Non-intrusive load sensing method fusing multi-time scale appliance signatures

Abstract

The application discloses a non-intrusive load sensing method fusing multi-time scale electrical appliance features. The method first carries out power consumption data collection and preprocessing, then detects whether an electrical appliance event occurs and whether a power sequence fluctuates. After the event ends and it is found that the power sequence is fluctuating, multi-time scale electrical appliance features are extracted, a sensing model is constructed and trained, and finally, the load sensing fusing the multi-time scale electrical appliance features is carried out. The application can comprehensively fuse the electrical appliance features on the multi-time scale, extract more comprehensive electrical appliance characteristics on different time scales, effectively balance the classification result errors on different time scales, and finally scientifically and reasonably perform the non-intrusive load sensing on the multi-time scale.

Description

Technical Field

[0001] This invention belongs to the field of smart electricity technology, specifically relating to a non-intrusive load sensing method that integrates the characteristics of electrical appliances across multiple time scales. Background Technology

[0002] Residential and commercial energy consumption accounts for nearly half of global energy consumption. Load monitoring and identification are of great importance, as they can help buildings improve energy efficiency and reduce carbon emissions. At the same time, for end users, load monitoring can provide them with information on electricity consumption at different times, helping them save energy.

[0003] Non-intrusive load monitoring (NILM) has become a promising new research topic in recent years. Unlike traditional invasive load monitoring, NILM aims to analyze the operating status of electrical appliances within a residence by analyzing aggregated current and voltage signals measured at the entrance. It does not require connection to the internal electrical circuits or individual sensors for appliances, thus saving time and costs. While NILM is relatively more challenging than invasive load monitoring, it offers advantages such as low cost and ease of operation. Furthermore, thanks to the rapid development of sensor and information technologies, NILM has gradually become a focus of research and attention.

[0004] Non-intrusive load sensing works by leveraging the uniqueness and repeatability of load characteristics. A typical algorithmic framework for a non-intrusive load sensing system includes steps such as data measurement, data preprocessing, electrical event and fluctuation detection, electrical feature extraction, sensing model training, and load identification. Feature extraction is a crucial technology. After data measurement, data preprocessing, and electrical event and fluctuation detection are completed, the quality of the feature extraction step directly affects the quality of subsequent steps.

[0005] Load characteristics mainly consist of two aspects: first, electrical characteristics, which are determined by the topology of the internal circuit and can be obtained by observing the voltage and current waveforms under the power frequency cycle; these are primarily load characteristics on a short time scale. Second, operating modes, which are determined by the control strategy and can be obtained by observing long-term operating data of the electrical equipment; these represent load characteristics on a long time scale. In the past, researchers focused too much on load identification at a single time scale, lacking a comprehensive consideration of load characteristics at different time scales. This resulted in past research methods being unable to handle loads that are similar at one time scale but different at other time scales, ultimately leading to low accuracy in load identification and matching.

[0006] In summary, how to scientifically and rationally extract non-intrusive load features across multiple time scales has become an urgent problem to be solved. Therefore, a non-intrusive load sensing method that integrates electrical appliance features across multiple time scales is needed. Summary of the Invention

[0007] The purpose of this invention is to provide a non-intrusive load sensing method that integrates multiple time-scale electrical appliance features in order to solve the above-mentioned problems.

[0008] To achieve the objectives of the invention described above, the technical solution adopted by the present invention is as follows:

[0009] A non-intrusive load sensing method that integrates multiple time-scale electrical appliance characteristics, the method comprising the following steps:

[0010] S1. Periodically sample the voltage and current at the entrance of the user's residence and calculate the effective value of active power per second;

[0011] S2. By performing electrical event and fluctuation detection on the effective value of active power per second calculated in step S1, it is possible to detect whether any electrical appliances are turned on or off or whether there are fluctuating electrical appliances in operation.

[0012] S3. When it is detected that an electrical appliance is turned on or off or that a fluctuating electrical appliance is in operation, the current data sampled in step S1 is processed at three different time scales: short time scale, transition process, and long time scale, to obtain the current sequence to be identified at each time scale; read the electrical appliance feature set in the load identification feature library, calculate the similarity between the current sequence to be identified at each time scale and each element in the electrical appliance feature set at the corresponding time scale, and obtain the electrical appliance feature vector of the current sequence to be identified.

