sequence is fitted within a sliding window, and a sudden increase is detected at the end. Once an input event is detected, the sampled value is replaced with a predicted value that conforms to the trend of the sequence value within the sliding window while marking the sampling period.

[0075] The formula for determining the occurrence of the finite-state load event is as follows:

[0076]

[0077] Where t is the sampling time, s represents the s-th sampling period, and α + For pre-assigned thresholds;

[0078] The fitted expression for the values ​​within the sliding window before the input event occurs is calculated using the least squares method:

[0079] P = k × t + b

[0080]

[0081] Where k is the weight of the expression, b is the bias of the expression, h is the sliding window length, and t is the mean of the sampling time t before the input event occurs;

[0082] Replace the sampled value with the predicted value P. n :

[0083] P n =k×t n +b

[0084] For loads that have multiple state transitions and intermittent operation during a single switching event, a time delay t is set. delay The time interval will be less than t delay Multiple marking events are merged into a single load switching event;

[0085] Different threshold ranges are set according to the rated power of the equipment to filter out events from different equipment. If there are overlapping load events, multiple event detections are performed to separate events from different loads.

[0086] The preset neural network uses a convolutional neural network with an attention mechanism and a deep learning neural network with a bidirectional long short-term memory network to establish a unique mapping relationship between the total load and each target load. The relationship between the total power load power sequence Y of a household at time t and the power sequence X of a single load is defined as follows:

[0087]

[0088] Where a is a state variable consisting of 0-1 variables, 0 indicates that the device is disconnected, 1 indicates that the device is running, N indicates that there are N loads, and the function f represents the NILM mapping method.

[0089] The decomposition process of the finite-state load is as follows:

[0090] Using the load active power slice sequence as input and the actual load active power as output, the time window is cut before the one-dimensional data is input into the neural network, thus cutting the long time series into short segments.

[0091] The data is divided into training and test sets according to a certain ratio. The training set is used to learn the pattern, weights and biases. After training, the generalization ability of the model and system is verified on the test set.

[0092] Before inputting the data into the model and system, normalization is performed to convert it to the same scale, mapping the data to (-1,1):

[0093]

[0094] Where x t and x t 'These are the data before and after normalization, x' and 'x' respectively. min and x max These are the minimum and maximum values, respectively.

[0095] Setting the CNN as the first layer of the neural network, local features are extracted from the input data through layer-by-layer filtering to learn the unique power consumption pattern of each load. The calculation formulas for the convolutional layer and pooling layer are as follows:

[0096]

[0097] Where x represents the input data, which in the convolutional layer formula is the load power data after slicing the total power, and W is the convolutional kernel matrix. For convolution operations, b is the bias of the convolutional layer. α is the activation function, β is the pooling layer weight, β is the pooling layer bias, and pool(·) is the pooling function.

[0098] By setting BiLSTM as the second layer of the neural network, both local and global information are considered in the temporal data. The output h is obtained by fitting the hidden layer computation results of the forward and backward iterations. t :

[0099]

[0100] Among them, w t and v t For the weight vector, The hidden layer states during forward iteration. For the hidden layer states in the backward iteration, n t This represents the offset of the hidden layer state at that moment;

[0101] In the attention mechanism, different parts of the input data are given different levels of attention, making the network more focused on the time period in which the event occurred. A linear mapping is performed on the input sequence X to obtain matrices Q, K, and V that are linearly mapped from the input matrix, as shown in the formula:

[0102] Q = W q X

[0103] K = W k X

[0104] V = W v X

[0105] Output:

[0106]

[0107] Where D K The matrix used to stabilize the training gradients is softmax(·), which is the normalization function.

[0108] The input x will be in the fully connected layer. l-1 With weight matrix w l Multiply, and add the bias term b l Then, it is mapped to the output value y through an activation function. l Integrate the features extracted from the aforementioned networks:

[0109] y l =w l x l-1 +b l

[0110] Data denormalization is performed to restore the (-1,1) interval data processed in the model to the original data range, and the final output is the active power decomposition sequence of the load.

[0111] S103: Separate the finite-state loads from the total power signal to obtain the remaining power signal. Based on a preset neural network, learn and separate the operating modes of the continuously changing loads in the remaining power to achieve power decomposition.

