A non-invasive load switch status identification method for low-frequency signals

Abstract

The low-frequency signal non-intrusive load switch state identification algorithm disclosed by the present invention comprises the following steps: collecting training data of the current, active power, and reactive power of each single electrical appliance, clustering the current, active power, and reactive power to obtain the switch state class set of the single electrical appliance; extracting training features according to the total meter current, active power, and reactive power at a certain moment and the moment immediately preceding it in the training data; constructing a standardized training set according to the training features and the known switch state sequences of each electrical appliance; feeding the training set into a neural network model for training to obtain a switch state identification model; during operation, constructing working features according to the total meter current, active power, and reactive power at the current moment and the moment immediately preceding it; feeding the working features into the switch state identification model to obtain a combined switch state and decomposing the switch states of each single electrical appliance therefrom. The method of the present invention can accurately obtain the working states of each single electrical appliance, laying a foundation for subsequent further decomposition and identification.

Description

Technical Field

[0001] The present invention belongs to the field of non-intrusive load monitoring, and particularly relates to a method for identifying the state of a non-intrusive load switch for low-frequency signals. Background Art

[0002] In recent years, with the continuous development of smart grids, the lean and intelligent analysis of electrical signals has attracted increasing attention. For example, non-intrusive load monitoring (NILM) technology uses feature extraction and machine learning algorithms to analyze the combined voltage and current and monitor the usage of electrical appliances without the need to install sub-meters. Many feature extraction methods have been proposed for non-intrusive load monitoring, such as wavelet features, voltage-current trajectories, current harmonics, active / reactive power, and automatically learned deep features, etc. In the literature [Non-Intrusive Load Monitoring Using Semi-Supervised Machine Learning and Wavelet Design], signal features are extracted by using wavelet design and Procrustes analysis to match a new set of wavelets to the load type. However, in their method, the wavelets used according to different load categories need to be selected from a large number of wavelets, which limits its use in unknown electrical appliance scenarios to a certain extent. V-I trajectory-based features extract different features from the trajectory diagram of Voltage-Current (V-I) to describe the operating conditions of different electrical appliances. The proposed V-I trajectory-based features include loop direction, enclosed area, average curve non-linearity, number of self-intersections, middle segmentation inclination, left and right segmentation areas, etc. However, the influence of noise signals on the V-I trajectory is relatively large, which easily leads to mutations in various features of the V-I trajectory. Current harmonic features extract features such as modulus and modulus ratio from high-frequency current harmonics to distinguish electrical appliances. However, to obtain better results, it is generally required that the sampling frequency of the signal reaches 1 MHz and above, which brings a huge obstacle to the application of current harmonic features. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for identifying the state of a non-intrusive load switch for low-frequency signals, which can accurately obtain the switch states of each electrical appliance according to the data collected by the total meter, thereby laying a foundation for subsequent further decomposition and identification.

[0004] The technical solution adopted by the present invention to solve the above technical problems is: a method for identifying the state of a non-intrusive load switch for low-frequency signals, comprising the following steps:

[0005] 1) Collect the training data of the current, active power, and reactive power of each single electrical appliance, and cluster the current, active power, and reactive power of each single electrical appliance to obtain the switch state class set of the single electrical appliance;

[0006] 2) Extract training features based on the total meter current, active power, and reactive power at a certain moment and the previous moment in the training data;

[0007] 3) Construct a standardized training set according to the training features and the known switch state sequences of each individual electrical appliance;

[0008] 4) Construct a neural network model, send the training set into the neural network model for training, and obtain a switch state identification model;

[0009] 5) Construct working features according to the total meter current, active power, and reactive power at the current moment and the previous moment during operation;

[0010] 6) Send the working features into the trained switch state identification model to obtain the combined switch state, and decompose the switch states of each individual electrical appliance from the combined switch state.