[0013] S4. For each time scale, construct a bagging classifier with a machine learning machine as the base learner. The input of the bagging classifier is the feature vector of the appliance, and the output is the predicted appliance category probability. The load sensing model is composed of the bagging classifiers at the three time scales. In the load sensing model, the appliance category probabilities output at the three time scales are averaged to obtain the final appliance category probability. The appliance category with the highest probability is taken as the final category of the appliance corresponding to the current input. The load sensing model is trained using historical appliance feature vector data with appliance category labels.

[0014] S5. Input the electrical appliance feature vector of the current sequence to be identified into the trained load sensing model to obtain the prediction result and realize load identification.

[0015] Based on the above technical solution, the present invention may further adopt the following preferred solutions.

[0016] Preferably, in step S1, the voltage and current at the entrance of the user's residence are periodically sampled at sampling frequency f1 and alternating current frequency f2 to form a voltage sampling sequence u1(k) and a current sampling sequence i1(k), where k is the sampling point number, the sampling frequency f1 ranges from 0.4kHz to 50kHz, and the alternating current frequency f2 is 50Hz; then the effective value of active power per second is calculated to form an active power sequence P1, where the effective value of active power corresponding to time t in the sequence is P1(t).

[0017] Preferably, the specific implementation of electrical event and fluctuation detection in step S2 is as follows:

[0018] When performing electrical event detection, the change in active power between adjacent moments (i.e., adjacent seconds) is calculated as shown in equation (1):

[0019] ΔP(t-1,t)=|P1(t)-P1(t-1)| (1)

[0020] Where ΔP(t-1,t) is the change in the effective value of active power between time t and the previous time t-1; if the condition shown in equation (2) is satisfied at time t1, it is considered that an electrical event has occurred:

[0021]

[0022] Where t1 and t2 are the start and end times of the electrical event, respectively, and T f T is the threshold for event mutation. e The threshold for the occurrence of the event;

[0023] When performing volatility detection, set a sliding window G. long The time length is M, the sliding interval is N, and the sliding window G... long The active power range R and variance s are calculated according to formula (3). 2 Then slide window G long The active power range R and variance s in 2 Each with its own threshold R e and By comparison, when the condition shown in equation (4) is met, it is determined that an appliance with a power change is operating at this time;

[0024]

[0025]

[0026] in It is a sliding window G long The average active power of M active power RMS values.

[0027] As a preferred option, the sliding window G long The preferred time duration M is 5 minutes, and the sliding interval is 1 minute; the active power threshold R e The preferred value is 100, and the variance threshold is... The preferred value is 400.

[0028] Preferably, in step S3, the short time scale is one cycle of the alternating current frequency f2; the transition process is preferably 2 seconds after the electrical event occurs; and the long time scale is preferably greater than 1 minute.

[0029] Preferably, the current sequence to be identified on the short time scale is a current sampling sequence of 0.02 seconds, the current sequence to be identified during the transition process is the effective value of the current per cycle within 2 seconds after the electrical event occurs, and the current sequence to be identified on the long time scale is a sequence of root mean square current values ​​per second.

[0030] Preferably, in step S3, the electrical feature set of the load identification feature library is calculated as follows:

[0031] The electrical appliance feature scheme is a shapelet, which is defined as a subsequence that can represent the shape features of a time series. If there are M1 types of electrical appliances, each type of electrical appliance has M2 historical current sequences collected at each time scale. For any time scale, the length of the historical current sequence at the current time scale is L. The steps for calculating the shapelet feature set at the current time scale include:

[0032] ①Similarity calculation: For each type of electrical appliance, generate all subsequences V with lengths ranging from 3 to L by iterating through the M2 historical current sequences W.

