Load state derivation method suitable for few-sample electric appliances

Abstract

The disclosed load state derivation method suitable for few-sample electric appliances adopts original current, active power and reactive power and respective relative change amount as input features, adopts an LSTM network combined with an attention mechanism to perform single-appliance state classification, ensures the accuracy of preliminary decomposition, adopts a weighted bias small-sample classification re-voting mechanism to obtain the final load state of the single appliance, and further improves the classification accuracy of the non-standby state. The method can accurately derive the load state of each electric appliance according to the current, active power and reactive power of the total electric meter, is particularly suitable for deriving the load state of small-sample electric appliances, is conducive to further processing, lays a foundation for further application, and has a wide application and promotion prospect. In addition, due to the low-frequency non-intrusive signal processing technology based on 1-minute sampling, the method does not need to modify the hardware of the electric meter, greatly saves the cost, and has a wide application and promotion prospect.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of non-intrusive load monitoring, and particularly relates to a load state derivation method suitable for few-sample electrical appliances. BACKGROUND

[0002] Currently, building energy efficiency plays an important role in reducing carbon emissions through energy monitoring and energy monitoring. Energy monitoring refers to accurately and in detail viewing the energy consumption of equipment. In addition, this information is also useful in the demand response of electrical appliances from the perspective of preventive maintenance and practicality. In order to achieve economic and efficient energy monitoring, non-intrusive load monitoring (NILM) is a promising solution. NILM is the process of disaggregating individual appliance consumption from a smart meter. Due to the large-scale deployment of smart meters, the hardware infrastructure of the NILM solution is easy to have. The current smart meter supports a sampling rate of 1 Hz and below, which can support a low-frequency NILM solution.

[0003] NILM can generally be divided into two methods: load classification and load disaggregation. Load classification usually involves feature extraction and classification to distinguish sub-loads in aggregated loads. Its main task is to identify the characteristics of a set of useful power loads. In this regard, various feature sets, including active power and reactive power features, are used to distinguish the profiles of sub-loads in aggregated loads, and in addition, the spikes of transient power can be used to identify power-on events. Recently, voltage-current (V-I) trajectories have been considered as image features for classification, which improves the uniqueness of the load and helps transfer learning. However, despite many efforts to determine the correct features of power loads, it is still difficult to find a set of features that can describe and distinguish different load categories. In load disaggregation, the energy decomposition problem is considered as a source separation task, in which a mixed signal is decomposed into the sum of multiple single signals. For the decomposition of multi-state loads, the literature [A practical solution for non-intrusive type II load monitoring based on deep learning and post-processing] proposes a convolutional neural network model based on post-processing. In addition, a CNN-based heat map related to the power load consumption profile is also applied to the decomposition of low-frequency loads. Although many methods have been proposed, their effects on multi-state load decomposition are not good. This is because the sample difference between different states of multi-state loads is huge, and the usual method is easy to make the state with more samples drown the state with less samples, and most electrical appliances are usually in standby state, so the decomposition effect of non-standby state is not good. SUMMARY

[0004] The technical problems to be solved by the present application are to provide a load state derivation method suitable for few-sample electrical appliances, to overcome the shortcomings of the prior art, to use original current, active power and reactive power and their respective relative change amounts as input features, and to use an LSTM network combined with an attention mechanism for single-appliance state classification, to ensure the accuracy of preliminary decomposition, and to use a weighted bias small-sample classification re-voting mechanism to obtain the final load state of the single appliance, and to further improve the classification accuracy of the non-standby state.

[0005] The technical solution adopted by the present application to solve the above technical problems is: a load state derivation method suitable for few-sample electrical appliances, comprising the following steps:

[0006] 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, and obtaining a working state set of the single appliance;

[0007] 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, and obtaining a working state set of the total circuit;

[0008] 3) Constructing a training data set for each single appliance, simplifying the data corresponding to the dominant appliance state, and obtaining preprocessed training data;

[0009] 4) Constructing a state classification deep neural network, training the state classification deep neural network using the preprocessed training data, and obtaining a single-appliance working state classification model;

[0010] 5) Calculating the main class average distribution probability of each single appliance in different working states of the total circuit;

[0011] 6) Collecting the current, active power and reactive power of the total electric meter when it is working, and inputting the collected data into the single-appliance working state classification model to classify the working state of each single appliance;

[0012] 7) Comparing the working states of the total circuit at different times, and correcting the working state of each single appliance to the small-sample state class according to the working state of the total circuit and the main class average distribution probability of each single appliance in different working states, to obtain the optimized working state class, i.e. the load state of each single appliance.