[0011] Preferably, the specific process of step 1) is as follows:

[0012] 1.1) Suppose there are N electrical appliances, and the current, active power, and reactive power sequences in their training data are I k , P k , Q k , where k = 1, 2,... N. Use the meanshift clustering algorithm to cluster them into M classes, and it is required that M <= 10; when the number of clusters M obtained by clustering is > 10, gradually increase the bandwidth parameter of the meanshift algorithm by 5 each time until the number of classes M <= 10; record the class number of each individual electrical appliance at any moment t as m t ∈[1, M];

[0013] 1.2) Sort the cluster centers according to the current magnitude. When the class number of a certain individual electrical appliance is the smallest class, i.e., m t = 1, it is regarded that the electrical appliance is in the off state, i.e., the first state class; when the class number m t of the electrical appliance ∈[2, M], it is regarded that the electrical appliance is in the on state, i.e., the second state class;

[0014] Let the state classification number of the off state be 1, and the state classification number of the on state be 0, and obtain the switch state class set of the individual electrical appliance:

[0015]

[0016] Preferably, the specific process of step 2) is as follows:

[0017] 2.1) Sample once per minute, and sample the total current i t , total active power p t and total reactive power qt As the 1st, 2nd, and 3rd dimensional features of the training feature vector at time t;

[0018] 2.2) Calculate the difference i in the total current between time t and the previous time t-1 t -i t-1 and the difference p in the total active power t -p t-1 and the difference q in the total reactive power t -q t-1 As the 4th, 5th, and 6th dimensional features of the training feature vector at time t;

[0019] 2.3) Obtain the other N-dimensional features of the training feature vector at time t using the following formula:

[0020]

[0021] where cluster1(k) and cluster2(k) represent the clustering centers of the 1st state class and the 2nd state class of the kth electrical appliance respectively, and i t k represents the current value of the kth electrical appliance at time t. In this way, a total of 6+N dimensional features are obtained, and these 6+N dimensional features are used as the training features.

[0022] Preferably, the specific process of step 3) is as follows:

[0023] 3.1) Group all the electrical appliances into groups of 4. If the number of electrical appliances in the last group is less than 4, it is also counted as a group, and all the electrical appliances are divided into W groups of electrical appliances;

[0024] 3.2) Number the 4 electrical appliances in each group of electrical appliances in order from 1 to 4. When each electrical appliance is in the 1st state class at time t during training, it is in the standby or off state, represented by 1. When each electrical appliance is in the 2nd state class at time t during training, it is in the running or on state, represented by 0. In this way, the states of the 4 electrical appliances in each group of electrical appliances form a 4-bit binary number, which is converted into a decimal number with a range of 0-15 to obtain the decimal number switch state sequence of each group of electrical appliances; if the number of electrical appliances in the last group of electrical appliances is less than 4, the state range of the last group is smaller than 0-15;

[0025] 3.3) Combine the training features extracted in step 2) with the switch state sequence of each group obtained in step 3.2) to form an independent variable-dependent variable training data record sequence as the normalized training set.

[0026] Preferably, the specific process of step 4) is as follows:

[0027] 4.1) Construct W neural network models, each of which consists of a sequence input layer, two GRU layers, a fully connected layer, a Dropout layer, a fully connected layer, a softmax layer, and a classification layer;

[0028] 4.2) Feed the training set into the constructed neural network models for training to obtain W switch state identification models.

[0029] Preferably, in step 5), the 6 + N-dimensional features at the current moment during operation are constructed as the working features according to the same process as in steps 2.1)-2.3).

[0030] Preferably, the specific process of step 6) is as follows:

[0031] 6.1) For each group of electrical appliances, feed the working features into the trained switch state identification model to obtain the combined switch state category number, which is a positive integer between 0 and 15;

[0032] 6.2) Convert the combined switch state category number into a 4-bit binary number, padding with 0s in front if it is less than 4 bits;

[0033] 6.3) According to the serial number of each electrical appliance in the group of electrical appliances, obtain the state of the electrical appliance from the corresponding position of the 4-bit binary number. 0 indicates that the state of the electrical appliance at the current moment is the off state of not working, and 1 indicates the on state of working.

