Low-frequency signal non-intrusive load state identification algorithm

Abstract

This invention discloses a non-intrusive load state identification algorithm for low-frequency signals, comprising the following steps: collecting training data on current, active power, and reactive power of each individual electrical appliance and clustering them to obtain a set of switching state classes for each appliance; extracting training features; constructing a deep neural network, constructing a standardized training set based on the extracted training features, and feeding the training set into the deep neural network for training to obtain a switching state identification model; constructing operating features during operation and feeding them into the trained switching state identification model to obtain the switching state of each individual electrical appliance; and using a recursive algorithm combined with the switching states to infer the operating state of each individual electrical appliance. This invention's algorithm can accurately determine the operating state of each individual electrical appliance, facilitating further processing. Based on a low-frequency non-intrusive signal processing technology sampling once per minute, this invention requires no hardware modification to the electricity meter, significantly saving costs and possessing broad application and promotion prospects.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of non-intrusive load monitoring, and particularly relates to a low-frequency signal non-intrusive load state recognition algorithm. BACKGROUND

[0002] With the continuous development of the economic society, the energy consumption of the whole society and the acquisition cost of traditional energy are also increasing. On the other hand, with the increasing emphasis on environmental protection around the world, higher requirements are put forward for the lean and effectiveness of energy utilization. Continuously optimizing and improving the relationship between the supply side and the demand side of electricity can alleviate the pressure on energy, improve energy utilization efficiency, and also reduce the cost of the demand side. In addition, China proposes to build a strong smart grid, and smart electricity is an important part of it. The first step is to be able to analyze the electricity consumption of residents in a lean way, understand the load and electricity consumption of each period in the residents' home, and guide residents to use electrical appliances reasonably and reduce unnecessary waste of electricity, so as to achieve the purpose of energy saving and sustainable development.

[0003] Lean analysis aims to accurately understand the load composition of the entire power system, the power distribution of each load, and the power-on time, and then reasonably regulate the load and arrange the operation time of the load. Non-intrusive load monitoring, also known as non-intrusive load decomposition / recognition, refers to installing a smart meter at the user's electricity inlet to obtain the user's total electricity information, and then using big data analysis technology to calculate the electricity consumption and energy consumption of various electrical appliances used by the user. It can directly use existing user meter data without adding equipment, has the advantages of easy monitoring and implementation, low cost, protection of user privacy, etc., and therefore has broad development prospects.

[0004] According to the frequency division of the processed signal, the non-intrusive load decomposition algorithm can be divided into two categories based on high-frequency signal (>=1HZ) and low-frequency signal (<1HZ). The algorithm based on high-frequency signal extracts wavelet features, voltage-current trajectory, current harmonics, active / reactive power, and other transient and steady-state features from high-frequency signals, and uses hidden Markov model, deep learning algorithm, etc. Model to decompose total power, total current, etc. However, the collection of high-frequency signals requires hardware modification of existing smart meters, resulting in certain additional economic costs, thereby restricting its large-scale application. The decomposition based on low-frequency signal can use the data collected by the existing meter as much as possible, and achieve little or no economic cost. SUMMARY

[0005] The purpose of the present application is to provide a low-frequency signal non-intrusive load state recognition algorithm, which can accurately obtain the working state of each single electrical appliance according to the total current, active power and reactive power collected by the total meter, so as to facilitate further processing.

[0006] The technical scheme adopted by the present application to solve the above technical problems is: a low-frequency signal non-invasive load state identification algorithm, comprising the following steps:

[0007] 1) Collecting training data of current, active power and reactive power of each single appliance, clustering the current, active power and reactive power of each single appliance to obtain a switch state class set of the single appliance;

[0008] 2) Extracting training features according to the total ammeter current, active power and reactive power at the current time and the previous time in the training data;

[0009] 3) Constructing a deep neural network, constructing a normalized training set according to the extracted training features, sending the training set into the deep neural network for training to obtain a switch state identification model;

[0010] 4) Constructing working features according to electrical parameters at the current time during work, sending the working features into the trained switch state identification model to obtain the switch state of each single appliance;

[0011] 5) Using a recursive algorithm to infer the working state of each single appliance combined with the switch state.