[0013] As a preferred embodiment, the specific process of step 1) is as follows:

[0014] 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 k'=1,2,...N, using meanshift clustering method to gather M classes, requiring M≤10;

[0015] When the number of classes M>10 is obtained by clustering, the bandwidth parameter of the meanshift method is increased successively, and the value is increased by 5 each time until the number of classes M≤10.

[0016] 1.2) Sort the cluster centers by current size, with small current first and large current last, and number them. The class with the smallest number is considered to be in the off state (also known as standby state), otherwise in the on state, obtaining the working state set of single appliances.

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

[0018] 2.1) Sample once a minute, collect the training data of current, active power and reactive power of the total electric meter, and take the current i t , active power p t , and reactive power q t of the total electric meter at time t in the training data as the features of the feature vector at time t.

[0019] 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 electrical features of the total electric meter. Let D classes be clustered, and the jth class be Obtain the working state set of the total circuit.

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

[0021] 3.1) For any single appliance, if the proportion of the number of samples of a certain working state S in the training data exceeds 80% and the data of the working state S is not simplified, record the number of samples of the working state S in the training data as n1, traverse the original sample data belonging to the working state S, and set the time from t i to t j The working state of the appliance is S, i.e. State[t i ,t j ]=S, then only the sample data at time t i and time t j are retained, and the sample data between time t i and time t j is deleted, obtaining the simplified sample data.

[0022] Delete the sample data between time t i and time t jAfter the sample data between the work states S is sampled, if the number n2 of the sample data of the work state S remaining in the work state S is less than the number of the sample data of the work state S in the training data by m, the sample data is sampled every (n1-n2) / m sample data from the original sample data of the work state S, and the sample data is added to the reduced sample data to form a sample set of the work state S composed of n2+m sample data, and the sample set of the work state S is used as the reduced sample data.

[0023] 3.2) The step 3.1) is repeatedly performed until the proportion of the sample number of one type of work state in the training data is more than 80%, and the preprocessed training data is obtained.

[0024] As preferred, the specific process of the step 4) is as follows:

[0025] 4.1) A state classification deep neural network is constructed by sequentially stacking a two-layer LSTM network, an LSTM network attention layer, a full connection layer, a full connection layer, and a classification layer;

[0026] 4.2) For each piece of preprocessed training data, the current i t , active power p t , and reactive power q t of the total electric meter at time t are taken as the first part of features;

[0027] 4.3) From time t-1, the time is traversed forward until the first time inconsistent with the work state of the total circuit at time t is found, which is recorded as time t-k, and the current, active power, and reactive power of the total electric meter at time t-k are recorded as i t-k , p t-k , and q t-k respectively, and the relative change amounts of the current, active power, and reactive power of the total electric meter at time t and time t-k are calculated as i t -i t-k , p t -p t-k , and q t -q t-k as the second part of features;

[0028] 4.4) The first part of features and the second part of features are combined as input features, which are sent into the state classification deep neural network for training to obtain a single-appliance work state classification model.

[0029] As preferred, the specific process of the step 5) is as follows:

[0030] 5.1) For each work state section of the total circuit in the training data , if the work state section corresponds to the work state of a certain single-appliance E, the category of the work state of the single-appliance E is , and the work state section Each time point of the corresponding time period is classified into a state by the single appliance working state classification model The number of samples is The working state section The number of samples contained is The working state section The main class distribution probability of the corresponding single appliance E The calculation formula is:

[0031]

[0032] 5.2) Calculate the main class average distribution probability of the working state section of the single appliance E in the total circuit

[0033] 5.3) Calculate the main class average distribution probability of the single appliance E in the total circuit in other different working state sections and the main class average distribution probability of other single appliances in the total circuit in different working state sections.

