Appliance state derivation method based on electrical change amount state transition probability matching

Abstract

The application discloses an electrical appliance state derivation method based on electrical change amount state conversion probability matching, combines a method based on a graph model and a method based on deep learning, derives the state change of the electrical appliance by the method based on deep learning, derives the specific state change condition of the electrical appliance by the method based on the graph model, and can better handle the derivation problem of the low-frequency electrical signal electrical appliance. Because the advantages of the two methods are combined, the required data amount is greatly reduced. In addition, the application adopts a low-frequency non-invasive signal processing technology based on 1-minute sampling once, does not need to make hardware transformation on the electric meter, greatly saves the cost, and has wide application and popularization prospects.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of non-intrusive load monitoring, and particularly relates to an electrical appliance state derivation method based on electrical change amount state transition probability matching. BACKGROUND

[0002] Non-intrusive load monitoring can realize the identification of each electrical equipment and its working state by analyzing and processing total load meter data, and can be widely applied to the fields of building energy saving, smart city and smart grid. In recent years, with the large-scale deployment of smart meters and the wide application of various machine learning algorithms, non-intrusive load monitoring has attracted the attention of both academia and industry. The difference between different non-intrusive load monitoring schemes mainly lies in the different load features and load decomposition algorithms. Among numerous load decomposition models and methods, the Hidden Markov model (HMM) has attracted a lot of attention and research due to its ability to model time series data and represent unobservable states. In the past five years, methods based on graph signal processing (GSP) and deep learning (DL) have also been gradually applied in this field and have achieved better performance than traditional methods. On the other hand, these three algorithms also have some shortcomings. The HMM-based method has large computational load and accurate reasoning is difficult; the graph-based method is currently less used and cannot be well modeled; the deep learning-based method has many model parameters and requires a large amount of data.

[0003] Most current non-intrusive load identification methods use high-frequency data sampling to extract and utilize transient characteristics, which requires modification of existing meters, hindering their large-scale practical application. When transient characteristics are not used, the accuracy is not high due to the small amount of data. SUMMARY

[0004] The technical problem to be solved by the present application is to provide an electrical appliance state derivation method based on electrical change amount state transition probability matching, which combines the graph-based method and the deep learning-based method, uses the deep learning method to derive the state change of the electrical appliance, and uses the graph model method to derive the specific state change of the electrical appliance, thereby better handling the electrical appliance derivation problem of low-frequency electrical signals.

[0005] The technical scheme adopted by the present application to solve the above technical problem is as follows: an electrical appliance state derivation method based on electrical change amount state transition probability matching, comprising the following steps:

[0006] 1) Collecting training data of current, active power and reactive power of each single appliance, clustering current, active power and reactive power of each single appliance to obtain a working state set of single appliance;

[0007] 2) Collecting training data of current, active power and reactive power of total electric meter, clustering current, active power and reactive power of total electric meter to obtain a working state set of total circuit;

[0008] 3) Clustering the electrical characteristic difference between adjacent time of different working states of total circuit in the working state set of total circuit to obtain a change quantity category set of total circuit;

[0009] 4) Constructing a state transition probability matrix according to the transition frequency of adjacent time between different working states of single appliance in the working state set of single appliance;

[0010] 5) Constructing a neural network unit with attention mechanism, constructing input features and output features according to the working state change of total circuit and the working state change of single appliance, inputting the input features and output features into the neural network unit for learning to obtain a total circuit state change coding model;

[0011] 6) Using the total circuit state change coding model to obtain a set of appliances whose working state changes, calculating the probability of each working state of each appliance in the set of appliances by using the state transition probability matrix, and taking the working state with the maximum probability as the current state of the appliance.