[0034] Compared with the prior art, the present invention has the following advantages: The low-frequency signal non-intrusive load switch state identification method of the present invention extracts the current, active power, reactive power, and related features of the total electricity meter, feeds them into the constructed neural network model, obtains the classification result, decomposes the classification result, and accurately obtains the switch states of each single electrical appliance, thus laying a foundation for subsequent further decomposition and identification. Further, in the preferred solution of the present invention, the electrical appliances are grouped in groups of 4, and the used switch state identification model is reduced to 1 / 4 of other known methods, greatly improving the processing efficiency. The present invention is based on the low-frequency non-intrusive signal processing technology with a sampling frequency of once per minute, without the need for hardware modification of the electricity meter, greatly saving costs, and having broad application and promotion prospects. Description of the Drawings

[0035] Figure 1 It is a process example diagram during the first group of grouped training in the embodiment. Detailed Embodiment

[0036] The following further describes the present invention in detail with reference to the embodiments of the drawings.

[0037] The low-frequency signal non-intrusive load switch state identification method of the embodiment includes the following steps:

[0038] 1) Collect training data on the current, active power, and reactive power of each single electrical appliance, cluster the current, active power, and reactive power of each single electrical appliance, and obtain a set of switching state classes of the single electrical appliance;

[0039] The specific process of step 1) is:

[0040] 1.1) Suppose there are N=6 electrical appliances, namely, basement lamp, dishwasher, washing machine, refrigerator, TV, and computer. The current, active power, and reactive power sequences in the training data are I k ,P k ,Q k , k=1,2,...N, use meanshift clustering algorithm to cluster them into M categories, requiring M<=10; when the number of categories M obtained by clustering is greater than 10, gradually increase the bandwidth parameter of meanshift algorithm by 5 each time until the number of categories M<=10; record the category number of each single appliance at any time t as m t ∈[1,M];

[0041] 1.2) Sort the cluster centers by current size. When the category number of a single electrical appliance is the smallest category, that is, m t =1, the appliance is considered to be in the off state, that is, the first state class; when the category number of the appliance is m t ∈[2, M], the appliance is considered to be in the on state, that is, the second state class;

[0042] Let the state classification number of the off state be 1, and the state classification number of the on state be 0, and the switch state class set of a single appliance is obtained:

[0043]

[0044] 2) Extract training features based on the total meter current, active power, and reactive power at a certain moment and the previous moment in the training data;

[0045] The specific process of step 2) is:

[0046] 2.1) Sampling once every minute, the total current i of the total meter at time t in the training data is obtained by sampling t , total active power p t and total reactive power q t As the 1st, 2nd, and 3rd dimension features of the training feature vector at time t;

[0047] 2.2) Calculate the difference i between the total current at time t and the previous time t-1 t -i t-1 , the difference in total active power p t -p t-1 The difference between the total reactive power and thet -q t-1 As the 4th, 5th, and 6th dimensional features of the training feature vector at time t;

[0048] 2.3) Obtain the other N-dimensional features of the training feature vector at time t using the following formula:

[0049]

[0050] where cluster1(k) and cluster2(k) represent the cluster centers of the first state class and the second state class of the kth electrical appliance respectively, and i t k represents the current value of the kth electrical appliance at time t. In this way, a total of 6 + N-dimensional features are obtained, and these 6 + N-dimensional features are used as the training features.