[0012] As a preferred, the specific process of step 1) is:

[0013] 1.1) Assuming that there are N appliances, the current, active power and reactive power sequences in the training data are I k ,P k ,Q k , k = 1, 2,... N, using meanshift clustering algorithm to cluster them into M classes, requiring M <= 10; when the number of classes M obtained by clustering is greater than 10, increase the bandwidth parameter of meanshift algorithm by 5 each time until the number of classes M is <= 10; the class number of each single appliance at any time t is recorded as m t ∈ [1, M];

[0014] 1.2) Sort the cluster centers by current size, when the class number of a single appliance is the smallest class, i.e. m t = 1, it is considered that the appliance is in the off state, i.e. the first state class; when the class number of the appliance is m t ∈ [1, M], it is considered that the appliance is in the on state, i.e. the second state class;

[0015] Let the state classification number of the off state be 1 and the state classification number of the on state be 0 to obtain a switch state class set of the single appliance:

[0016]

[0017] As preferred, the specific process of step 2) is as follows:

[0018] 2.1) Sample once per minute, and sample the total current i of the total electric meter at time t in the training data t , total active power p t and total reactive power q t as the 1st, 2nd and 3rd dimensional features of the training feature vector at time t;

[0019] 2.2) Calculate the difference i t -i t-1 , p t -p t-1 and q t -q t-1 of the total current at time t and the previous time t-1 as the 4th, 5th and 6th dimensional features of the training feature vector at time t.

[0020] As preferred, the specific process of step 3) is as follows:

[0021] 3.1) Construct a deep neural network composed of a sequence input layer, an LSTM layer, a full connection layer, a softmax layer and a classification layer;

[0022] 3.2) For each single appliance, input the extracted training features and the state classification number sequence into the constructed deep neural network for learning, to obtain a switch state recognition model, and N appliances correspond to N switch state recognition models.

[0023] As preferred, the specific process of step 4) is as follows:

[0024] 4.1) Cluster the total electric meter current, active power and reactive power sequences in the training data using the meanshift algorithm, and set the bandwidth parameter to be the minimum value of the bandwidth parameters of each single appliance in step 1.1), to obtain a clustering set of the total electric meter state categories;

[0025] 4.2) Calculate the distance dis(F, C i ) between the total electric meter current, active power and reactive power at each time t and each category in the clustering set in step 4.1) during work, and the calculation formula is as follows:

[0026]

[0027] In the above formula, F represents the total electric meter current, active power and reactive power at time t; C i represents the center of the clustering set in step 4.1); DN represents the feature dimension; F n represents the component value on the nth feature; and C i nThis represents the component value on the nth dimension of the i-th cluster center; This indicates that when all cluster centers are sorted according to the nth dimension, the value in C is... i n The value of the previous one in the nth dimension; Indicates ranking in C i n The value of the next one in the nth dimension; 1(C n ) n This represents the value in the nth dimension of the first element; -1(C n ) n This represents the value in the nth dimension that is the last one in the sequence;

[0028] 4.3) Assign the total meter status category at time t to the category with the smallest distance. If the total meter status category at the current time t is the same as the total meter status category at the previous time t-1, then continue to take the data at the next time t+1 for loop processing; otherwise, proceed to the next step.

[0029] 4.4) Construct the features at the current time t as working features using the same process as steps 2.1)-2.2);

[0030] 4.5) For each single electrical appliance, the operating characteristics are fed into the trained switch state identification model to obtain the switch state at the current time t.