[0112] When separating the finite-state loads from the total power signal to obtain the remaining power signal:

[0113] Noise is divided into random noise and non-random noise. It is assumed that when no load is applied, the total active power data collected by the main meter is non-random noise, with a set duration t. noise The detected single occurrence time is less than t noise The marked event is determined to be random noise caused by glitch pulses during sampling. The formula for separating the remaining power is:

[0114]

[0115] Where P j The total power is represented by j, where j represents the j-th sampling period, n represents the number of loads that are active and consuming power during the j-th sampling period, and N represents the noise.

[0116] Specifically, the present invention will be further illustrated below through embodiments:

[0117] The load of residential users was simulated for 24 hours in the LVNS simulation platform to obtain the corresponding electricity consumption data as an application scenario. The method and system of this invention were compared with the results of load decomposition using only CNN and CNN-BiLSTM neural network frameworks. The residual value was separated and the neural network prediction value of photovoltaics was compared. The settings of each load model are shown in Table 1.

[0118] Table 1 Load Model Settings

[0119] equipment Load type Rated power (W) quantity dryer Multi-state load 4000 1 stove Multi-state load 1500 1 Photovoltaics Loads with continuously varying power 1500 1 washing machine Multi-state load 180 1 coffee machine Multi-state load 920 1 Incandescent lamp 1 Single-state load 40 8 Incandescent lamp 2 Single-state load 40 6 food processor Single-state load 800 1 water heater Multi-state load 1200 1

[0120] Evaluation metrics used in machine learning include F1 score, mean squared error (MSE), and coefficient of determination (R²). 2 This serves as an indicator of the accuracy of load decomposition. Table 2 shows a comparison of the load decomposition results for each method of this invention, and the decomposition prediction curves are as follows: Figure 5 As shown;

[0121] Table 2 Load decomposition results of the method and system of the present invention

[0122]

[0123] The parameter performance indicators and fitting trends of the decomposition results show that the model and system in this paper have good decomposition performance for each load. The model and system can effectively understand and learn the unique operating characteristics of each load under the superimposed influence of continuously changing photovoltaic power and sampling noise, and achieve accurate decomposition from the total active power slice sequence to the actual active power sequence of each load.

[0124] Example 2: Second aspect, such as Figure 6 As shown, in order to achieve the above objectives, this invention discloses a non-intrusive load decomposition system based on deep learning, comprising:

[0125] The data processing module 11 is used to acquire total table data and sub-table data, fill missing values ​​in the sub-table data based on the total table data, and unify the sampling frequency of power data in the total table and sub-table. The total table data is responsible for sampling the total power, and the sub-table data is responsible for sampling the power of each load.

[0126] The load detection module 12 is used to detect finite-state load events based on the sliding window and step decision method, and to refine the total power to finite-state loads based on a preset neural network.

[0127] The load decomposition module 13 is used to separate finite-state loads from the total power signal to obtain the remaining power signal. Based on a preset neural network, it learns and separates the operating modes of continuously changing loads in the remaining power to achieve power decomposition.

