Non-intrusive load decomposition method and system based on deep learning

Abstract

The invention discloses a non-intrusive load decomposition method and system based on deep learning, and relates to the technical field of power load management analysis, and the method comprises the following steps: obtaining general table data and sub-table data, filling the missing value of the sub-table data based on the general table data, and unifying the sampling frequency of the power data of the general 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; detecting a finite state type load event based on a sliding window and a step judgment method, and refining total power to finite state type load based on a preset neural network; and separating a finite state type load from the total power signal to obtain a residual power signal, and learning and separating an operation mode of a power continuous change type load in the residual power based on a preset neural network to realize power consumption decomposition.

Description

Technical Field

[0001] The present invention relates to the technical field of power load management and analysis, and specifically to a non-invasive load decomposition method and system based on deep learning. Background Art

[0002] With the popularization of Internet of Things technology, integration technology, and communication technology, the share of power continuously changing devices is increasing. Especially in recent years, with the development of clean energy and the transformation of energy structure, photovoltaics, as a typical representative of power continuously changing loads, are gradually entering households. This trend has brought new challenges to non-invasive load monitoring and identification. Decomposing finite state loads under continuous power superposition and modeling the characteristics of continuously changing power loads are of great significance to the optimization of distribution and utilization systems. Real-time monitoring of photovoltaic power generation can provide power companies with data support for optimizing energy structure and improving power quality. At the same time, it can also enable users to understand the photovoltaic power generation in their homes, and then optimize their own electricity consumption behavior to maximize the use of solar energy.

[0003] The patent with publication number CN118606670A is a load identification method and load monitoring device based on genetic algorithms. It uses genetic algorithms to perform crossover or mutation operations on each individual in the initial identification population of electrical appliances to obtain the results of power consumption behavior identification. It does not explicitly consider the normally open type of load with continuous power changes, and the model generalization ability is insufficient. The patent with publication number CN117996838B is a distributed photovoltaic identification device based on non-invasive load monitoring. It uses non-invasive load monitoring technology for distributed photovoltaic power generation systems, and uses low-complexity sliding windows and support vector machine technology to decompose distributed photovoltaic power generation from the main feeder of the load. It needs to process a large amount of electrical parameter and non-electrical parameter data, and has high requirements for data processing and storage. Summary of the invention

[0004] In order to address the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide a non-intrusive load decomposition method and system based on deep learning, which can separate finite state loads and power continuously changing loads and realize the refinement of the power consumption of each load, and intuitively compare the predicted power consumption of each load with the actual power consumption. The power consumption decomposition result obtained has a high accuracy rate, which can provide more comprehensive data support for optimizing energy management and energy conservation and emission reduction.

[0005] In a first aspect, the purpose of the present invention can be achieved by the following technical solution: a non-intrusive load decomposition method based on deep learning, the method comprising the following steps:

[0006] Obtain the total table data and the sub-table data, fill 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, wherein 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;

[0007] Finite state load events are detected based on the sliding window and step judgment method, and the total power is refined into finite state loads based on the preset neural network;

[0008] The finite state load is separated from the total power signal to obtain the residual power signal. Based on the preset neural network learning, the operating mode of the load with continuously changing power in the residual power is separated to realize the decomposition of power consumption.

[0009] In combination with the first aspect, in certain implementations of the first aspect, the method also 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 combination with the first aspect, in some implementations of the first aspect, the method further includes: a process of detecting a finite state load event based on the sliding window and step determination method:

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

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

[0013]

[0014] Where t is the sampling time, s represents the sth sampling period, α + is a pre-assigned threshold;

[0015] The fitting expression of the value in the sliding window before the input event occurs is calculated by 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 Pn :

[0020] P n = k × t n +b

[0021] For loads that have multiple state switching and intermittent operation in one switching event, set the delay t delay , will occur at intervals less than t delay Multiple marking events are merged into one load switching event;

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

[0023] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the preset neural network uses a convolutional neural network that introduces an attention mechanism Attention and a deep learning neural network of a bidirectional long short-term memory network to establish a unique mapping relationship between the total load and each target load, and defines 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 means the device is cut off, 1 means the device is in operation, N means there are N loads, and f function represents the NILM mapping method.