[0031] As a preferred option, the specific process of step 5) is as follows:

[0032] 5.1) The total electrical parameters at the current time t Assign to the remaining total electrical parameters The following recursive algorithm is used to determine the operating state of each individual appliance at time t:

[0033] 5.1.1) When the current electrical appliance If it's not the last appliance, there are two possibilities: a) and b)

[0034] a) If the current electrical appliance If the current electrical appliance is in the off state, then the current electrical appliance is in the off state. Let the state at the current time t be m. t =1, remaining total electrical parameters Subtract current appliances Electrical parameters in the off state Right now Then continue recursively processing the next appliance until the current appliance is processed. For the last appliance, calculate the probability of each appliance taking the current switch state, i.e., the current probability P. C ;

[0035] b) If the current electrical appliance is not the off state, then for the current appliance all possible states m t ∈ [2, M], let the current appliance be in state m , and let the remaining total electrical parameters be reduced by the electrical parameters of the current appliance in state m . Then continue the recursive process for the next appliance until the current appliance is the last appliance, and then calculate the probabilities of each individual appliance taking the current switch state, i.e. the current probability P C .

[0036] The current probability with the largest calculated value is the maximum probability Pm, and if the current probability P C > the maximum probability P m , then the current probability P C is set to the maximum probability P m , and the current state sequence of each individual appliance is saved, the state sequence of each individual appliance that achieves the maximum probability is set as the current appliance state sequence, and the maximum probability P m is set as the current probability P C .

[0037] 5.1.2) When the current appliance is the last appliance, for each possible state m t ∈ [1, M], calculate the remaining electrical parameters , and then calculate the current probability P from the remaining electrical parameters C . The current probability with the largest calculated value is the maximum probability Pm, and if the current probability P C > the maximum probability P m , then the current probability P C is set to the maximum probability P m , and the current state sequence of each individual appliance is saved, the state sequence of each individual appliance that achieves the maximum probability is set as the current appliance state sequence, and the maximum probability P m is set as the current probability P C .

[0038] The calculation formula of the above-mentioned current probability P C is:

[0039]

[0040]

[0041] Compared with the prior art, the low-frequency signal non-intrusive load state recognition algorithm has the advantages that the low-frequency signal non-intrusive load state recognition algorithm can accurately obtain the working state of each single electric appliance according to the total current, active power and reactive power collected by the total electric meter, so as to facilitate further processing. The low-frequency non-intrusive signal processing technology based on 1-minute sampling once does not need to modify the hardware of the electric meter, greatly saves the cost, and has a wide application and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a total electric meter current graph in the embodiment;

[0043] Figure 2 is a state recognition result graph (television) in the embodiment. DETAILED DESCRIPTION

[0044] The application will be further described in detail below with reference to the embodiment and the accompanying drawings.

[0045] The low-frequency signal non-intrusive load state recognition algorithm of the embodiment includes the following steps:

[0046] 1) Collect the training data of the current, active power and reactive power of each single electric appliance, cluster the current, active power and reactive power of each single electric appliance, and obtain the on-off state class set of the single electric appliance; the specific process of step 1) is as follows:

[0047] 1.1) There are N=6 electric appliances, which are basement lamp, dishwasher, washing machine, refrigerator, television and computer, and the current, active power and reactive power sequences in the training data are I k ,P k ,Q k , k=1, 2,..., N, which are clustered into M classes by using the meanshift clustering algorithm, and it is required that M<=10; when the number of classes M obtained by clustering is greater than 10, the bandwidth parameter of the meanshift algorithm is increased by 5 each time until the number of classes M is less than or equal to 10; the class number of each single electric appliance at any time t is recorded as m t ∈[1, M];

[0048] 1.2) Sort the clustering centers according to the current size; when the class number of a single electric appliance is the smallest class, that is, m t =1, it is regarded as the off state, that is, the first state class; when the class number of the electric appliance is m t ∈[2, M], it is regarded as the on state, that is, the second state class;

[0049] 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 on-off state class set of the single electric appliance:

[0050]

