Non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM

Through the Fryze-current decomposition and CNN-BiLSTM method, the problem of low accuracy of electrical appliance categories in non-invasive load monitoring is solved, and high-accuracy electrical appliance recognition is achieved, which is suitable for automatic feature learning of complex load data and electrical appliances. Status monitoring.

CN118940094APending Publication Date: 2024-11-12GUANGZHOU POWER SUPPLY BUREAU GUANGDONG POWER GRID CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411154094.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In the field of non-invasive load monitoring (NILM), load recognition accuracy is low when there are many electrical categories. Existing deep learning methods are difficult to automatically learn the most characterization features in complex load data, resulting in a high recognition error rate.

Method used

Using a non-invasive household appliance recognition method based on Fryze-current decomposition and CNN-BiLSTM, the CNN-BiLSTM neural network is trained by obtaining sample data sets, and the current signal is decomposed using Fryze power theory, combined with Euclidean distance function and image conversion, input The feature vector is extracted from the convolutional neural network, and the probability vector is calculated through the bidirectional long and short-term memory neural network to realize electrical identification.

Benefits of technology

The load identification accuracy rate when there are many electrical appliance categories in the field of non-invasive load monitoring has been improved, and effective monitoring of the operating status of indoor electrical equipment has been achieved. The average identification accuracy rate is 97.8%, meeting actual production needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940094A_ABST
    Figure CN118940094A_ABST
Patent Text Reader

Abstract

The invention discloses a non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM, and relates to the technical field of intelligent power grids, and the method comprises the steps: carrying out the processing of collected total current and voltage waveforms, and obtaining to-be-identified load data; decomposing the activation current i into an orthogonal component related to electric energy in a time domain by using a method based on a Fryze power theory; reducing the dimensions of the decomposed signal ia and the decomposed signal if from the original dimension of the sampling period Ts to a predefined dimension w; obtaining a distance matrix according to the active current i (t) a and the inactive component i (t) f; performing image conversion on the distance matrix, and inputting an obtained image into a convolutional neural network in a CNN-BiLSTM neural network to obtain a one-dimensional feature vector of the electric equipment; and calculating the one-dimensional feature vector of the electric equipment through a bidirectional long-short term memory neural network in the CNN-BiLSTM neural network to obtain a probability vector of the to-be-identified household appliance. According to the invention, the problem that the load identification accuracy is low when there are many types of electric appliances in the non-intrusive load monitoring field can be solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of smart grid technology, and in particular to a non-invasive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM. Background Art

[0002] According to the United Nations Environment Programme Sustainable Buildings and Climate Initiative (UNEP-SBCI), residential and commercial buildings consume about 60% of the world's electricity. Energy conservation in buildings can generally be improved by upgrading more efficient equipment to improve the energy efficiency of buildings (for example, using better insulation materials), or by changing user energy behavior to optimize electricity use. In recent years, advances in energy sensors and smart meter technology have led to the rise of non-intrusive load monitoring (NILM). Non-intrusive load monitoring (NILM) is a computing technology that uses total power data such as smart meters, current or voltage sensor plugs to infer the terminal devices running in the building and estimate their respective power consumption. Non-intrusive load monitoring (NILM) relies on signal processing and machine learning techniques to analyze appliance patterns from total power measurements. Non-intrusive load monitoring (NILM) provides a low-cost and convenient way to monitor household load energy consumption, which can be easily integrated into existing buildings without any inconvenience to users. The strategy can also provide electricity retailers with the power consumption of terminal appliances, thereby gaining insight into what appliances customers use, when they use them, how much electricity they consume, and why they consume it in this way. Identifying and decomposing the power consumption information of terminal appliances through non-intrusive load monitoring (NILM) technology helps public utility managers evaluate the energy demand of residential loads, thereby effectively carrying out public facility construction and demand-side response businesses, which plays an important role and significance for energy conservation, environmental protection and low-carbon transformation of society.

[0003] Non-intrusive load identification is the core task of non-intrusive load monitoring (NILM). Currently, the commonly used algorithms include supervised and unsupervised. Among them, supervised algorithms mainly realize load identification by training labeled data sets, such as decision tree-based methods, support vector machine-based methods, and neural network-based methods. Although such methods have improved the accuracy of load identification to a certain extent, they are often prone to misidentification when there are many types of loads. Unsupervised algorithms are mainly used in non-intrusive load monitoring (NILM) scenarios where it is difficult to obtain load labels. Load identification is achieved by automatically learning the structural information hidden in the data. Currently, commonly used methods include random forest method, K-means method, and DBSCAN density clustering method, which eliminate the tediousness of manual labeling, but the accuracy of load identification is generally low.