[0034] As a preferred, the specific process of step 6) is:

[0035] 6.1) Sample once a minute, collect the data of current, active power and reactive power when the total electric meter is working, and take the current i' t , active power p' t and reactive power q' t as feature one from the collected data;

[0036] 6.2) Traverse from time t-1 until the first time point inconsistent with the working state of the total circuit at time t is found, recorded as time t-k1, and the current, active power and reactive power of the total electric meter at time t-k1 are recorded as i' t-k1 , p' t-k1 , q' t-k1 respectively, calculate the relative change of current, active power and reactive power of the total electric meter at time t and time t-k1 i' t -i' t-k1 , p' t -p' t-k1 , q' t -q' t-k1 as feature two;

[0037] 6.3) Combine feature one and feature two into input features and input into the single appliance working state classification model to classify the working state of each single appliance.

[0038] As a preferred, the specific process of step 7) is:

[0039] ​7.1) the working state of the total circuit at time t is classified into the working state set of the total circuit, if the working state of the total circuit at time t is inconsistent with the category of the working state of the total circuit at time t-1, the working state of the single appliance obtained in step 6) is directly taken as the final load state of the single appliance; if the working state of the total circuit at time t is consistent with the category of the working state of the total circuit at time t-1, step 7.2) is entered;

[0040] 7.2) the time t is traversed forward until the first time inconsistent with the working state of the total circuit at time t is found, recorded as time t-k2, the number of occurrences of each working state of the single appliance between time t-k2+1 and time t is counted, and the number of occurrences of a working state C is recorded as then the scalar scale correction value of the possibility size of the working state C is wherein P(C S ) indicates the probability of the single appliance appearing in the working state C in all working state segments of the total circuit;

[0041] 7.3) the scalar scale correction value of the possibility size of the other working states of the single appliance is calculated, and the working state corresponding to the maximum scalar scale correction value of all the scalar scale correction values of the single appliance is taken as the load state of the single appliance between time t-k2+1 and time t.

[0042] Compared with the prior art, the present application has the following advantages: the present application adopts the original current, active power and reactive power and their respective relative change amounts as input features, and adopts the LSTM network combined with the attention mechanism to classify the single appliance state, thereby ensuring the accuracy of the preliminary decomposition, and adopting the weighted biased small sample classification re-voting mechanism to obtain the final load state of the single appliance, thereby further improving the classification accuracy of the non-standby state. The method of the present application can accurately deduce the load state of each appliance according to the current, active power and reactive power of the total meter, is particularly suitable for deducing the load state of a small sample appliance, is conducive to further processing, and lays a foundation for further application. In addition, due to the low-frequency non-intrusive signal processing technology based on 1-minute sampling, the method of the present application does not need to modify the hardware of the meter, greatly saves the cost, and has a broad application and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0043] Fig. 1 the state of the television set in the embodiment and the current diagram corresponding thereto;

[0044] Fig. 2 the current diagram of the total meter;

[0045] Fig. 3 the actual current diagram of the television set. DETAILED DESCRIPTION

[0046] The application will be described in further detail below with reference to the drawings.

[0047] The load state derivation method of the embodiment suitable for few-sample electrical appliances comprises the following steps:

[0048] 1) Collecting training data of current, active power and reactive power of each single electrical appliance, clustering the current, active power and reactive power of each single electrical appliance to obtain a working state set of the single electrical appliance;

[0049] The specific process of step 1) is as follows:

[0050] 1.1) Assuming that there are N=6 electrical appliances, which are basement light, dishwasher, washing machine, refrigerator, television and computer, collecting training data of current, active power and reactive power of each electrical appliance at time t, the current, active power and reactive power sequences in the training data are respectively k′=1, 2,...N, adopting meanshift clustering method to cluster them into M classes, and requiring that M≤10;

[0051] 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;

[0052] 1.2) Sorting the clustering centers according to the current size, with small current in front and large current behind, and numbering them, regarding the class with the smallest number as the off state of the electrical appliance, otherwise as the on state, to obtain a working state set of the single electrical appliance;

[0053] 2) Collecting training data of current, active power and reactive power of the total meter, clustering the current, active power and reactive power of the total meter to obtain a working state set of the total circuit;

[0054] The specific process of step 2) is as follows:

[0055] 2.1) Sampling once every minute to collect training data of current, active power and reactive power of the total meter, taking the current i t , active power p t and reactive power q t of the total meter at time t in the training data as the features of the training feature vector at time t;

[0056] 2.2) Taking the minimum bandwidth parameter value in the clustering of each single electrical appliance in step 1.1) as the current bandwidth parameter value, clustering the electrical features of the total meter, assuming that there are D classes, and the jth class is to obtain a working state set of the total circuit;

[0057] 3) constructing a training data set of each single appliance, refining the data corresponding to the dominant appliance state, and obtaining preprocessed training data;

[0058] The specific process of step 3) is as follows:

[0059] 3.1) For any single appliance, if the proportion of the number of samples of a certain working state S in the training data exceeds 80% and the data of the working state S is not refined, record the number of samples of the working state S in the training data as n1, traverse the original sample data belonging to the working state S, and set the time from t i to t j The working state of the appliance is S, that is, State[t i ,t j ] = S, then only the sample data at time t i and time t j are retained, the sample data between time t i and time t j are deleted, and the refined sample data is obtained;

[0060] After deleting the sample data between time t i and time t j , if the remaining number of samples n2 of the working state S is less than m than the number of samples of the third working state in the training data, then sample the sample data every (n1-n2) / m samples from the original sample data of the working state S, add the refined sample data, and form a sample set of the working state S consisting of n2+m sample data. The sample set of the working state S is used as the refined sample data.

[0061] 3.2) cyclically execute step 3.1) until the proportion of the number of samples of any working state in the training data does not exceed 80%, and obtain the preprocessed training data;

[0062] 4) constructing a state classification deep neural network, training the state classification deep neural network using the preprocessed training data, and obtaining a single appliance working state classification model;

[0063] The specific process of step 4) is as follows:

[0064] 4.1) constructing a state classification deep neural network stacked by two layers of LSTM network, LSTM network attention layer, fully connected layer, fully connected layer, and classification layer in sequence;

[0065] 4.2) for each piece of preprocessed training data, taking the current i t , active power p t and reactive power q t of the total meter at time t as the first part of the features;

[0066] 4.3) Traverse from time t-1 to the first time point that is inconsistent with the working state of the total circuit at time t, denoted as time t-k, and record the current, active power and reactive power of the total meter at time t-k as i t-k ,p t-k ,q t-k , Calculate the relative change of the current, active power and reactive power of the total meter at time t and time t-k i t -i t-k ,p t -p t-k ,q t -q t-k as the second part of the features;

[0067] 4.4) Combine the first part of the features and the second part of the features into input features, and input them into the state classification deep neural network for training to obtain a single appliance working state classification model;

[0068] 5) Calculate the main class average distribution probability of each single appliance in different working states under different total circuit working states;

[0069] The specific process of step 5) is as follows:

[0070] 5.1) For each working state segment of the total circuit in the training data such as the working state segment The class of the working state of a certain single appliance E corresponding to the working state segment is And the working state segment Each time point in the corresponding time segment is classified by the single appliance working state classification model as state The number of samples of state The number of samples contained in the working state segment is recorded as Then the working state segment The main class distribution probability of the single appliance E corresponding to the working state segment is The calculation formula is:

[0071]

[0072] 5.2) Calculate the main class average distribution probability of the single appliance E in the working state segment of the total circuit

[0073] 5.3) Calculate the main class average distribution probability of the single appliance E in other different working state segments of the total circuit and the main class average distribution probability of other single appliances in different working state segments of the total circuit;

[0074] 6) Collect the current, active power and reactive power of the main meter when it is working, and send the collected data into the single appliance working status classification model to classify the working status of each single appliance.