[0051] 2) Extract training features according to the total ammeter current, active power and reactive power of the current moment and the previous moment in the training data; the specific process of step 2) is as follows:

[0052] 2.1) Sampling once per minute, and sampling the total current i t , the total active power p t and the total reactive power q t of the total ammeter moment t in the training data as the 1st, 2nd and 3rd dimensional features of the training feature vector at t moment;

[0053] 2.2) Calculating the difference i t -i t-1 , the difference p t -p t-1 and the difference q t -q t-1 of the total current, the total active power and the total reactive power between the moment t and the previous moment t-1 as the 4th, 5th and 6th dimensional features of the training feature vector at t moment;

[0054] 3) Constructing a deep neural network, constructing a normalized training set according to the extracted training features, sending the training set into the deep neural network for training to obtain a switch state recognition model; the specific process of step 3) is as follows:

[0055] 3.1) Constructing a deep neural network composed of a sequence input layer, an LSTM layer, a full connection layer, a softmax layer and a classification layer;

[0056] 3.2) For each single electrical appliance, sending the extracted training features and the state classification number sequence into the constructed deep neural network for learning to obtain a switch state recognition model, and N electrical appliances correspond to N switch state recognition models;

[0057] 4) During work, constructing working features according to the electrical parameters at the current moment, sending the working features into the trained switch state recognition model to obtain the switch state of each single electrical appliance; the specific process of step 4) is as follows:

[0058] 4.1) Clustering the total ammeter current, active power and reactive power sequence in the training data by meanshift algorithm, and setting the bandwidth parameter to be the minimum value of the bandwidth parameter of each single electrical appliance clustering in step 1.1) to obtain a clustering set of the total ammeter state category;

[0059] 4.2) During work, calculating the distance dis(F, C i ) between the total ammeter current, active power and reactive power at each moment t and each category in the clustering set in step 4.1), and the calculation formula is as follows:

[0060]

[0061] F represents the total electric meter current, active power, reactive power at time t; C i represents the center of the cluster set of step 4.1); DN represents the feature dimension; F n represents the component value on the nth dimension of the feature; C i n represents the component value on the nth dimension of the i-th cluster center; represents the value on the nth dimension of the first cluster center; -1(C i n represents the value on the nth dimension of the first cluster center; -1(C represents the value on the nth dimension of the first cluster center; -1(C i n represents the value on the nth dimension of the first cluster center; -1(C n ) n represents the value on the nth dimension of the first cluster center; -1(C n ) n represents the value on the nth dimension of the first cluster center; -1(C

[0062] 4.3) The total electric meter state category at time t is classified into the category with the smallest distance. If the total electric meter state category at the current time t is the same as that at the previous time t-1, the next time t+1 data is taken for loop processing, otherwise the next step is entered;

[0063] 4.4) The feature at the current time t is constructed as the working feature according to the same process of steps 2.1)-2.2);

[0064] 4.5) For each single appliance, the working feature is sent into the trained switch state recognition model to obtain the switch state at the current time t;

[0065] 5) The working state of each single appliance is inferred by using a recursive algorithm combined with the switch state. The specific process of step 5) is as follows:

[0066] 5.1) The total electric parameter at the current time t is assigned to the remaining total electric parameter The following recursive algorithm is used to determine the working state of each single appliance at the current time t:

[0067] 5.1.1) When the current appliance is not the last appliance, there are two cases a) and b) as follows:

[0068] a) If the state of the current appliance is the off state, the state of the current appliance at the current time t is set to mt =1, remaining total electrical parameters Subtract current appliances Electrical parameters in the off state Right now Then continue recursively processing the next appliance until the current appliance is processed. For the last appliance, calculate the probability of each appliance taking the current switch state, i.e., the current probability P. C ;