[0004] With the development of a new generation of artificial intelligence technology represented by deep learning, more and more scholars have tried to apply deep learning technology to load identification in the field of non-intrusive load monitoring (NILM). For example, some researchers have proposed using the switching transient waveform of current as a feature. The non-intrusive load identification method based on convolutional neural network has shown better classification effect than the traditional BP (backpropagation) neural network; some researchers have also proposed using handwritten data sets to train convolutional neural networks, and then transferring the trained convolutional neural networks to the field of household appliance load identification, and then identifying electrical loads. Although the existing load identification methods based on deep learning have improved the accuracy of load identification to a certain extent, in scenarios with more types of power loads, they are prone to identification errors due to the limited model learning ability. Therefore, it is an urgent problem to carry out research on new deep learning methods or combined methods to enhance the ability of algorithms to automatically learn the most representative features from complex load data, thereby improving the accuracy of load identification. Summary of the invention

[0005] The purpose of this application is to provide a non-invasive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM, which can solve the problem of low load identification accuracy when there are many types of electrical appliances in the field of non-invasive load monitoring (NILM).

[0006] To achieve the above-mentioned purpose, the present application provides a non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM, comprising the following steps: S1: obtaining a sample data set to train a neural network to obtain a CNN-BiLSTM neural network; wherein the sample data set includes: current and voltage signals of multiple types of electrical equipment; the CNN-BiLSTM neural network includes: a convolutional neural network and a bidirectional long short-term memory neural network; S2: collecting the total current and voltage waveforms, processing the total current and voltage waveforms, and obtaining load data to be identified, wherein the load data to be identified includes: an activation current i and an activation voltage v; S3: using a method based on Fryze power theory to decompose the activation current i in the time domain into orthogonal components related to electric energy, wherein the orthogonal components include: active current i(t) a and non-active component i(t) f ; S4: Use the segmented aggregation approximation method to decompose the decomposed signal i a and signal i f The dimension of the sampling period T s The original dimension is reduced to a predefined dimension w, where signal i a is the sampling period T s Internal active current i(t) a The collection of signal i fis the sampling period T s Internal non-active component i(t) f S5: Apply the Euclidean distance function to measure the active current i(t) a and non-active component i(t) f The similarity or correlation between them is calculated to obtain a distance matrix; S6: image conversion is performed on the distance matrix to obtain an image, and the image is input into the convolutional neural network in the CNN-BiLSTM neural network, and the convolutional neural network extracts the image to obtain a one-dimensional feature vector of the electrical equipment; S7: the one-dimensional feature vector of the electrical equipment is calculated by the bidirectional long short-term memory neural network in the CNN-BiLSTM neural network to obtain a probability vector of the home appliance to be identified.

[0007] As above, the measurement target is measured by a non-invasive measurement device to obtain the total current and voltage waveforms, and the total current and voltage waveforms are processed using a high-frequency feature extraction method, and the single-cycle steady-state signals extracted from the total current and voltage waveforms are used as activation current i and activation voltage v, and the activation current i and activation voltage v are used as load data to be identified.

[0008] As above, the activation current i is decomposed in the time domain as follows: i(t)=i(t) a +i(t) f ; Where i(t) is the activation current at sampling time t; i(t) a is the active current; i(t) f It is a non-active component.

[0009] As above, where the active current i(t) a The expression is: Among them, the current i a (t) represents the resistance information, which is a pure sine curve; v(t) is the activation voltage at the sampling time t; v rms is the effective value of voltage; Pa is the average power.

[0010] As above, where the non-active component i(t) f The expression is: i(t) f =i(t)-i(t) a ; Where i(t) is the activation current at sampling time t; i(t) a is the active current.