[0033] The distance between sequences is defined as dist(T,R), and the distance between a subsequence and a current sequence is defined as subdist(T,R). The calculation formulas for both are shown in equations (5) and (6):

[0034]

[0035]

[0036] In the formula: T and R are current sequences of length m, and S is a subsequence of length l in the current sequence T;

[0037] For all subsequences V, calculate the distance subdist(W,V) between the subsequence V and its corresponding original historical current sequence W as the similarity δ1 between the subsequence V and the original historical current sequence W; simultaneously, calculate the historical current sequences W of the subsequence V and all other categories of electrical appliances. other1 W other2 …、Wothern The similarity between subdist(W) other1 ,V), subdist(W other2 ,V), ...,subdist(W othern ,V), where n is the total number of all historical current sequences for all other categories of electrical appliances;

[0038] ② Binary Information Gain Calculation and Ranking: For each subsequence V, among the distances between subsequence V and other original historical current sequences of the same category of electrical appliances, the number of sequences with a distance greater than δ1 is a1, and the number of sequences with a distance less than or equal to δ1 is a2; at the same time, among the distances between subsequence V and all original historical current sequences of other categories of electrical appliances, the number of sequences with a distance greater than δ1 is b1, and the number of sequences with a distance less than δ1 is b2; then the information gain Gain(V) of subsequence V is calculated using equations (7) and (8):

[0039]

[0040]

[0041] For each category of electrical appliances, k-means clustering is performed on all subsequences to remove subsequences with similar shapes. Then, their information gains are ranked, and the top 5 subsequences by information gain are used to form the shapelet feature set J(k,m1) for the corresponding category at the current time scale. Here, k is the time scale type index (k=1 represents a short time scale, k=2 represents a transition process, k=3 represents a long time scale), and m1 is the category index of the electrical appliance. Finally, the shapelet feature set J at the current time scale... k ={J(k,1),J(k,2),...,J(k,M1)}.

[0042] Preferably, in step S3, the electrical feature vector of the current sequence to be identified is calculated as follows:

[0043] For each time scale k, calculate the current sequence to be identified at that time scale and the shapelet feature set J corresponding to that time scale. k The distance sim of each subsequence element k 5M1 distance simulations were obtained at each time scale. k This forms a 5M1-dimensional feature vector of electrical appliances at this time scale.

[0044] Preferably, in step S4, when constructing and training the perception model, the number of base learners is M at each time scale. D During training, the samples input to each base learner are extracted using a bootstrap method; the maximum number of features that each base learner can learn is set to k.Di , and k Di It needs to be less than 5M1; for each base learner, all training features still need to be extracted from the original features using a bootstrapping method.

[0045] Preferably, the machine learning machine employs a decision tree or a support vector machine.

[0046] The beneficial effects of this invention are as follows: This invention can integrate electrical appliance features at multiple time scales, extract more comprehensive electrical appliance characteristics at different time scales, and effectively balance the classification results errors at different time scales, ultimately enabling scientific and reasonable non-intrusive load sensing at multiple time scales. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the non-intrusive load sensing method that integrates multiple time-scale electrical appliance features according to the present invention;

[0048] Figure 2 It is a classification confusion matrix with two time scales and integration;

[0049] Figure 3 It is the feature vector of the electrical appliances in the training samples;

[0050] Figure 4 These are the operating current waveforms and shapelet sequences of two types of air conditioners. Detailed Implementation

[0051] The technical solution of the present invention will be further described and illustrated clearly and completely below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention, and the technical features of each embodiment of the present invention can be combined accordingly without mutual conflict.

[0052] like Figure 1 As shown, in a preferred embodiment of the present invention, a non-intrusive load sensing method integrating multi-timescale electrical appliance characteristics is provided, the method comprising the following steps S1 to S5:

[0053] S1. Measurement and preprocessing of user electricity consumption data: Periodically sample the voltage and current at the entrance of the user's residence and calculate the effective value of active power per second.

[0054] In an embodiment of the present invention, in step S1 above, the voltage and current at the entrance of the user's residence are periodically sampled at sampling frequency f1 and alternating current frequency f2 to form a voltage sampling sequence u1(k) and a current sampling sequence i1(k), where k is the sampling point number, the sampling frequency f1 ranges from 0.4kHz to 50kHz, and the alternating current frequency f2 is 50Hz; then the effective value of active power per second is calculated to form an active power sequence P1, where the effective value of active power corresponding to time t in the sequence is P1(t).

[0055] In step S1 above, the voltage and current at the entrance of the user's residence are periodically sampled at sampling frequency f1 and alternating current frequency f2 to form a voltage sampling sequence u1(k) and a current sampling sequence i1(k), where k is the sampling point number, the sampling frequency f1 ranges from 0.4kHz to 50kHz, and the alternating current frequency f2 is 50Hz; then the effective value of active power per second is calculated to form an active power sequence P1, where the effective value of active power corresponding to time t in the sequence is P1(t).