[0128] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0129] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0130] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0131] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A deep learning-based non-intrusive load decomposition method, characterized by, The method comprises the following steps: The method comprises the following steps: Obtain total table data and sub-table data, fill in missing values of the sub-table data based on the total table data, and unify the sampling frequencies of the power data of the total table and the sub-table, wherein the total table data is responsible for sampling total power, and the sub-table data is responsible for sampling each load power; Detect the finite state type load event based on the sliding window and step judgment method, and refine the total power to the finite state type load based on a preset neural network; The process of detecting the finite state type load event based on the sliding window and step judgment method: Fit the change trend of the total power sequence in the sliding window, detect the sudden increase at the end, and once the switching-on event is detected, replace the sampling value with the predicted value conforming to the change trend of the sequence value in the sliding window while marking the sampling period; wherein is the sampling time, denotes the sample period, is a preassigned threshold value; The judgment formula for the occurrence of the finite state type load event is: wherein is a weight for the expression, is a bias for the expression, is a length of the sliding window, is a mean of the sampling time before the occurrence of the input event, is a mean of the sampling time before the occurrence of the input event. Replacing sample values with predicted values : For the load with multiple state switching and intermittent operation in one switching event, set the time delay The multiple marking events with time interval less than are combined into one load switching event. Calculate the fitting expression of the value in the sliding window before the switching-on event occurs by the least square method: Set different threshold intervals according to the rated power of the equipment, screen out the events of different equipment, and if there are overlapping load events, detect the events of different loads multiple times; Separate the finite state type load from the total power signal to obtain a residual power signal, learn the operation mode of the power continuous change type load in the residual power based on a preset neural network, and realize the decomposition of power consumption; The decomposition process of the finite state type load is as follows: Cut the load active power slice sequence as input, and the load actual active power as output, before inputting one-dimensional data into the neural network, cut the long time sequence into short segments through time window cutting; Normalization is performed before inputting data into the model and system, converting to the same scale, mapping data to : wherein and are the pre-normalized and normalized data, respectively, and are the minimum and maximum values, respectively; Divide the data into a training set and a test set according to a proportion, learn the mode, weight and bias through the training set, and verify the model and the generalization ability of the system on the test set after the training is completed. wherein is the input data, in the convolution layer formula, it is the load power data after total power slicing, is the convolution kernel matrix, is the convolution operation, is the convolution layer bias quantity, is the activation function, is the pooling layer weight, is the pooling layer bias quantity, is the pooling function; The BiLSTM is set as the second layer of the neural network, local and global information in the time series data is considered simultaneously, and the output is fitted by calculating the results of the hidden layers of the forward iteration and the backward iteration : wherein, and is a weight vector, is a hidden layer state of the forward iteration, is a hidden layer state of the backward iteration, is a bias of the hidden layer state at the time. Different parts of the input data are given different degrees of attention in the attention mechanism, and the input sequence is linearly mapped to obtain a matrix for linearly mapping the input matrix , and , the formula is: Set the CNN as the first layer of the neural network, extract the local features in the input data through layer-by-layer filtering, learn the unique power consumption mode of each load, and the calculation formulas of the convolution layer and the pooling layer are as follows: wherein is a matrix that stabilizes the training gradient, is a normalization function; At the fully connected layer, the input is multiplied by a weight matrix plus a bias term and then mapped through an activation function to an output value that integrates the features extracted by the preceding network: Data denormalization, the processed in the model The interval data is restored to the original data range to obtain the final output, which is the active power decomposition sequence of the load.

2. The deep learning based non-intrusive load decomposition method according to claim 1, wherein, Obtain the output: 3.The deep learning based non-intrusive load decomposition method of claim 1, wherein, The preset neural network uses a convolutional neural network with an introduced attention mechanism, a deep learning neural network of a bidirectional long short-term memory network, establishes a unique mapping relationship between total load and each target load, and defines the relationship between the total power load power sequence of a certain family at a certain time and the power sequence of each single load. total power load power sequence power sequence of each single load The relationship between them is: wherein is a state variable consisting of 0-1 variables, 0 indicating that the device is cut off, 1 indicating that the device is in operation, denotes that there is a load, the function denotes a NILM mapping method. 4.The deep learning based non-intrusive load decomposition method of claim 1, wherein, The total table data and the sub-table data respectively preprocess the original total power sequence and the load power sequence, fill in the missing values, and unify the sampling frequencies of the sub-table at each load and the total table at the entrance. The noise is divided into random noise and non-random noise. It is considered that the total active power data collected by the total power meter when there is no load input is non-random noise, and the time length is set The detected single occurrence time lower than The labeled event is determined as random noise caused by the burr pulse at the sampling time, and the formula for separating the remaining power is: wherein Ptotal is the total power, Pn represents the nth sample period, Pn represents the nth sample period has loads in an on state and consumes power, Pnoise is the noise.

5. A deep learning based non-intrusive load disaggregation system employing the deep learning based non-intrusive load disaggregation method of claim 1, characterized in that, When the finite state type load is separated from the total power signal to obtain a residual power signal: It comprises: A data processing module is configured to obtain total table data and sub-table data, fill in missing values of the sub-table data based on the total table data, and unify the sampling frequencies of the power data of the total table and the sub-table, wherein the total table data is responsible for sampling total power, and the sub-table data is responsible for sampling each load power; A load detection module is configured to detect the finite state type load event based on the sliding window and step judgment method, and refine the total power to the finite state type load based on a preset neural network; A load decomposition module is configured to separate the finite state type load from the total power signal to obtain a residual power signal, learn the operation mode of the power continuous change type load in the residual power based on a preset neural network, and realize the decomposition of power consumption.

6. A terminal device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program capable of running on the processor, and the processor loads and executes the computer program, and adopts the deep learning-based non-intrusive load decomposition method in any one of claims 1 to 4.

7. A computer-readable storage medium having stored therein a computer program, characterized in that, The computer program is loaded and executed by the processor, and adopts the deep learning-based non-intrusive load decomposition method in any one of claims 1 to 4.

Images