[0011] As above, the distance matrix expression is: Among them, D w,w is the distance matrix; d 2,1 ,d w,1 ,d1,2 ,d w,2 and d 2,w are all element values ​​in the distance matrix; d 2,1 =||i(t)2-i(t)1||2; d w,1 =||i(t) w -i(t)1||2;d 1,2 =||i(t)1-i(t)2||2; d w,2 =||i(t) w -i(t)2||2;d 1,w =||i(t)1-i(t) w ||2;d 2,w =||i(t)2-i(t) w ||2, i(t)1 is the value of the first dimension, i(t)2 is the value of the second dimension, i(t) w is the value of the w-th dimension, and w represents the dimension after dimensionality reduction.

[0012] As mentioned above, the convolutional neural network in the CNN-BiLSTM neural network includes: an input layer, a hidden layer and an output layer; the image is input to the input layer, and after feature conversion and extraction by the hidden layer, it is output to the output layer, and the output layer outputs a one-dimensional feature vector of the electrical equipment.

[0013] As above, the expression of the bidirectional long short-term memory neural network in the CNN-BiLSTM neural network is: Where: x t represents the input information of the cell at time t; h t is the output information of the cell at time t; h t-1 is the output information of the cell at time t-1; r t is the output of the reverse layer at time t; r t+1 is the output of the reverse layer at time t+1; y t is the output of the output layer at time t; f(·) represents the activation function of the forward layer and the reverse layer; r=g(·) is the activation function of the output layer; W1 is the weight matrix mapping the input layer to the forward layer; W2 is the weight matrix mapping the output of the forward layer at the previous calculation moment to the current calculation moment; W3 is the weight matrix mapping the input layer to the reverse layer; W4 is the weight matrix mapping the output of the forward layer to the output layer; W5 is the weight matrix mapping the output of the reverse layer at the previous calculation moment to the current calculation moment; W6 is the weight matrix mapping the output of the reverse layer to the output layer.

[0014] As above, if the device to be identified changes, the changed device to be identified is used as the current device to be identified, sample data of the current device to be identified is obtained, step S1 is executed, a new CNN-BiLSTM neural network is obtained, and steps S2-S7 are executed using the new CNN-BiLSTM neural network; wherein the changes include: adding new electrical equipment, reducing electrical equipment and / or replacing electrical equipment

[0015] As above, if the device to be identified does not change, the CNN-BiLSTM neural network obtained in step S1 is directly used to execute steps S2 to S7; wherein the changes include: adding new electrical equipment, reducing electrical equipment and / or replacing electrical equipment.

[0016] The beneficial effects achieved by this application are as follows:

[0017] (1) The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM of the present application can solve the problem of low load identification accuracy when there are many types of electrical appliances in the field of non-intrusive load monitoring (NILM), that is, it improves the load identification accuracy when there are many types of electrical appliances in the field of non-intrusive load monitoring (NILM).

[0018] (2) The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM of the present application can monitor the operating status of indoor electrical equipment according to the total current and voltage measurement values ​​entering the house. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0020] Figure 1 The present invention is a flowchart of an embodiment of a non-invasive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM. DETAILED DESCRIPTION

[0021] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 those skilled in the art without creative work are within the scope of protection of the present invention.

[0022] like Figure 1As shown, the present application provides a non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM, comprising the following steps:

[0023] S1: Obtain a sample data set to train a neural network and obtain a CNN-BiLSTM neural network; wherein the sample data set includes: current and voltage signals of multiple types of electrical equipment; the CNN-BiLSTM neural network includes: a convolutional neural network and a bidirectional long short-term memory neural network.

[0024] Specifically, the neural network is trained by collecting and acquiring the historical current and voltage data of the equipment to be identified, thereby obtaining and saving the CNN-BiLSTM neural network, which provides model support for subsequent electrical load identification. CNN-BiLSTM neural network: used to process the load data to be identified and obtain the current and voltage signals.

[0025] Furthermore, the sample data set is a public data set PLAID, but is not limited to a public data set PLAID. The present application preferably uses a public data set PLAID.

[0026] Specifically, the dataset PLAID is a public dataset commonly used in the field of load identification. The public dataset PLAID is obtained by sampling at 30Hz. The public dataset PLAID contains measurement data of multiple types of electrical appliances such as air conditioners, blenders, coffee machines, energy-saving lamps, fans, refrigerators, hair dryers, curling irons, heaters, incandescent lamps, notebooks, microwave ovens, electric soldering irons, vacuum cleaners, washing machines, and electric kettles. The measurement data includes current and voltage signals.