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

[0013] 1.1) There are N appliances, and the training data of current, active power and reactive power at time t is collected, and the current, active power and reactive power sequences in the training data are respectively Using meanshift clustering method to cluster them into M classes, and requiring M≤10;

[0014] When the number of classes M obtained by clustering is greater than 10, the bandwidth parameter of meanshift method is increased by 5 each time until the number of classes M is less than or equal to 10;

[0015] 1.2) The clustering centers are sorted according to the current size, the larger the current, the larger the class number, and the smallest class is regarded as the off state of the appliance, otherwise it is the on state, and the working state set of single appliance is obtained.

[0016] As preferred, the specific process of step 2) is:

[0017] 2.1) Sampling once every minute, and the current i of total electric meter at time t in the training data of current, active power and reactive power of total electric meter is sampledt active power p t reactive power q t as the feature of the training feature vector at time t;

[0018] 2.2) Take the minimum bandwidth parameter value in the clustering of each single appliance in step 1.1) as the current bandwidth parameter value, and cluster the total meter features, assuming that there are D classes, and the jth class is Get the working state set of the total circuit.

[0019] As a preferred, the specific process of step 3) is:

[0020] 3.1) Calculate the electrical feature difference between different working states of the total circuit at adjacent times Where j, l are two state classes of the total circuit appearing at adjacent times;

[0021] 3.2) Cluster the above electrical feature differences, control the number of classes within 50, and get the change class set of the total circuit.

[0022] As a preferred, the specific process of step 4) is:

[0023] 4.1) Calculate the transition frequency f between different working states of single appliances at adjacent times uv , u, v are adjacent working states, u≠v;

[0024] 4.2) Convert the above transition frequency into probability to get the state transition probability matrix Where v' is all possible next working states, when u=v, let f' uv =1.

[0025] As a preferred, the specific process of step 5) is:

[0026] 5.1) Construct three encoding-decoding neural network units with attention mechanism, each encoding-decoding neural network unit is composed of an encoder and a decoder, the structure of the encoder includes two layers of LSTM network, the structure of the decoder includes two layers of LSTM network and attention layer, and the three encoding-decoding neural network units are sequentially stacked;

[0027] 5.2) Construct input features and output features, in the class sequence of the change class set of the total circuit, take a sliding window with a length of 10, take the electrical feature difference in the window as the current input feature, take the list of single appliances whose working state changes when the working state of the total circuit changes the last time as the current output feature, and take a placeholder number as the current output feature when no single appliance changes the working state;

[0028] 5.3) Input and output features are fed into the encoding / decoding neural network unit as training data for learning to obtain the encoding / decoding model of the overall circuit state change.

[0029] As a preferred option, the specific process of step 6) is as follows:

[0030] 6.1) The set of electrical appliances whose operating states change is obtained by using the overall circuit state change encoding and decoding model;

[0031] 6.2) There are c electrical appliances e1, e2, ..., e c The working states of the elements simultaneously changed, with their previous working states being u1, u2, ..., u... c Since the change in electrical characteristics corresponding to the change in the operating state of the electrical appliance starting from this point may not be an optimal match with the change in the electrical characteristics of the overall circuit, an intermediate state w is introduced. k and the transition of the target state v k The state u of each electrical appliance is calculated using the following formula. k via w k Convert to v k Maximum probability:

[0032] P(u1→w1→v1,u2→w2→v2,…,u c →w c →v c )=

[0033]

[0034] In the above formula, d s d(w) represents the change in the total electrical characteristics of the circuit, while d(w) represents the change in the total electrical characteristics of the circuit. k ,v k ) indicates that the k-th electrical appliance is in state w k Change to v k Changes in electrical characteristics over time This represents the state transition probability of the k-th appliance;

[0035] Directly v k Electrical characteristic quantity minus w k The electrical characteristic quantities at that time are used to obtain the changed electrical components e1, e2, ..., e c The working states with the highest probability v1, v2, ..., v c The working states with the highest probability v1, v2, ..., v c That is, electrical appliances e1, e2, ..., e c The optimal working state, with the optimal working state v1, v2, ..., v c As electrical appliances e1, e2, ..., e cThe current state of the electrical appliances is determined according to the current state of the electrical appliances, and the electrical appliances whose working states do not change keep the original working states.