[0056] S2. Electrical Event and Fluctuation Detection: By performing electrical event and fluctuation detection on the effective value of active power per second calculated in step S1, it can detect whether electrical appliances are turned on or off or whether there are fluctuating electrical appliances in operation.

[0057] In an embodiment of the present invention, the specific implementation of electrical event and fluctuation detection in step S2 above is as follows:

[0058] When performing electrical event detection, the change in active power between adjacent moments (adjacent seconds) is calculated as shown in equation (1):

[0059] ΔP(t-1,t)=|P1(t)-P1(t-1)| (1)

[0060] Where ΔP(t-1,t) is the change in the effective value of active power between time t and the previous time t-1; if the condition shown in equation (2) is satisfied at time t1, it is considered that an electrical event has occurred:

[0061]

[0062] Where t1 and t2 are the start and end times of the electrical event, respectively, and T f T is the threshold for event mutation. e The threshold for the occurrence of the event;

[0063] The principle for volatility detection is as follows:

[0064] ① Set the sliding window G long The time length is M, and the sliding interval is N;

[0065] ② In the sliding window G long The active power range R and variance s are calculated according to formula (3). 2 ;

[0066] ③Then slide window G long The active power range R and variance s in 2 Each with its own threshold R e and By comparison, when the condition shown in equation (4) is met, it is determined that an appliance with a power change is operating at this time;

[0067]

[0068]

[0069] in It is a sliding window G long The average active power of the M active power RMS values.

[0070] In an embodiment of the present invention, the sliding window G long The preferred time duration M is 5 minutes, and the sliding interval is 1 minute; the active power threshold R e The preferred value is 100, and the variance threshold is... The preferred value is 400.

[0071] S3. Electrical Appliance Feature Extraction: When an electrical appliance is detected to be turned on or off, or when a fluctuating electrical appliance is in operation, the current data sampled in step S1 is processed at three different time scales: short time scale, transition process, and long time scale, to obtain the current sequence to be identified at each time scale; the electrical appliance feature set in the load identification feature library is read, and the similarity between the current sequence to be identified at each time scale and each element in the electrical appliance feature set at the corresponding time scale is calculated to obtain the electrical appliance feature vector of the current sequence to be identified.

[0072] In an embodiment of the present invention, in step S3 above, the short time scale is preferably one cycle of the alternating current frequency f2; the transition process is preferably 2 seconds after the electrical event occurs; and the long time scale is preferably greater than 1 minute. Since the preferred value of the alternating current frequency in step S1 is 50Hz, the preferred value of the short time scale is 0-0.02 seconds. In an embodiment of the present invention, the current sequence to be identified on the short time scale is preferably a current sampling sequence with a length of 0.02 seconds, the current sequence to be identified on the transition process is preferably the effective value of the current per cycle within 2 seconds after the electrical event occurs, and the current sequence to be identified on the long time scale is preferably a sequence of root mean square current values ​​per second.

[0073] In addition, in step S3, the calculation method for the electrical feature set of the load identification feature library is as follows:

[0074] The electrical appliance feature scheme is a shapelet, which is defined as a subsequence that can represent the shape features of a time series. If there are M1 types of electrical appliances, each type of electrical appliance has M2 historical current sequences collected at each time scale. For any time scale, the length of the historical current sequence at the current time scale is L. Then, the calculation steps for the shapelet feature set (i.e., the shapelet shape set) at the current time scale include:

[0075] ①Similarity calculation: For each type of electrical appliance, generate all subsequences V with lengths ranging from 3 to L by iterating through the M2 historical current sequences W.