[0027] S2: Collecting total current and voltage waveforms, processing the total current and voltage waveforms, and obtaining load data to be identified, wherein the load data to be identified includes: an activation current i and an activation voltage v.

[0028] Furthermore, the measurement target is measured by a non-invasive measurement device to obtain the total current and voltage waveforms, and the total current and voltage waveforms are processed using a high-frequency feature extraction method. The single-cycle steady-state signals extracted from the total current and voltage waveforms are used as activation current i and activation voltage v, and the activation current i and activation voltage v are used as load data to be identified.

[0029] Specifically, the measurement target is a residential building, but not limited to a residential building, and the present application is preferably a residential building. The non-intrusive measurement device is a smart meter, but not limited to a smart meter, and the present application is preferably a smart meter.

[0030] Specifically, considering the characteristics of the power consumption device extracted from the high-frequency accumulated voltage and current measurement signals in a short time window, the characteristics can distinguish multiple power consumption devices operating simultaneously. The activation current i and activation voltage v are defined as single-cycle steady-state signals extracted from the total current and voltage waveforms. In order to obtain the activation component waveform from the aggregated measurement signal, N is measured after the event occurs. c = 20 cycles of current and voltage {i (a) ,v (a)}, where N c is the period, i (a) N c The activation current dataset element for each cycle, v (a) N c The activation voltage dataset element for each cycle. N c The period corresponds to the steady state, which is equivalent to T s ×N s samples, among which T s is the sampling period, f s is the sampling frequency, f is the main frequency; N s is the number of sampling cycles. Then extract a cycle T s The activation current i and the activation voltage v.

[0031] Specifically, in the field of non-intrusive load monitoring, a method based on Fryze power theory extracts sub-waveforms from the total current and voltage waveforms.

[0032] S3: Using a method based on Fryze power theory, the activation current i is decomposed into orthogonal components related to electric energy in the time domain, where the orthogonal components include: active current i(t) a and non-active component i(t) f .

[0033] Specifically, the decomposed current characteristics are obtained by applying the Fryze power theory. The method based on the Fryze power theory is a prior art, but applying the method based on the Fryze power theory to the field of non-intrusive load monitoring is one of the improvements of this application.

[0034] Furthermore, the activation current i is decomposed in the time domain as follows:

[0035] i(t)=i(t) a +i(t) f ;

[0036] Where i(t) is the activation current at sampling time t; i(t) a is the active current; i(t) f It is a non-active component.

[0037] Furthermore, due to the active current i(t) a is the current of a resistive load with the same active power at the same activation voltage v.

[0038] In Fryze's theory, the active power calculation i(t)·v(t) is performed in a basic period T s The average value on:

[0039]

[0040] Where Pa is the average power; T s Sampling period; t is the sampling time; i(t) is the activation current at the sampling time t; v(t) is the activation voltage at the sampling time t.

[0041] Therefore, the active current i(t) a The expression is:

[0042]

[0043] Among them, the current i a (t) represents the resistance information, which is a pure sine curve; v(t) is the activation voltage at the sampling time t; v rms is the effective value of voltage; Pa is the average power.

[0044] Furthermore, the effective value of the voltage v rms The expression is:

[0045]

[0046] Among them, T s is the sampling period; t is the sampling time; v(t) is the activation voltage at the sampling time t. Furthermore, the non-active component i(t) f The expression is:

[0047] i(t) f =i(t)-i(t) a ;

[0048] Where i(t) is the activation current at sampling time t; i(t) a is the active current.

[0049] S4: Use the piecewise aggregation approximation method (PAA) to decompose the decomposed signal i a and signal i f The dimension of the sampling period T s The original dimension is reduced to a predefined dimension w, where signal i a is the sampling period T s Internal active current i(t)a The collection of signal i f is the sampling period T s Internal non-active component i(t) f A collection of .

[0050] Specifically, the segmented aggregation approximation method (PAA) is used to transform the decomposed signal i a and signal i f The dimension of the sampling period T s The specific method of reducing the dimension to the predefined dimension w can be implemented by using the existing technology, so it will not be described in detail.

[0051] S5: Apply the Euclidean distance function to measure the active current i(t) a and non-active component i(t) f The similarity or correlation between them is calculated to obtain the distance matrix.