[0036] Compared with the prior art, the present application has the following advantages: the present application combines the method based on the graph model and the method based on deep learning, uses the method based on deep learning to derive the state change of the electrical appliances, and uses the method based on the graph model to derive the specific state change of the electrical appliances, so that the derivation problem of the electrical appliances with low-frequency electrical signals can be better handled. Because the advantages of the two methods are combined, the required data amount is greatly reduced. In addition, the present application uses a low-frequency non-intrusive signal processing technology based on 1-minute sampling, 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

[0037] Fig. 1 is the current state of the refrigerator and the corresponding current diagram thereof;

[0038] Fig. 2 is the current diagram of the total electric meter;

[0039] Fig. 3 is the actual current diagram of the refrigerator. DETAILED DESCRIPTION

[0040] The present application will be further described in detail below in combination with the embodiments of the drawings.

[0041] The electrical appliance state derivation method based on the state transition probability matching of the electrical change amount of the embodiment includes the following steps:

[0042] 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 working state set of the single electrical appliance;

[0043] The specific process of step 1) is as follows:

[0044] 1.1) There are N=6 electrical appliances, which are the basement lamp, dishwasher, washing machine, refrigerator, television and computer, and the training data of the current, active power and reactive power of each electrical appliance at time t is collected. The current, active power and reactive power sequences in the training data are The meanshift clustering method is used to cluster them into M classes, and it is required that M≤10;

[0045] When the number of classes M obtained by clustering is greater than 10, the bandwidth parameter of the meanshift method is increased by 5 each time until the number of classes M is less than or equal to 10;

[0046] 1.2) Sort the cluster centers by the current size, the larger the current, the larger the category number, the smallest class is considered as the off state of the electrical appliance, otherwise it is the on state, and the working state set of single electrical appliance is obtained;

[0047] 2) Collect the training data of the current, active power and reactive power of the total meter, cluster the current, active power and reactive power of the total meter, and obtain the working state set of the total circuit;

[0048] The specific process of step 2) is as follows:

[0049] 2.1) Sample once a minute, and sample the training data of the current, active power and reactive power of the total meter at time t of the total meter t i t , active power p t , and reactive power q t as the features of the training feature vector at time t;

[0050] 2.2) Take the minimum bandwidth parameter value in step 1.1) as the current bandwidth parameter value, cluster the total meter features, set the total number of clusters to D, and the jth cluster is to obtain the working state set of the total circuit;

[0051] 3) Cluster the electrical feature difference between adjacent time points of different working states of the total circuit in the working state set of the total circuit, and obtain the change amount category set of the total circuit;

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

[0053] 3.1) Calculate the electrical feature difference between adjacent time points of different working states of the total circuit Where j and l are two state classes of the total circuit at adjacent time points;

[0054] 3.2) Cluster the above electrical feature differences, and control the number of categories within 50 to obtain the change amount category set of the total circuit;

[0055] 4) Construct a state transition probability matrix according to the transition frequency of adjacent time points between different working states of single electrical appliance in the working state set of single electrical appliance;

[0056] The specific process of step 4) is as follows:

[0057] 4.1) Calculate the transition frequency f uv of adjacent time points between different working states of single electrical appliance

[0058] 4.2) Convert the above transition frequency to probability to obtain the state transition probability matrix where v' is all possible next working states, let f u v = 1;

[0059] 5) Constructing the coding-decoding neural network unit with attention mechanism, constructing the input feature and the output feature according to the working state change of the total circuit and the working state change of the single appliance, sending the input feature and the output feature into the coding-decoding neural network unit as training data for learning, and obtaining the total circuit state change coding-decoding model;

[0060] The specific process of step 5) is as follows:

[0061] 5.1) Constructing three coding-decoding neural network units with attention mechanism, each coding-decoding neural network unit is composed of an encoder and a decoder, the structure of the encoder includes two layers of LSTM network, the structure of the decoder includes two layers of LSTM network and an attention layer, and the three coding-decoding neural network units are sequentially stacked;