[0051] 3) Construct a normalized training set according to the training features and the known switch state sequences of each single electrical appliance;

[0052] The specific process of step 3) is as follows:

[0053] 3.1) Group all the electrical appliances in groups of 4. If the number of electrical appliances in the last group is less than 4, it is also counted as a group. All the electrical appliances are divided into W groups of electrical appliances. In this embodiment, all the electrical appliances are divided into W = 2 groups of electrical appliances. The first group has 4 electrical appliances, and the second group has 2 electrical appliances;

[0054] 3.2) Number the 4 electrical appliances in each group of electrical appliances in order from 1 to 4. When each electrical appliance is in the first state class at time t during training, it is in the standby or off state, represented by 1. When each electrical appliance is in the second state class at time t during training, it is in the running or on state, represented by 0. In this way, the states of the 4 electrical appliances in each group of electrical appliances form a 4-bit binary number, which is converted into a decimal number, ranging from 0 to 15, to obtain the decimal number switch state sequence of each group of electrical appliances; if the number of electrical appliances in the last group of electrical appliances is less than 4, the state range of the last group is smaller than 0 - 15;

[0055] In this embodiment, the first group has 4 electrical appliances, and the state range is 0 - 15; the second group has 2 electrical appliances, and the state range is 0 - 3;

[0056] 3.3) Combine the training features extracted in step 2) with the switch state sequences of each group obtained in step 3.2) to form an independent variable - dependent variable training data record sequence as the normalized training set;

[0057] 4) Construct a neural network model, send the training set into the neural network model for training, and obtain a switch state identification model;

[0058] The specific process of step 4) is as follows:

[0059] 4.1) Construct W neural network models, each of which consists of a sequence input layer, two GRU layers, a fully connected layer, a Dropout layer, a fully connected layer, a softmax layer, and a classification layer;

[0060] 4.2) Feed the training set into the constructed neural network models for training to obtain W switch state identification models;

[0061] In this embodiment, the number of constructed neural network models is 2, and the number of obtained switch state identification models is also 2. Among them, the process example diagram during the first group of grouped training is shown in Figure 1 ;

[0062] 5) When working, construct working features based on the total meter current, active power, and reactive power at the current moment and its previous moment. Specifically, construct the 6 + N-dimensional features at the current moment during work as working features according to the same process in steps 2.1)-2.3);

[0063] 6) Feed the working features into the trained switch state identification model to obtain the combined switch state, and decompose the switch states of each single electrical appliance from the combined switch state;

[0064] The specific process of step 6) is as follows:

[0065] 6.1) For each group of electrical appliances, feed the working features into the trained switch state identification model to obtain the combined switch state category number, which is a positive integer between 0 and 15;

[0066] 6.2) Convert the combined switch state category number into a 4-bit binary number, and fill 0 in front when it is less than 4 bits;

[0067] 6.3) According to the serial number of each electrical appliance in this group of electrical appliances, obtain the state of this electrical appliance from the corresponding position of the 4-bit binary number. 0 indicates that the state of this electrical appliance at the current moment is the off state of not working, and 1 indicates the on state of working.

Claims

1. A non-invasive load switch state identification method for low-frequency signals, characterized in that, Including the following steps: 1) Collect the training data of the current, active power, and reactive power of each single electrical appliance, cluster the current, active power, and reactive power of each single electrical appliance, and obtain the switch state class set of the single electrical appliance; 2) Extract training features according to the total meter current, active power, and reactive power at a certain moment and the previous moment in the training data. The specific process of step 2) is as follows: 2.1) Sample once per minute, and sample the total current i at the total electricity meter moment t in the training data t , the total active power p t and the total reactive power q t as the 1st, 2nd, and 3rd dimensional features of the training feature vector at moment t; 2.2) Calculate the difference i between the total current at time t and the total current at the previous time t-1 t -i t-1 and the difference p in the total active power t -p t-1 and the difference q in the total reactive power t -q t-1 as the 4th, 5th, and 6th dimensional features of the training feature vector at time t; 2.3) Use the following formula to obtain the other N-dimensional features of the training feature vector at time t: where cluster1(k) and cluster2(k) respectively represent the cluster centers of the first state class and the second state class of the k-th electrical appliance, and i t k represents the current value of the k-th electrical appliance at time t. In this way, a total of 6 + N-dimensional features are obtained, and these 6 + N-dimensional features are used as training features; 3) Construct a normalized training set according to the training features and the known switch state sequences of each single electrical appliance; 4) Construct a neural network model, send the training set into the neural network model for training, and obtain a switch state identification model; 5) Construct working features according to the total meter current, active power, and reactive power at the current moment and the previous moment during operation; 6) Send the working features into the trained switch state identification model to obtain a combined switch state, and decompose the switch states of each single electrical appliance from the combined switch state. The specific process of step 6) is as follows: 6.1) For each group of electrical appliances, send the working features into the trained switch state identification model to obtain a combined switch state category number, which is a positive integer between 0 and 15; 6.2) Convert the combined switch state category number into a 4-bit binary number, and fill in 0s in front when it is less than 4 bits; 6.3) Obtain the state of the electrical appliance according to the serial number of each electrical appliance in the group of electrical appliances from the corresponding position of the 4-bit binary number. 0 indicates that the state of the electrical appliance at the current moment is the off state of non-operation, and 1 indicates the on state of operation.