[0076] The distance between sequences is defined as dist(T,R), and the distance between a subsequence and a current sequence is defined as subdist(T,R). The calculation formulas for both are shown in equations (5) and (6):

[0077]

[0078]

[0079] In the formula: T and R are current sequences of length m, and S is a subsequence of length l in the current sequence T;

[0080] For all subsequences V, calculate the distance subdist(W,V) between the subsequence V and its corresponding original historical current sequence W as the similarity δ1 between the subsequence V and the original historical current sequence W; simultaneously, calculate the historical current sequences W of the subsequence V and all other categories of electrical appliances. other1 W other2 …、W othern The similarity between subdist(W) other1 ,V), subdist(W other2 ,V), ...,subdist(W othern ,V), where n is the total number of all historical current sequences for all other categories of electrical appliances;

[0081] ② Binary Information Gain Calculation and Ranking: For each subsequence V, among the distances between subsequence V and other original historical current sequences of the same category of electrical appliances, the number of sequences with a distance greater than δ1 is a1, and the number of sequences with a distance less than or equal to δ1 is a2; at the same time, among the distances between subsequence V and all original historical current sequences of other categories of electrical appliances, the number of sequences with a distance greater than δ1 is b1, and the number of sequences with a distance less than δ1 is b2; then the information gain Gain(V) of subsequence V is calculated using equations (7) and (8):

[0082]

[0083]

[0084] For each category of electrical appliances, k-means clustering is performed on all subsequences to remove subsequences with similar shapes. Then, their information gains are ranked, and the top 5 subsequences by information gain are used to form the shapelet feature set J(k,m1) for the corresponding category at the current time scale. Here, k is the time scale type index (k=1 represents a short time scale, k=2 represents a transition process, k=3 represents a long time scale), and m1 is the category index of the electrical appliance. Finally, the shapelet feature set J at the current time scale... k ={J(k,1),J(k,2),...,J(k,M1)}.

[0085] In addition, in step S3, the electrical feature vector of the current sequence to be identified is calculated as follows:

[0086] For each time scale k, calculate the current sequence to be identified at that time scale and the shapelet feature set J corresponding to that time scale. k The distance sim of each subsequence element (each element is a subsequence) k 5M1 distance simulations were obtained at each time scale. k This forms a 5M1-dimensional feature vector of electrical appliances at this time scale.

[0087] S4. Load Sensing Model Construction and Training: The load sensing model is constructed using an ensemble bagging classifier with a general machine learning machine as the base learner. Specifically, for each time scale, a bagging classifier with a machine learning machine as the base learner is constructed. The input of the bagging classifier is the appliance feature vector, and the output is the predicted appliance category probability. The load sensing model is composed of bagging classifiers at three time scales. The appliance category probabilities output at the three time scales are averaged to obtain the final appliance category probability. The appliance category with the highest probability is selected as the final appliance category corresponding to the current input. The load sensing model is trained using historical appliance feature vector data containing appliance category labels.

[0088] In an embodiment of the present invention, in step S4 above, when constructing and training the perceptual model, the number of base learners is M at each time scale. D During training, the samples input to each base learner are extracted using a bootstrap method; the maximum number of features that each base learner can learn is set to k. Di , and k DiIt needs to be less than 5M1; for each base learner, all training features still need to be extracted from the original features using a bootstrapping method.

[0089] In this invention, the machine learning machine can be implemented using general machine learning models, such as decision trees or support vector machines. In the embodiments of this invention, a decision tree is preferably used as the machine learning machine. The specific training methods for the machine learning machine are prior art and will not be elaborated upon here.

[0090] S5. Load Identification: The feature vectors of the electrical appliances in the current sequence to be identified are input into the trained load sensing model to obtain the prediction results, thus achieving load identification. That is, the structure and parameters of the model remain unchanged, and the input during model training is replaced by the feature vectors of the electrical appliances in the current sequence to be identified, instead of the historical feature vectors of the electrical appliances with appliance labels.

[0091] The non-intrusive load sensing method that integrates multiple time-scale electrical appliance features as described in S1 to S5 of the above embodiments will be applied to a specific real-world scenario to demonstrate its technical effectiveness. The specific steps are as described above, and only its specific implementation and technical effects will be shown below.

[0092] Example

[0093] To verify the effectiveness of the non-intrusive load sensing method that integrates multiple time-scale electrical appliance features proposed in this invention, this embodiment uses the PyCharm development platform and the Python 3.6 programming language to develop and implement the method. The test and verification of this embodiment were completed using a PC equipped with an Intel Xeon-X5650 2.6GHz CPU and 24GB of memory.