[0075] The specific process of step 6) is as follows:

[0076] 6.1) Sample once per minute to collect data on the current, active power, and reactive power of the main meter during operation, and extract the current i' of the main meter at time t from the collected data. t Active power p' t and reactive power q' t As a characteristic one;

[0077] 6.2) Starting from time t-1, iterate backwards until the first time that the operating state of the total circuit at time t is inconsistent with that at time t, and denote it as time t-k1. Record the current, active power, and reactive power of the main meter at time t-k1 as i' respectively. t-k1 ,p' t-k1 ,q' t-k1 Calculate the relative changes i' of current, active power, and reactive power at time t and time t-k1 of the main meter. t -i' t-k1 ,p' t -p' t-k1 ,q' t -q' t-k1 As feature two;

[0078] 6.3) Combine feature one and feature two into input features and feed them into the single appliance working state classification model to classify the working state of each single appliance;

[0079] 7) Compare the working states of the total circuit at different times. Based on the working states of the total circuit and the average distribution probability of the main class of different working states of each individual appliance, perform state correction on the working states of each individual appliance with a bias towards the small sample state class to obtain the optimized working state class, which is the load state of each individual appliance.

[0080] The specific process of step 7) is as follows:

[0081] 7.1) Assign the working state of the total circuit at time t to the working state set of the total circuit. If the working state of the total circuit at time t is inconsistent with the working state at time t-1, then directly take the working state of the single appliance obtained in step 6) as the final load state of the single appliance; if the working state of the total circuit at time t is consistent with the working state at time t-1, proceed to step 7.2).

[0082] 7.2) Traverse forward from time t until the first time that is inconsistent with the working state of the total circuit at time t, denoted as time t-k2, count the number of occurrences of each working state of the single appliance between time t-k2+1 and time t, and let the number of occurrences of a working state C be Then the scalar scale correction value of the possibility size of the working state C is Where P(C S ) refers to the probability of the single appliance appearing in the working state C in all working state segments of the total circuit;

[0083] 7.3) Calculate the scalar scale correction value of the possibility size of the other working states of the single appliance, and take the working state corresponding to the largest scalar scale correction value among all the calculated scalar scale correction values of the single appliance as the load state of the single appliance between time t-k2+1 and time t.

[0084] In this embodiment, the load state derivation result of the television set is shown in Figs. 1-3 . Fig. 1 The state of the television set in the embodiment and the corresponding current diagram. Fig. 2 The current diagram of the total ammeter. Fig. 3 The actual current diagram of the television set, represented by the current of the corresponding cluster center.

Claims

1. A method for load state derivation suitable for few sample electric appliances, characterized in that, 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) constructing a training data set of each single appliance, simplifying the data corresponding to the dominant appliance state to obtain preprocessed training data; 4) constructing a state classification deep neural network, training the state classification deep neural network using the preprocessed training data to obtain a single appliance working state classification model; The specific process of step 4) is as follows: 4.1) constructing a state classification deep neural network stacked by two layers of LSTM network, LSTM network attention layer, full connection layer, full connection layer and classification layer in sequence; 4.2) For each training data, take the current i at the total meter time t t , the active power p t and the reactive power q t as the first part of features; 4.3) Traverse from time t-1 forward until the first time is found that is inconsistent with the operating state of the total circuit at time t, denoted as time t-k, and record the current, active power and reactive power of the total meter at time t-k as i t-k ,p t-k ,q t-k , Calculate the relative change of current, active power and reactive power of the total meter at time t and time t-k i t -i t-k ,p t -p t-k ,q t -q t-k As the second part of the feature; 4.4) merging the first part of features and the second part of features into input features, and inputting the input features into the state classification deep neural network for training to obtain a single appliance working state classification model; 5) calculating the main class average distribution probability of each single appliance in different working states of the total circuit; The specific process of step 5) is as follows: 5.1) For each working state section of total circuit in training data As working state section The category of working state of corresponding certain single appliance E is And the single appliance E is in working state section Each time point of corresponding time section is classified as state The sample number of state The working state section The sample number contained is recorded as The working state section The main category distribution probability of corresponding single appliance E The calculation formula is: 5.2) Calculate the single appliance E in the total circuit operating state segment of the main class average distribution probability 5.3) calculating the main class average distribution probability of the single appliance E in other different working state segments of the total circuit and the main class average distribution probability of other single appliances in different working state segments of the total circuit; 6) collecting current, active power and reactive power when the total electric meter is working, and inputting the collected data into the single appliance working state classification model to classify the working state of each single appliance; 7) comparing the working states of the total circuit at different times, and correcting the working state of each single appliance to a small sample state class according to the working state of the total circuit and the main class average distribution probability of each single appliance in different working states to obtain an optimized working state class, which is the load state of each single appliance.