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

[0027] The load active power slice sequence is used as input, and the load actual active power is used as output. Before the one-dimensional data is input into the neural network, time window cutting is performed to cut the long time series into short segments;

[0028] Divide the data into training sets and test sets according to the proportion, learn the pattern, weight and bias through the training set, and verify the generalization ability of the model and system on the test set after the training is completed;

[0029] Before inputting the data into the model and system, normalize it, convert it to the same scale, and map the data to (-1,1):

[0030]

[0031] where x t and x t ' are the data before and after normalization, x minand x max are the minimum and maximum values ​​respectively;

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

[0033]

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

[0035] BiLSTM is set as the second layer of the neural network. Both local and global information are considered in the time series data. The output h is obtained by fitting the hidden layer calculation results of the forward iteration and the backward iteration. t :

[0036]

[0037] Among them, w t and v t is the weight vector, is the hidden layer state of the forward iteration, is the hidden layer state of the backward iteration, n t is the offset of the hidden layer state at that moment;

[0038] In the attention mechanism, different parts of the input data are given different degrees of attention, and the input sequence X is linearly mapped to obtain matrices Q, K and V that linearly map the input matrix. The formula is:

[0039] Q=W q X

[0040] K=W k X

[0041] V=W v X

[0042] The output is:

[0043]

[0044] Where D K is the matrix for stabilizing training gradients, and softmax(·) is the normalization function;

[0045] In the fully connected layer, the input x l-1 With the weight matrix wl Multiply, and add the bias term b l , and then mapped to the output value y through the activation function l , integrating the features extracted by the aforementioned network:

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

[0047] The data is denormalized 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 combination with the first aspect, in some implementations of the first aspect, the method further includes: when separating the finite state 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 put into operation, the total active power data collected by the total electric meter is non-random noise. The setting time t noise , the detected single occurrence time is less than t noise The marking event is determined to be the random noise caused by the glitch pulse during sampling, and the formula for separating the remaining power is:

[0050]

[0051] Where P j is the total power, j represents the jth sampling period, n represents that in the jth sampling period, n loads are turned on and consume power, and N is the noise.

[0052] In a second aspect, in order to achieve the above-mentioned object, the present invention discloses a non-intrusive load decomposition system based on deep learning, comprising:

[0053] A data processing module is used to obtain the total table data and the sub-table data, fill 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, wherein 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;

[0054] A load detection module is used to detect finite state load events based on a sliding window and step determination method, and to refine the total power into finite state loads based on a preset neural network;

[0055] The load decomposition module is used to separate the finite state load from the total power signal to obtain the residual power signal. Based on the preset neural network learning, the operating mode of the load with continuous power change in the residual power is separated to realize the decomposition of power consumption.

[0056] In another aspect of the present invention, in order to achieve the above-mentioned purpose, 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, wherein the memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, a non-invasive load decomposition method based on deep learning as described above is adopted.

[0057] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a computer-readable storage medium is disclosed, in which a computer program is stored. 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 adopted.

[0058] Beneficial effects of the present invention:

[0059] The present invention can detect finite state load switching events through real-time power mutation under the interference of continuous power superposition, taking into account the diversity of load characteristics, and can accurately locate the time when different load switching occurs while reducing the amount of data storage, providing more accurate input data for subsequent load decomposition models and systems. The present invention considers the actual composition of total power data, and in the case of using only total active power data as source data, with the help of the mapping ability of neural networks, to a certain extent eliminates the errors of event detection, the errors of other load decompositions and the interference of unknown noise, and realizes the accurate decomposition of finite state loads and power continuous change loads, greatly reducing the demand for data sampling. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0061] Figure 1 It is a schematic flow chart of the method of the present invention;

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

[0063] Figure 3 It is a schematic diagram of event detection prediction value replacement;

[0064] Figure 4 It is a finite state load decomposition step diagram;

[0065] Figure 5 It is the load active power decomposition effect diagram;

[0066] Figure 6 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0067] 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.