[0052] Specifically, in order to further enhance the uniqueness of the decomposed current feature, the Euclidean distance function is applied to the active current and the non-active component to measure the similarity or correlation between the two values ​​of the active current and the non-active component. w The distance matrix D w,w It is a square matrix of Euclidean distances, representing the distance between sets of points of dimension w in Euclidean space. Where i(t)1 is the value of the first dimension, i(t)2 is the value of the second dimension, and i(t) w is the value of the w-th dimension, and w represents the dimension after dimensionality reduction.

[0053] Furthermore, the expression of the Euclidean distance function is:

[0054] d u,v =||i(t) u -i(t) v ||2;

[0055] Among them, d u,v is the Euclidean distance function; i(t) u is the u-th value; i(t) v is the vth value.

[0056] Furthermore, the distance matrix is ​​expressed as:

[0057]

[0058] Among them, D w,w is the distance matrix; d 2,1 ,d w,1 ,d 1,2 ,d w,2 and d2,w are all element values ​​in the distance matrix; d 2,1 =||i(t)2-i(t)1||2; d w,1 =||i(t) w -i(t)1||2;d 1,2 =||i(t)1-i(t)2||2; d w,2 =||i(t) w -i(t)2||2;d 1,w =||i(t)1-i(t) w ||2;d 2,w =||i(t)2-i(t) w ||2.

[0059] Specifically, the distance matrix D w,w It is formed by calculating the distance between two values ​​of dimension w.

[0060] S6: Perform image conversion on the distance matrix to obtain an image, and input the image into the convolutional neural network (CNN) in the CNN-BiLSTM neural network. The convolutional neural network extracts the image to obtain a one-dimensional feature vector of the electrical equipment.

[0061] Specifically, the essence of an image is a matrix. The image can be obtained by normalizing the element values ​​of the distance matrix to a set numerical range. The distance matrix is ​​converted into an image representation, and then the image is input into a pre-trained convolutional neural network (CNN) to extract the one-dimensional feature vector of the electrical device. The one-dimensional feature vector of the electrical device is the feature vector obtained after the output feature map of the convolutional neural network is vectorized.

[0062] Among them, the convolutional neural network in the CNN-BiLSTM neural network includes: an input layer, a hidden layer and an output layer; the image is input to the input layer, and after feature conversion and extraction by the hidden layer, it is output to the output layer, and the output layer outputs the one-dimensional feature vector of the electrical equipment.

[0063] Among them, the hidden layers include: convolutional layer, pooling layer and fully connected layer.

[0064] Among them, the convolution layer outputs several feature maps, and the expression of the convolution layer is:

[0065]

[0066] in, is the jth feature image of the nth layer in the convolutional neural network, is the jth feature image of the n-1th layer in the convolutional neural network; represents the convolution kernel function between the jth feature map of the nth layer and the n-1th layer, i is the i-th feature map of the n-1th layer; M i is the total number of feature maps in the n-1th layer; i∈M i Indicates that i is M i Elements in is the j-th bias term of the n-th layer; b represents the j-th weight of the n-th layer; f(·) is the activation function, and the function of the activation function f(·) is to convert the convolution result into a nonlinear output.

[0067] The pooling layer outputs the reduced-dimensional feature map, and the expression of the pooling layer is:

[0068]

[0069] in, is the jth feature image of the nth layer in the convolutional neural network; is the jth feature image of the n-1th layer in the convolutional neural network; represents the j-th bias number of the n-th layer, and β is the bias number; represents the jth weight of the nth layer, and α is the weight. pool() represents the pooling function, which includes: maximum pooling function or mean pooling function. f(·) is the activation function, which converts the convolution result into a nonlinear output.

[0070] After being processed by the convolutional layer and the pooling layer, the output data (i.e., the feature map with lower dimension and vectorization) is passed to the fully connected layer to map the output data to the classification label space. The output of the fully connected layer is converted into a classification label using the softmax function to obtain a one-dimensional feature vector of the electrical equipment, which is then output by the output layer.

[0071] Specifically, the softmax function is a well-known activation function. After the calculation of the convolution layer and the pooling layer, the feature map of the input data is obtained, and the feature map of the input data is vectorized to obtain a one-dimensional feature vector of the electrical device, and the one-dimensional feature vector of the electrical device is used as the input of the subsequent BiLSTM.