2. The load state derivation method for a few sample appliance suitable according to claim 1, characterized in that, The specific process of step 1) is as follows: 1.1) Set up N electrical appliances, collect the training data of current, active power and reactive power at time t, the current, active power and reactive power sequences in the training data are respectively Use the meanshift clustering method to cluster them into M classes, and require M≤10; When the number of clusters M>10, the bandwidth parameter of the meanshift method is increased by 5 each time until the number of clusters M≤10; 1.2) sorting the cluster centers according to the current size, with small current first and large current last, and numbering them, with the smallest one being considered as the off state of the appliance, otherwise as the on state, to obtain a working state set of the single appliance.

3. The load state derivation method for a few sample appliance suitable according to claim 2, characterized in that, The specific process of step 2) is as follows: 2.1) Sample once per minute, collect the training data of the current, active power and reactive power of the total electric meter, take the current i t , active power p t , reactive power q t of the total electric meter at time t in the training data as the features of the 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 electrical characteristics of the total meter, set the total cluster as D class, and the jth class as Get the working state set of the total circuit.

4. The load state derivation method for a few sample appliance suitable according to claim 3, characterized in that, The specific process of step 3) is as follows: 3.1) For any single appliance, if the proportion of the number of samples of its certain working state S in the training data exceeds 80% and the data of working state S is not refined, the number of samples of working state S in the training data is recorded as n1, and the original sample data belonging to working state S is traversed. Set from time t i to time t j , the working state of the appliance is S, that is, State[t i ,t j ] = S, only the sample data at time t i and time t j are retained, and the sample data between time t i and time t j is deleted, and the refined sample data is obtained; deletion time t i and time t j After the sample data between the time t is deleted, if the remaining sample number n2 of the working state S is less than the sample number of the third working state in the training data by m, then sample the sample data every (n1-n2) / m samples from the original sample data of the working state S, add the simplified sample data, and form a sample set of the working state S consisting of n2+m sample data, and use the sample set of the working state S as the simplified sample data. 3.2) cyclically executing step 3.1) until the sample number of each working state in the training data is less than 80%, to obtain preprocessed training data.

5. The load state derivation method for a few sample appliance suitable according to claim 1, characterized in that, The specific process of step 6) is as follows: 6.1) Sampling once per minute, collecting data of the current, active power and reactive power of the total meter in operation, taking the current i' of the total meter at time t from the collected data as a feature one; i' = i(t) t p' = p(t) t q' = q(t) t ​ 6.2) Traversing from time t-1 forward until the first time is found that is inconsistent with the operating state of the total circuit at time t, denoted as time t-k1, and the current, active power and reactive power of the total meter at time t-k1 are denoted as i' t-k1 ,p' t-k1 ,q' t-k1 , the relative change amounts of the current, active power and reactive power of the total meter at time t and time t-k1 are calculated as i' t -i' t-k1 ,p' t -p' t-k1 ,q' t -q' t-k1 as feature two; 6.3) merging feature one and feature two into input features, and inputting the input features into the single appliance working state classification model to classify the working state of each single appliance.

6. The load state derivation method for a few sample appliance suitable according to claim 5, characterized in that, The specific process of step 7) is as follows: 7.1) the working state of the total circuit at time t is classified into the working state set of the total circuit, if the working state of the total circuit at time t is not consistent with the category of the working state of the total circuit at time t-1, the working state of the single appliance obtained in step 6) is directly taken as the final load state of the single appliance; if the working state of the total circuit at time t is consistent with the category of the working state of the total circuit at time t-1, step 7.2) is entered; 7.2) Traverse from time t forward until the first time t-k2 is found which is inconsistent with the working state of the total circuit at time t, count the number of occurrences of each working state of the single appliance between time t-k2+1 and time t, and let the number of occurrences of a working state C be Then the scalar scale correction value of the possibility size of the working state C is Where P(C S ) refers to the probability of the occurrence of the working state C of the single appliance in all working state segments of the total circuit; 7.3) the scalar scale correction value of the possibility size of the other working state of the single appliance is calculated, and the working state corresponding to the maximum of all the calculated scalar scale correction values of the single appliance is taken as the load state of the single appliance at time t-k2+1 to time t.

Images