[0068] Embodiment 1:

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

[0070] S101: Obtaining total table data and sub-table data, filling missing values ​​of sub-table data based on the total table data, and unifying the sampling frequency of power data of the total table and sub-table, wherein the total table data is responsible for sampling total power, and the sub-table data is responsible for sampling power of each load;

[0071] The original total power sequence and each load power sequence are preprocessed in the total table data and the sub-table data, missing values ​​are filled, and the sampling frequency of the sub-table at each load is unified with the sampling frequency of the total table at the household;

[0072] S102: Detecting finite state load events based on a sliding window and step determination method, and refining total power into finite state loads based on a preset neural network;

[0073] The process of detecting finite state load events based on the sliding window and step judgment method:

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

[0075] The judgment formula for the occurrence of the finite state load event is:

[0076]

[0077] Where t is the sampling time, s represents the sth sampling period, α + is a pre-assigned threshold;

[0078] The fitting expression of the value in the sliding window before the input event occurs is calculated by 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 switching and intermittent operation in one switching event, set the delay t delay , will occur at intervals less than t delay Multiple marking events are merged into one load switching event;

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

[0086] The preset neural network uses a convolutional neural network that introduces the attention mechanism Attention and a deep learning neural network of a bidirectional long short-term memory network to establish a unique mapping relationship between the total load and each target load, and defines 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:

[0087]

[0088] Where a is a state variable consisting of 0-1 variables, 0 means the device is cut off, 1 means the device is in operation, N means there are N loads, and f function represents the NILM mapping method.

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

[0090] The load active power slice sequence is used as input, and the load actual active power is used as output. Before the one-dimensional data is input into the neural network, time window cutting is performed to cut the long time series into short segments;

[0091] Divide the data into training sets and test sets according to the proportion, learn the pattern, weight and bias through the training set, and verify the generalization ability of the model and system on the test set after the training is completed;

[0092] Before inputting the data into the model and system, normalize it, convert it to the same scale, and map the data to (-1,1):

[0093]

[0094] where x t and x t ' are the data before and after normalization, x min and x max are the minimum and maximum values ​​respectively;

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

[0096]

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

[0098] BiLSTM is set as the second layer of the neural network. Both local and global information are considered in the time series data. The output h is obtained by fitting the hidden layer calculation results of the forward iteration and the backward iteration. t :

[0099]

[0100] Among them, w t and v t is the weight vector, is the hidden layer state of the forward iteration, is the hidden layer state of the backward iteration, n t is the offset of the hidden layer state at that moment;

[0101] In the attention mechanism, different parts of the input data are given different degrees of attention, so that the network pays more attention to the time period when the event occurs. , linear mapping is performed on the input sequence X to obtain matrices Q, K and V that linearly map the input matrix. The formula is:

[0102] Q=W q X

[0103] K=W k X

[0104] V=W v X

[0105] The output is:

[0106]

[0107] Where D K is the matrix for stabilizing training gradients, and softmax(·) is the normalization function;

[0108] In the fully connected layer, the input x l-1 With the weight matrix w l Multiply, and add the bias term b l , and then mapped to the output value y through the activation function l , integrating the features extracted by the aforementioned network:

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

[0110] The data is denormalized 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 load from the total power signal to obtain a residual power signal, and separate the operation mode of the power continuously changing load from the residual power based on a preset neural network learning to achieve power consumption decomposition.