[0094] This invention's embodiments focus on technical verification within a residential household's appliance operation scenario. The embodiments utilize the French Controlled On / Off Load Library (COOLL) dataset and data from a location in Hangzhou, Zhejiang Province, for verification. The current sampling frequency is 6.4 kHz. The appliances included in the load library are Drill, Fan, Grinder, Hair_dryer, Hedge_trimmer, Lamp, Paint_stripper, Planer, Router, Sander, and Vacuum_cleaner.

[0095] The F-1 score is used to evaluate the model's performance. The F-1 score is the harmonic mean of recall and precision, representing the model's overall performance. Additionally, this embodiment includes a confusion matrix to visualize the classification performance.

[0096]

[0097]

[0098]

[0099] TP, TN, FP, and FN represent the number of true positive, true negative, false positive, and false negative samples, respectively. Precision and Recall are the precision and recall rates, respectively.

[0100] Because the appliances in the COOLL database lack long-term operational characteristics, Table 1 shows the classification performance (F-1 score) of all appliances in the COOLL dataset at two time scales and after ensemble. Figure 2 The image shows detailed classification performance across two time scales and after ensemble integration. (See attached image.) Figure 3 We can see the electrical feature phasors of the training samples.

[0101] Table 1. Classification performance (F-1 score) of all electrical appliances at two time scales.

[0102]

[0103]

[0104] Because the appliances in the COOLL database lack long-term operational characteristics, we selected frequency-controlled air conditioners and inverter air conditioners, which possess long-term operational characteristics, from a real-world test in Hangzhou for verification. We also ensured that these two types of appliances exhibited identical characteristics across the other two time scales. Through the extraction of long-term appliance characteristics, the construction of the perception model, and load identification, the F-1 score was 0.98. (Appendix) Figure 4 The operating current waveforms and their shapelet sequences for two types of air conditioners are shown.

[0105] Although the present invention has been described in detail through the above embodiments, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above content. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A non-intrusive load sensing method fusing multi-time scale appliance signatures, characterized in that, The method comprises the following steps: S1, periodically sampling the voltage and current at the user's home entrance, and calculating the active power effective value per second; S2, detecting whether an appliance is turned on or off or a fluctuating appliance is in operation by detecting the appliance event and fluctuation of the active power effective value per second calculated in step S1; S3, when detecting that an appliance is turned on or off or a fluctuating appliance is in operation, processing the current data sampled in step S1 on three different time scales of a short time scale, a transition process and a long time scale respectively to obtain a to-be-identified current sequence on each time scale; reading the appliance feature set in the load identification feature library, calculating the similarity between the to-be-identified current sequence on each time scale and each element in the appliance feature set on the corresponding time scale, and obtaining the appliance feature vector of the to-be-identified current sequence; S4, for each time scale, a bagging classifier with a machine learning classifier as a base learner is constructed, wherein the input of the bagging classifier is the appliance feature vector, and the output is the predicted appliance category probability; a load perception model is composed of the bagging classifiers of the three time scales, the appliance category probabilities output on the three time scales are averaged in the load perception model to obtain the final appliance category probability, and the electrical category with the maximum probability is taken as the classification to which the current input finally corresponds; the load perception model is trained using historical appliance feature vector data containing appliance category labels; S5, the appliance feature vector of the to-be-identified current sequence is input into the trained load perception model to obtain the prediction result, and load identification is realized.

2. The method of claim 1, wherein: In step S1, the voltage and current at the user's home entrance are periodically sampled at a sampling frequency f1 and an alternating current frequency f2 to form a voltage sampling sequence u1(k) and a current sampling sequence i1(k), where k is the sampling point number, the sampling frequency f1 is in the range of 0.4 kHz to 50 kHz, and the alternating current frequency f2 is 50 Hz; then the active power effective value per second is calculated to form an active sequence P1, and the active power effective value corresponding to the time t in the sequence is P1(t).

3. The method of claim 1, wherein: In step S2, the specific method of appliance event and fluctuation detection is as follows: When performing appliance event detection, the change of active power between adjacent time points is calculated as shown in formula (1): ΔP(t-1,t) = |P1(t)-P1(t-1)| (1) Where ΔP(t-1,t) is the change of active power effective value between time t and previous time t-1; if t1 satisfies the condition as shown in formula (2), it is considered that an appliance event occurs: where t1 and t2 are the start and end times of the appliance event, respectively, T f is the event mutation threshold, and T e is the event occurrence threshold. When detecting fluctuation, a sliding window G long is set with a time length of M and a sliding interval of N, and the active power range R and variance s long in the sliding window G 2 are calculated according to formula (3), and the active power range R and variance s long in the sliding window G 2 are compared with the respective threshold values R e and respectively, and when the condition shown in formula (4) is met, it is determined that an electrical appliance with power change is running at this time. wherein is the sliding window G long the average active power of M active power effective values.