[0069] b) If the current electrical appliance If the state is not off, then for the current appliance... All possible states m t Let the current electrical appliance be ∈[2, M]. The current state at time t is this state, and the remaining total electrical parameters Subtract current appliances Electrical parameters of the current state Right now Then continue recursively processing the next appliance until the current appliance is processed. For the last appliance, calculate the probability of each appliance taking the current switch state, i.e., the current probability P. C ;

[0070] The maximum probability Pm is the current probability with the largest calculated value. C Maximum probability P m Then the current probability P C Let P be the maximum probability. m And save the current state sequence of each individual appliance, set the state sequence of each individual appliance with the highest probability as the current state sequence of the appliance, and set the highest probability P m Set to the current probability P C ;

[0071] 5.1.2) When the current electrical appliance When it is the last electrical appliance, for each possible state m t ∈[1, M], calculate the remaining electrical parameters Then based on the remaining electrical parameters Calculate the current probability P C The maximum probability Pm is the current probability with the largest calculated value. If the current probability P... C Maximum probability P m Then the current probability P C Let P be the maximum probability. m And save the current state sequence of each individual appliance, set the state sequence of each individual appliance with the highest probability as the current state sequence of the appliance, and set the highest probability P m Set to the current probability PC ;

[0072] The above current probability P C The calculation formula is:

[0073]

[0074]

[0075] Figure 2 Figure 2 shows the state recognition results in this embodiment, i.e., the states of the television and their corresponding current diagrams obtained in this embodiment. In Figure 2, the upper figure represents the current values ​​corresponding to the television state sequence identified by this algorithm, and the lower figure represents the current values ​​corresponding to the actual state sequence. Each class is represented by the current of the cluster center of that class.

Claims

1. A non-intrusive load condition identification algorithm for low-frequency signals, characterized in that, Includes the following steps: 1) Collect training data on the current, active power and reactive power of each individual electrical appliance, and cluster the current, active power and reactive power of each individual electrical appliance to obtain a set of switching state classes of the individual electrical appliances. 2) Extract training features based on the total meter current, active power, and reactive power at the current time and the previous time in the training data; 3) Construct a deep neural network, construct a standardized training set based on the extracted training features, and feed the training set into the deep neural network for training to obtain the switch state recognition model. 4) During operation, the working characteristics are constructed based on the electrical parameters at the current moment, and the working characteristics are fed into the trained switch state recognition model to obtain the switch state of each individual electrical appliance. 5) The working state of each individual electrical appliance is inferred by using a recursive algorithm combined with the switch status; The specific process of step 5) is as follows: 5.1) The total electrical parameters at the current time t Assign to the remaining total electrical parameters The following recursive algorithm is used to determine the operating state of each individual appliance at time t: 5.1.1) When the current electrical appliance If it's not the last appliance, there are two possibilities: a) and b) a) If the current electrical appliance If the current electrical appliance is in the off state, then the current electrical appliance is in the off state. Let the state at the current time t be m. t =1, remaining total electrical parameters Subtract current appliances Electrical parameters in the off state Right now Then continue recursively processing the next appliance until the current appliance is processed. For the last appliance, calculate the probability of each appliance taking the current switch state, i.e., the current probability P. C ; b) If the current electrical appliance If the state is not off, then for the current appliance... All possible states m t Let the current electrical appliance be ∈[2, M]. The current state at time t is this state, and the remaining total electrical parameters Subtract current appliances Electrical parameters of the current state Right now Then continue recursively processing the next appliance until the current appliance is processed. For the last appliance, calculate the probability of each appliance taking the current switch state, i.e., the current probability P. C ; The maximum probability Pm is the current probability with the largest calculated value. C Maximum probability P m Then the current probability P C Let P be the maximum probability. m And save the current state sequence of each individual appliance, set the state sequence of each individual appliance with the highest probability as the current state sequence of the appliance, and set the highest probability P m Set to the current probability P C ; 5.1.2) When the current electrical appliance When it is the last electrical appliance, for each possible state m t ∈[1, M], calculate the remaining electrical parameters Then based on the remaining electrical parameters Calculate the current probability P C The maximum probability Pm is the current probability with the largest calculated value. If the current probability P... C Maximum probability P m Then the current probability P C Let P be the maximum probability. m And save the current state sequence of each individual appliance, set the state sequence of each individual appliance with the highest probability as the current state sequence of the appliance, and set the highest probability P m Set to the current probability P C ; The above current probability P C The calculation formula is:

2. The low-frequency signal non-intrusive load condition identification algorithm according to claim 1, characterized in that, The specific process of step 1) is as follows: 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 Given k = 1, 2, ..., N, the mean shift clustering algorithm is used to cluster them into M clusters, where M <= 10. When the number of clusters M > 10, the bandwidth parameter of the mean shift algorithm is increased by 5 each time until the number of clusters M <= 10. The cluster number of each individual appliance at any time t is denoted as m. t ∈[1, M]; 1.2) Sort the cluster centers according to the magnitude of the current. The cluster center with the smallest category number for a single electrical appliance is m. t When m = 1, the appliance is considered to be in the off state, i.e., state type 1; when the appliance's category number m t When ∈[2, M], the electrical appliance is considered to be in the on state, which is the second state class; Assigning the state classification number of the off state to 1 and the state classification number of the on state to 0, we obtain the set of switch state classes for a single electrical appliance:

3. The low-frequency signal non-intrusive load condition identification algorithm according to claim 2, characterized in that, The specific process of step 2) is as follows: 2.1) Sample once per minute to obtain the total current i at time t of the total meter in the training data. t Total active power p t Total reactive power q t As the first, second, and third dimensions of the training feature vector at time t; 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 p in total active power t -p t-1 The difference q between the total reactive power and the total reactive power t -q t-1 The 4th, 5th, and 6th dimensions of the feature vector used for training at time t.

4. The low-frequency signal non-intrusive load condition identification algorithm according to claim 3, characterized in that, The specific process of step 3) is as follows: 3.1) Construct a deep neural network consisting of a sequence input layer, an LSTM layer, a fully connected layer, a softmax layer, and a classification layer; 3.2) For each single electrical appliance, the extracted training features and state classification number pair sequence are fed into the constructed deep neural network for learning to obtain the switch state identification model. N electrical appliances correspond to N switch state identification models.

5. The low-frequency signal non-intrusive load condition identification algorithm according to claim 4, characterized in that, The specific process of step 4) is as follows: 4.1) Cluster the total meter current, active power and reactive power sequences in the training data using the meanshift algorithm, and set the bandwidth parameter to the minimum value of the bandwidth parameter of each single appliance cluster in step 1.1) to obtain the cluster set of the total meter status categories; 4.2) During operation, calculate the total meter current, active power, and reactive power at each time t, and the distance dis(F, C) between these values ​​and each cluster in the cluster set from step 4.1). i The calculation formula is as follows: In the above formula, F represents the total meter current, active power, and reactive power at time t; C i The cluster centers are represented by DN (in step 4.1); DN represents the feature dimension. F n C represents the component value in the nth dimension of the feature; i n This represents the component value on the nth dimension of the i-th cluster center; This indicates that when all cluster centers are sorted according to the nth dimension, the value in C is... i n The value of the previous one in the nth dimension; Indicates ranking in C i n The value of the next one in the nth dimension; 1(C n ) n This represents the value in the nth dimension of the first element; -1(C n ) n This represents the value in the nth dimension that is the last one in the sequence; 4.3) Assign the total meter status category at time t to the category with the smallest distance. If the total meter status category at the current time t is the same as the total meter status category at the previous time t-1, then continue to take the data at the next time t+1 for loop processing; otherwise, proceed to the next step. 4.4) Construct the features at the current time t as working features using the same process as steps 2.1)-2.2); 4.5) For each single electrical appliance, the operating characteristics are fed into the trained switch state identification model to obtain the switch state at the current time t.

Images