[0112] When the finite state load is separated from the total power signal to obtain the residual power signal:

[0113] Noise is divided into random noise and non-random noise. It is assumed that when no load is put into operation, the total active power data collected by the total electric meter is non-random noise. The setting time t noise , the detected single occurrence time is less than t noise The marking event is determined to be the random noise caused by the glitch pulse during sampling, and the formula for separating the remaining power is:

[0114]

[0115] Where P j is the total power, j represents the jth sampling period, n represents that in the jth sampling period, n loads are turned on and consume power, and N is the noise

[0116] Specifically, the present invention is further described below by way of embodiments:

[0117] In the LVNS simulation platform, the load of residential users is simulated for 24 hours to obtain the corresponding power consumption data as the application scenario. The method and system of the present invention are compared with the results of load decomposition using only the CNN and CNN-BiLSTM neural network frameworks, and the separated residual values are compared with the neural network prediction values of photovoltaic power. 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 Photovoltaic Power continuously changing load 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] Taking the evaluation metrics F1 value, mean square error (MSE), and coefficient of determination (R 2 ) in the field of machine learning as indicators to measure the accuracy of load decomposition. The comparison of the load decomposition results of the method of the present invention is shown in Table 2, and the decomposition prediction curve is as Figure 5 shown;

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

[0122]

[0123] From the parameter performance indicators and fitting trends of the decomposition results, it can be seen 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 influence of the superposition of continuously changing photovoltaic power and sampling noise, and achieve accurate decomposition from the total active power slice sequence to the true active power sequences of each load.

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

[0125] A data processing module 11, configured to obtain the total meter data and the sub-meter data, fill in the missing values of the sub-meter data based on the total meter data, and unify the sampling frequencies of the power data of the total meter and the sub-meter, where the total meter data is responsible for sampling the total power, and the sub-meter data is responsible for sampling the power of each load;

[0126] A load detection module 12, configured to detect finite state class load events based on the method of sliding window and step determination, and refine the total power to the finite state class load based on a preset neural network;

[0127] A load decomposition module 13, configured to separate the finite state class load from the total power signal to obtain a residual power signal, and learn the operating mode of the power continuously changing class load in the separated residual power based on a preset neural network to achieve the decomposition of power consumption.

[0128] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically used to load and execute one or more instructions in a computer storage medium to implement the above method.

[0129] It needs to be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by a processor to execute the above method. The storage medium can adopt 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, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0130] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0131] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure may have various changes and improvements, and these changes and improvements fall within the scope of the present disclosure to be protected.

Claims

1. A non-intrusive load decomposition method based on deep learning, characterized in that: The method comprises the following steps: Obtain the total table data and the sub-table data, fill 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, wherein 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; Finite state load events are detected based on the sliding window and step judgment method, and the total power is refined into finite state loads based on the preset neural network; The finite state load is separated from the total power signal to obtain the residual power signal. Based on the preset neural network learning, the operating mode of the load with continuously changing power in the residual power is separated to realize the decomposition of power consumption.

2. According to the non-intrusive load decomposition method based on deep learning in claim 1, it is characterized in that: The original total power sequence and each load power sequence are preprocessed in the total table data and the sub-table data, missing values ​​are filled, and the sampling frequency of the sub-table at each load is unified with the sampling frequency of the total table at the household entrance.

3. The non-intrusive load decomposition method based on deep learning according to claim 1, characterized in that: The process of detecting finite state load events based on the sliding window and step judgment method: The changing trend of the total power sequence is fitted within the sliding window, and a sudden increase is detected at the end. Once an input event is detected, the sampling value is replaced with a predicted value that conforms to the numerical trend of the sequence within the sliding window while marking the sampling period.

4. The non-intrusive load decomposition method based on deep learning according to claim 3, characterized in that: The judgment formula for the occurrence of the finite state load event is: Where t is the sampling time, s represents the sth sampling period, α + is a pre-assigned threshold; The fitting expression of the value in the sliding window before the input event occurs is calculated by the least squares method: 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; Replace the sampled value with the predicted value P n : P n =k×t n +b For loads that have multiple state switching and intermittent operation in one switching event, set the delay t delay , will occur at intervals less than t delay Multiple marking events are merged into one load switching event; Different threshold intervals are set according to the rated power of the equipment to filter out events of different equipment. If there are overlapping load events, multiple event detections are performed to separate events of different loads.