[0072] S7: The one-dimensional feature vector of the electrical device is calculated by the bidirectional long short-term memory (BiLSTM) neural network in the CNN-BiLSTM neural network to obtain the probability vector of the home appliance to be identified.

[0073] Specifically, BiLSTM consists of two LSTM (Long Short-Term Memory Artificial Neural Network) layers, one is a positive layer and the other is a negative layer; the positive layer is used for forward propagation, and the negative layer is used for reverse propagation. The forward layer iterates the input from the beginning of the sequence, while the reverse layer iterates the input from the end of the sequence. By fitting the output of the forward layer with the output of the negative layer, the recognition result can be obtained.

[0074] The long short-term memory artificial neural network (LSTM) is composed of a series of identical cells, each of which consists of a forget gate, an input gate, and an output gate. The forget gate is used to control the retention and forgetting of memory in the cell. The specific mathematical expression is:

[0075] f t =sigmoid(W f ·[h t-1 ,x t ]+b f );

[0076] In the formula, f t is the output of the forget gate; sigmoid() is the activation function; W f is the weight of the forget gate; b f is the bias of the forget gate; x t is the input information of the cell at time t; h t is the output information of the cell at time t; h t-1 is the output information of the cell at time t-1.

[0077] Among them, the output of the input gate is expressed as:

[0078] i t =σ(w i ·[h t-1 ,x t ]+b i );

[0079]

[0080] Where: i t is a control signal between 0 and 1 generated by the input gate; is the candidate cell image at the current moment generated by the tanh layer; σ() is the activation function, usually the sigmoid function; tanh is the activation function; W i and W C Both represent weights; b i and b C Both represent bias; C t is the cell state at time t; h t-1is the output information of the cell at time t-1; x t is the input information of the cell at time t.

[0081] Specifically, W i and W C Usually a randomly generated small random number that follows a normal distribution, and the values ​​are generally different; b i and b C It is usually a randomly generated small random number that follows a normal distribution, and the values ​​are generally different.

[0082] Updated cell state C t The expression is:

[0083]

[0084] Among them, f t is the output of the forget gate; C t-1 is the cell state at time t-1; i t is a control signal between 0 and 1 generated by the input gate; is the candidate cell image at the current moment generated by the tanh layer.

[0085] Among them, the output gate is expressed as:

[0086]

[0087] Where: O t Represents the intermediate information of the output gate; σ() is the activation function, usually the sigmoid function; h t is the output information of the cell at time t; W O is the weight; b O is bias; C t is the cell state at time t; tanh is the activation function; h t-1 is the output information of the cell at time t-1; x t is the input information of the cell at time t.

[0088] Furthermore, the expression of the bidirectional long short-term memory (BiLSTM) neural network in the CNN-BiLSTM neural network is:

[0089]

[0090] Where: x t represents the input information of the cell at time t; h t is the output information of the cell at time t; h t-1 is the output information of the cell at time t-1; rt is the output of the reverse layer at time t; r t+1 is the output of the reverse layer at time t+1; y t is the output of the output layer at time t; f(·) represents the activation function of the forward layer and the reverse layer; r=g(·) is the activation function of the output layer; W1 is the weight matrix mapping the input layer to the forward layer; W2 is the weight matrix mapping the output of the forward layer at the previous calculation moment to the current calculation moment; W3 is the weight matrix mapping the input layer to the reverse layer; W4 is the weight matrix mapping the output of the forward layer to the output layer; W5 is the weight matrix mapping the output of the reverse layer at the previous calculation moment to the current calculation moment; W6 is the weight matrix mapping the output of the reverse layer to the output layer.

[0091] Specifically, after the load data to be identified is processed through steps S2 to S6, a one-dimensional feature vector of the electrical equipment is obtained, which is then input into the trained neural network in step S7 to obtain a probability vector P of the household appliance to be identified = {p1, p2, ..., p n}, where P is the number of home appliances, p1, p2, p n are all elements of the probability vector of the household appliance to be identified; p1 is the electrical appliance 1; p2 is the electrical appliance 2; p n is the electrical device n; the element is the probability of identifying the corresponding possible household appliance, among which the largest corresponding element is the electrical device with the highest probability.