[0062] 5.2) Constructing the input feature and the output feature, in the category sequence of the total circuit change amount category set, taking a sliding window with a length of 10, taking the electrical feature difference in the window as the current input feature, taking the single appliance list whose working state changes when the working state of the total circuit changes last time as the current output feature, and taking a placeholder number as the current output feature when no single appliance changes its working state;

[0063] 5.3) Sending the input feature and the output feature into the coding-decoding neural network unit as training data for learning, and obtaining the total circuit state change coding-decoding model;

[0064] 6) Obtaining the appliance set whose working state changes by using the total circuit state change coding-decoding model, calculating the probability of each working state of each appliance in the appliance set by using the state transition probability matrix, and taking the working state with the maximum probability as the current state of the appliance;

[0065] The specific process of step 6) is as follows:

[0066] 6.1) Obtaining the appliance set whose working state changes by using the total circuit state change coding-decoding model;

[0067] 6.2) Assuming that the working states of c appliances e1, e2, …, e c have changed at the same time, and the working states before the change are u1, u2, …, u c , the working state change of the appliance starting from this point may not be the optimal matching with the electrical feature change of the total circuit, so the intermediate state w k and the conversion target state v​k The state u of each appliance is calculated by the following formula k The maximum probability of converting from w k to v k

[0068]

[0069] In the above formula, d s represents the total circuit electrical characteristic variation, and d(w k , v k ) represents the electrical characteristic variation when the kth appliance changes from state w k to v k , represents the state transition probability of the kth appliance;

[0070] The electrical characteristic quantity at v k is directly subtracted from the electrical characteristic quantity at w k , so that the most probable working state v1, v2, …, v c of the changed appliances e1, e2, …, e c is obtained. The most probable working state v1, v2, …, v c is the optimal working state of the appliances e1, e2, …, e c , and the optimal working state v1, v2, …, v c is taken as the current state of the appliances e1, e2, …, e c , and the appliances whose working states do not change remain in their original working states.

[0071] Figs. 1-3 is a diagram of the state of the refrigerator in the embodiment. Fig. 1 is a diagram of the state of the refrigerator and its corresponding current in the embodiment, indicating the current value corresponding to the category sequence of the state of the refrigerator recognized by the method. Fig. 2 is a diagram of the current of the total meter. Fig. 3 is a diagram of the actual current of the refrigerator, represented by the current of the corresponding cluster center.​

Claims

1. An electrical appliance state derivation method based on electrical change amount state transition probability matching, characterized by, The method comprises the following steps: 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 working state set of the single appliance; 2) collecting training data of current, active power and reactive power of the total electric meter, clustering the current, active power and reactive power of the total electric meter to obtain a working state set of the total circuit; 3) clustering the electrical characteristic difference between different working states of the total circuit at adjacent time points in the working state set of the total circuit to obtain a change quantity category set of the total circuit; 4) constructing a state transition probability matrix according to the conversion frequency of adjacent time points between different working states of the single appliance in the working state set of the single appliance; 5) constructing a coding-decoding neural network unit with an attention mechanism, constructing input features and output features according to the working state change of the total circuit and the working state change of the single appliance, and inputting the input features and the output features into the coding-decoding neural network unit as training data for learning to obtain a total circuit state change coding-decoding model; 6) obtaining an appliance set with working state change by using the total circuit state change coding-decoding model, calculating the probability of each working state of each appliance in the appliance set by using the state transition probability matrix, and taking the working state with the maximum probability as the current state of the appliance.

2. The method of claim 1, wherein, The specific process of step 1) is as follows: 1.1) Set up N electrical appliances, collect their training data of current, active power and reactive power at time t, the training data of current, active power and reactive power sequences are respectively Use meanshift clustering method to cluster them into M classes, and require M≤10; When the number of categories M obtained by clustering is greater than 10, the bandwidth parameter of the meanshift method is increased by 5 each time until the number of categories M is less than or equal to 10; 1.2) sorting the clustering centers according to the current size, the larger the current, the larger the category number, and the smallest category is regarded as the off state of the appliance, otherwise as the on state, to obtain the working state set of the single appliance.