5. The non-intrusive load decomposition method based on deep learning according to claim 1, characterized in that: The preset neural network uses a convolutional neural network that introduces the attention mechanism Attention and a deep learning neural network of a bidirectional long short-term memory network to establish a unique mapping relationship between the total load and each target load, and defines 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: Where a is a state variable consisting of 0-1 variables, 0 means the device is cut off, 1 means the device is in operation, N means there are N loads, and f function represents the NILM mapping method.

6. The non-intrusive load decomposition method based on deep learning according to claim 1, characterized in that: The decomposition process of the finite state class load is as follows: The load active power slice sequence is used as input, and the load actual active power is used as output. Before the one-dimensional data is input into the neural network, time window cutting is performed to cut the long time series into short segments; Divide the data into training sets and test sets according to the proportion, learn the pattern, weight and bias through the training set, and verify the generalization ability of the model and system on the test set after the training is completed; Before inputting the data into the model and system, normalize it, convert it to the same scale, and map the data to (-1,1): where x t and x t ' are the data before and after normalization, x min and x max are the minimum and maximum values ​​respectively; Set CNN as the first layer of the neural network, extract local features from the input data through layer-by-layer filtering, and learn the unique power consumption pattern of each load. The calculation formulas for the convolution layer and the pooling layer are: Where x is the input data, in the convolutional layer formula it is the load power data after slicing the total power, W is the convolution kernel matrix, is the convolution operation, b is the convolution layer bias, is the activation function, α is the pooling layer weight, β is the pooling layer bias, and pool(·) is the pooling function; BiLSTM is set as the second layer of the neural network. Both local and global information are considered in the time series data. The output h is obtained by fitting the hidden layer calculation results of the forward iteration and the backward iteration. t : Among them, w t and v t is the weight vector, is the hidden layer state of the forward iteration, is the hidden layer state of the backward iteration, n t is the offset of the hidden layer state at that moment; In the attention mechanism, different parts of the input data are given different degrees of attention, and the input sequence X is linearly mapped to obtain matrices Q, K and V that linearly map the input matrix. The formula is: Q=W q X K=W k X V=W v X The output is: Where D K is the matrix for stabilizing training gradients, and softmax(·) is the normalization function; In the fully connected layer, the input x l-1 With the weight matrix w l Multiply, and add the bias term b l , and then mapped to the output value y through the activation function l , integrating the features extracted by the aforementioned network: y l =w l x l-1 +b l The data is denormalized 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.

7. The non-intrusive load decomposition method based on deep learning according to claim 1, characterized in that: When the finite state load is separated from the total power signal to obtain the residual power signal: Noise is divided into random noise and non-random noise. It is assumed that when no load is put into operation, the total active power data collected by the total electric meter is non-random noise. The setting time t noise , the detected single occurrence time is less than t noise The marking event is determined to be the random noise caused by the glitch pulse during sampling, and the formula for separating the remaining power is: Where P j is the total power, j represents the jth sampling period, n represents that in the jth sampling period, n loads are turned on and consume power, and N is the noise.

8. A non-intrusive load decomposition system based on deep learning, characterized in that: include: A data processing module is used to obtain the total table data and the sub-table data, fill 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, wherein 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; A load detection module is used to detect finite state load events based on a sliding window and step determination method, and to refine the total power into finite state loads based on a preset neural network; The load decomposition module is used to separate the finite state load from the total power signal to obtain the residual power signal. Based on the preset neural network learning, the operating mode of the load with continuous power change in the residual power is separated to realize the decomposition of power consumption.

9. 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 that can be run on the processor. When the processor loads and executes the computer program, a non-intrusive load decomposition method based on deep learning according to any one of claims 1 to 7 is adopted.

10. A computer-readable storage medium having a computer program stored therein, characterized in that: When the computer program is loaded and executed by the processor, a non-intrusive load decomposition method based on deep learning as described in any one of claims 1 to 7 is adopted.

Images