[0092] This application is mainly to solve the problem of low accuracy of load identification when there are many types of electrical appliances in the field of non-intrusive load monitoring (NILM). As an example, the test results of the non-intrusive home appliance identification method based on Fryze-current decomposition and CNN-BiLSTM in this application on the public dataset PLAID are shown in Table 1.

[0093] Electrical appliances Recognition rate air conditioner 97.15% Blender 99.73% Coffee machine 99.65% Energy-saving lamps 97.64% fan 97.89% refrigerator 98.23% hair dryer 96.46% Heater 98.65% notebook 96.41% Micro-wave oven 97.24% Vacuum cleaner 96.75% washing machine 96.67% Electric kettle 98.99%

[0094] Table 1 shows the typical household appliances recognition results of the PLAID dataset

[0095] Specifically, it can be seen from Table 1 that when there are many categories of household appliances, the non-invasive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM of the present application can still achieve relatively ideal recognition effects in terms of recognition accuracy of various categories of appliances, with an average recognition accuracy of 97.8%, which can meet actual production needs.

[0096] The content of non-intrusive load monitoring (NILM) is described. Specifically, the goal of non-intrusive load monitoring (NILM) is to aggregate (total) measurements x from multiple devices of smart meters. t Identify the working status of multiple electrical devices Where m = 1, 2, ... M, representing M different electrical devices, such that:

[0097]

[0098] Among them, x t is the input information of the cell at time t; t It represents both the impact of unconsidered electrical equipment and the measurement noise of the meter; is the power of the mth electrical device at time t; is the working status of the mth electrical equipment at time t, m∈[1,M], and M is the total number of energy-consuming equipment.

[0099] use represents a set of input features derived from the total measured values ​​of M terminal electrical devices, and is represented by Represents the measured value of the relevant electrical equipment. If each electrical equipment has k states, it is recorded as Make in, The mth electrical device is in state 1; The mth power consumer is in state k. Matrix represents the associated multi-label state. Therefore, given D = {x t ,s t |t=1,...,T} dataset, where s t is the state set of electrical equipment at time t; T is the sampling time; the goal is to learn a multi-label classifier to extract the aggregate power feature vector x from the input t The predicted state vector in, is the state of the electrical equipment M at time t.

[0100] Furthermore, as an embodiment, if the device to be identified changes, the changed device to be identified is used as the current device to be identified, sample data of the current device to be identified is obtained, step S1 is executed, a new CNN-BiLSTM neural network is obtained, and steps S2 to S7 are executed using the new CNN-BiLSTM neural network; wherein the changes include: adding new electrical equipment, reducing electrical equipment, and replacing electrical equipment.

[0101] Furthermore, as another embodiment, if the device to be identified does not change, the CNN-BiLSTM neural network obtained in step S1 is directly used to execute steps S2 to S7; wherein the changes include: adding new electrical equipment, reducing electrical equipment, and replacing electrical equipment.

[0102] The beneficial effects achieved by this application are as follows:

[0103] (1) The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM of the present application can solve the problem of low load identification accuracy when there are many types of electrical appliances in the field of non-intrusive load monitoring (NILM), that is, it improves the load identification accuracy when there are many types of electrical appliances in the field of non-intrusive load monitoring (NILM).

[0104] (2) The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM of the present application can monitor the operating status of indoor electrical equipment according to the total current and voltage measurement values ​​entering the house.

[0105] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic creative concepts. Therefore, the scope of protection of the present application is intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application. Obviously, those skilled in the art may make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the protection scope of the present application and its equivalent technology, the present application is also intended to include these modifications and variations.

Claims

1. A non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM, characterized in that: The steps include: S1: Obtain a sample data set to train a neural network and obtain a CNN-BiLSTM neural network; wherein the sample data set includes: current and voltage signals of multiple types of electrical equipment; the CNN-BiLSTM neural network includes: a convolutional neural network and a bidirectional long short-term memory neural network; S2: Collecting total current and voltage waveforms, processing the total current and voltage waveforms, and obtaining load data to be identified, wherein the load data to be identified includes: activation current i and activation voltage v; S3: Using a method based on Fryze power theory, the activation current i is decomposed into orthogonal components related to electric energy in the time domain, where the orthogonal components include: active current i(t) a and non-active component i(t) f ; S4: Use the segmented aggregation approximation method to decompose the decomposed signal i a and signal i f The dimension of the sampling period T s The original dimension is reduced to a predefined dimension w, where signal i a is the sampling period T s Internal active current i(t) a The collection of signal i f is the sampling period T s Internal non-active component i(t) f A collection of; S5: Apply the Euclidean distance function to measure the active current i(t) a and non-active component i(t) f The similarity or correlation between them is calculated to obtain the distance matrix; S6: Perform image conversion on the distance matrix to obtain an image, and input the image into the convolutional neural network in the CNN-BiLSTM neural network, and the convolutional neural network extracts the image to obtain a one-dimensional feature vector of the electrical equipment; S7: The one-dimensional feature vector of the electrical equipment is calculated through the bidirectional long short-term memory neural network in the CNN-BiLSTM neural network to obtain the probability vector of the home appliance to be identified.

2. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: The measurement target is measured by a non-invasive measurement device to obtain the total current and voltage waveforms, and the total current and voltage waveforms are processed using a high-frequency feature extraction method. The single-cycle steady-state signals extracted from the total current and voltage waveforms are used as the activation current i and the activation voltage v, and the activation current i and the activation voltage v are used as the load data to be identified.

3. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: The activation current i is decomposed in the time domain as follows: i(t)=i(t) a +i(t) f ; Where i(t) is the activation current at sampling time t; i(t) a is the active current; i(t) f It is a non-active component.

4. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: Active current i(t) a The expression is: Among them, the current i a (t) represents the resistance information, which is a pure sine curve; v(t) is the activation voltage at the sampling time t; v rms is the effective value of voltage; Pa is the average power.

5. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: Non-active component i(t) f The expression is: i(t) f =i(t)-i(t) a ; Where i(t) is the activation current at sampling time t; i(t) a is the active current.

6. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: The distance matrix expression is: Among them, D w,w is the distance matrix; d 2,1 ,d w,1 ,d 1,2 ,d w,2 and d 2,w are all element values ​​in the distance matrix; d 2,1 =||i(t)2-i(t)1||2; d w,1 =||i(t) w -i(t)1||2;d 1,2 =||i(t)1-i(t)2||2; d w,2 =||i(t) w -i(t)2||2;d 1,w =||i(t)1-i(t) w ||2;d 2,w =||i(t)2-i(t) w ||2, i(t)1 is the value of the first dimension, i(t)2 is the value of the second dimension, i(t) w is the value of the w-th dimension, and w represents the dimension after dimensionality reduction.

7. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: The convolutional neural network in the CNN-BiLSTM neural network includes: an input layer, a hidden layer and an output layer; the image is input to the input layer, and after feature conversion and extraction by the hidden layer, it is output to the output layer, and the output layer outputs the one-dimensional feature vector of the electrical equipment.

8. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: The expression of the bidirectional long short-term memory neural network in the CNN-BiLSTM neural network is: Where: x t represents the input information of the cell at time t; h t is the output information of the cell at time t; h t-1 is the output information of the cell at time t-1; r t is the output of the reverse layer at time t; r t+1 is the output of the reverse layer at time t+1; y t is the output of the output layer at time t; f(·) represents the activation function of the forward layer and the reverse layer; r = g(·) is the activation function of the output layer; W1 is the weight matrix that maps the input layer to the forward layer; W2 is the weight matrix that maps the output of the forward layer at the previous calculation time to the current calculation time; W3 is the weight matrix that maps the input layer to the reverse layer; W4 is the weight matrix that maps the output of the forward layer to the output layer; W5 is the weight matrix that maps the output of the reverse layer at the previous calculation time to the current calculation time; W6 is the weight matrix that maps the output of the reverse layer to the output layer.

9. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: If the device to be identified changes, the changed device to be identified is used as the current device to be identified, sample data of the current device to be identified is obtained, step S1 is executed, a new CNN-BiLSTM neural network is obtained, and steps S2 to S7 are executed using the new CNN-BiLSTM neural network; wherein the changes include: adding new electrical equipment, reducing electrical equipment and / or replacing electrical equipment.

10. The non-intrusive household appliance identification method based on Fryze-current decomposition and CNN-BiLSTM according to claim 1, characterized in that: If the device to be identified does not change, the CNN-BiLSTM neural network obtained in step S1 is directly used to execute steps S2 to S7; wherein the change includes: adding new electrical equipment, reducing electrical equipment and / or replacing electrical equipment.