3. The method of claim 2, wherein the state of the electrical appliance is determined based on the state transition probability matching the electrical variation amount. The specific process of step 2) is as follows: 2.1) Sample once per minute, sample the total electric meter current, active power and reactive power training data of the total electric meter time t t , active power p t , reactive power q t as a feature of the training feature vector at time t; 2.2) Take the minimum bandwidth parameter value of each single appliance cluster in step 1.1) as the current bandwidth parameter value, and cluster the total meter characteristics. Let the total number of clusters be D, and the jth cluster be Get the working state set of the total circuit.

4. The method of claim 3, wherein the state of the electrical appliance is derived based on a probability matching of state transitions of the electrical change quantity. The specific process of step 3) is as follows: 3.1) Calculate the difference of electrical characteristics between different states of the total circuit at adjacent time instants where j, l are two state classes of the total circuit at adjacent time instants. 3.2) clustering the above electrical characteristic difference, the number of categories is controlled within 50, and the change quantity category set of the total circuit is obtained.

5. The method of claim 4, wherein the state of the electrical appliance is determined based on the state transition probability matching the amount of change in the electrical quantity. The specific process of step 4) is as follows: 4.1) Calculate the switching frequency f between adjacent time instants of different operating states of the single appliance uv u, v are adjacent time instants of operating states, u≠v; 4.2) Convert the above transition frequencies into probabilities to obtain a state transition probability matrix where v' is the set of all possible next operating states, and f' = 1 when u = v. uv = 1.

6. The method of claim 5, wherein the state of the electrical appliance is determined based on the state transition probability matching the electrical variation amount. The specific process of step 5) is as follows: 5.1) constructing three coding-decoding neural network units with an attention mechanism, each coding-decoding neural network unit is composed of an encoder and a decoder, the structure of the encoder includes two layers of LSTM network, the structure of the decoder includes two layers of LSTM network and attention layer, and the three coding-decoding neural network units are sequentially stacked; 5.2) constructing input features and output features, taking a sliding window with a length of 10 in the category sequence of the change quantity category set of the total circuit, taking the electrical characteristic difference in the window as the current input feature, and taking the list of single appliances with working state change when the working state of the total circuit changes last time as the current output feature, when there is no single appliance with working state change, using a placeholder number as the current output feature; 5.3) inputting the input features and the output features into the coding-decoding neural network unit as training data for learning to obtain a total circuit state change coding-decoding model.

7. The method of claim 6, wherein the state of the electrical appliance is determined based on the state transition probability matching the electrical variation amount. The specific process of step 6) is as follows: 6.1) obtaining an appliance set with working state change by using the total circuit state change coding-decoding model; 6.2) Set of c appliances e1, e2, …, e c whose working states have changed simultaneously, and the working states before the change are u1, u2, …, u c , the change of the working state of the appliance starting from this point corresponds to the change of the electrical characteristics of the total circuit, which is not necessarily the optimal match, so the intermediate state w k and the conversion target state v k are introduced, and the maximum probability of each appliance state u k converted to v k through w k is calculated: In the above formula, d s represents the total circuit electrical characteristic variation, and d(w k , v k ) represents the electrical characteristic variation when the kth electrical appliance changes from state w k to v k , represents the state transition probability of the kth electrical appliance. directly subtracting the electrical characteristic quantity at time w k from the electrical characteristic quantity at time v k , so as to obtain the working state v1, v2, …, v c with the largest probability of the electrical appliances e1, e2, …, e c , the working state v1, v2, …, v c with the largest probability is the optimal working state of the electrical appliances e1, e2, …, e c , and the optimal working state v1, v2, …, v c is taken as the current state of the electrical appliances e1, e2, …, e c , and the electrical appliances whose working states do not change maintain their original working states.

Images