2. The non-invasive load switch state identification method for low-frequency signals according to claim 1, characterized in that, The specific process of step 1) is as follows: 1.1) There are N electrical appliances, and the current, active power, and reactive power sequences in their training data are I k , P k , Q k , where k = 1, 2,... N. The meanshift clustering algorithm is used to cluster them into M classes, with the requirement that M <= 10; when the number of classes M obtained by clustering is > 10, the bandwidth parameter of the meanshift algorithm is increased successively, with each value increased by 5 until the number of classes M <= 10; the class number of each single electrical appliance at any moment t is denoted as m t ∈[1, M]; 1.2) Sort the cluster centers according to the magnitude of the current. When the class number of a single electrical appliance is the smallest category, i.e., m t = 1, it is regarded that the electrical appliance is in the off state, that is, the first state class; when the class number m of the electrical appliance t ∈ [2, M], it is regarded that the electrical appliance is in the on state, that is, the second state class; Let the state classification number of the off state be 1, and the state classification number of the on state be 0, and obtain the switch state class set of the single electrical appliance:

3. The non-intrusive load switch state identification method for low-frequency signals according to claim 1, characterized in that, The specific process of step 3) is as follows: 3.1) Group all electrical appliances in groups of 4. If the number of electrical appliances in the last group is less than 4, it is also counted as a group, and all electrical appliances are divided into W groups of electrical appliances; 3.2) Number the 4 electrical appliances in each group of electrical appliances in order from 1 to 4. When each electrical appliance is in the first state class at time t during training, it is in the standby or off state, represented by 1. When each electrical appliance is in the second state class at time t during training, it is in the running or on state, represented by 0. In this way, the states of the 4 electrical appliances in each group of electrical appliances form a 4-bit binary number, which is converted into a decimal number, ranging from 0 to 15, to obtain the decimal number switch state sequence of each group of electrical appliances; if the number of electrical appliances in the last group of electrical appliances is less than 4, the state range of the last group is smaller than 0-15; 3.3) Combine the training features extracted in step 2) with the switch state sequences of each group obtained in step 3.2) to form an independent variable-dependent variable training data record sequence as a normalized training set.

4. A non-intrusive load switch state identification method for low-frequency signals according to claim 3, characterized in that, The specific process of step 4) is as follows: 4.1) Construct W neural network model numbers, each neural network model number consists of a sequence input layer, two GRU layers, a fully connected layer, a Dropout layer, a fully connected layer, a softmax layer, and a classification layer; 4.2) Send the training set into the constructed neural network model for training to obtain W switch state identification models.

5. A non-intrusive load switch state identification method for low-frequency signals according to claim 4, characterized in that In step 5), the 6+N-dimensional feature at the current moment during operation is constructed as the working feature according to the same process as in steps 2.1)-2.3).

Images