4. The method of claim 3, wherein: Sliding window G long has a value of 5 minutes, and a sliding interval of 1 minute; the active power threshold R e has a value of 100, and the variance threshold has a value of 400.

5. The method of claim 1, wherein: In step S3, the short time scale is one period of the alternating current frequency f2; the transition process is 2 seconds after the appliance event occurs; and the long time scale is greater than 1 minute.

6. The method of claim 5, wherein: The current sequence to be identified on the short time scale is a current sampling sequence of 0.02 seconds, the current sequence to be identified on the transition process is the effective value of the current per cycle within 2 seconds after the occurrence of the appliance event, and the current sequence to be identified on the long time scale is a sequence of root mean square values of the current per second.

7. The method of claim 1, wherein: In step S3, the calculation of the appliance feature set of the load identification feature library is as follows: The scheme of the appliance feature is shapelet, which is defined as a subsequence capable of representing the shape feature of a time sequence; if there are M1 types of appliances, there are M2 historical current sequences collected for each type of appliance on each time scale; for any time scale, the length of the historical current sequence on the current time scale is L, and the calculation steps of the shapelet feature set on the current time scale include: ① Similarity calculation: for M2 historical current sequences W of each type of appliance, all sub-sequences V with lengths ranging from 3 to L are traversed; The distance between sequences is defined as dist(T, R), and the distance between a sub-sequence and a current sequence is subdist(T, R), and the calculation formulas are shown in formulas (5) and (6): In the formula, T and R are current sequences with a length of m, and S is a sub-sequence with a length of l in the current sequence T; For all subsequences V, the distance subdist(W, V) of the original historical current sequence W corresponding to the subsequence V is calculated as the similarity δ1 between the subsequence V and the original historical current sequence W; meanwhile, the similarity subdist(W other1 , W other2 …, W othern subdist(W other1 , V), subdist(W other2 , V), …, subdist(W othern , V) between the subsequence V and all historical current sequences W of all other category appliances is calculated, where n is the total number of all historical current sequences of all other category appliances; ② Binary information gain calculation and sorting: for each sub-sequence V, the number of distances greater than δ1 between the sub-sequence V and other original historical current sequences of the same type of appliance is a1, and the number of distances less than or equal to δ1 is a2; at the same time, the number of distances greater than δ1 between the sub-sequence V and all original historical current sequences of other types of appliances is b1, and the number of distances less than δ1 is b2; then the information gain Gain(V) of the sub-sequence V is calculated by formulas (7) and (8): All sub-sequences of each type of electrical appliance are subjected to k-means clustering, and sub-sequences with similar shapes are removed, and then the information gain thereof is sorted, and the top 5 sub-sequences with the highest information gain are taken to form a shapelet feature set J(k, m1) of the corresponding type of electrical appliance at the current time scale, where k is a time scale type sequence number, k = 1 represents a short time scale, k = 2 represents a transition process, k = 3 represents a long time scale, and m1 is a type sequence number of the electrical appliance; finally, the shapelet feature set J at the current time scale is k ={J(k, 1), J(k, 2),..., J(k, M1)}.

8. The method of claim 1, wherein: In step S3, the calculation of the appliance feature vector of the current sequence to be identified is as follows: For each time scale k, calculate the distance sim between each element of the subsequence and the shapelet feature set J corresponding to the time scale k The distance sim of each element of the subsequence k The distance sim of each element of the subsequence k The distance sim of each element of the subsequence 9. The method of claim 1, wherein: In the step S4, the number of base learners is M at each time scale when constructing and training the perception model D The samples input to each base learner during training are extracted using bootstrapping Di The maximum number of features learned by each base learner is set to k Di k Di needs to be less than 5M1; for each base learner, all training features still need to be extracted from the original features using bootstrapping 10. The method of claim 1, wherein: The machine learner adopts a decision tree or a support vector machine.

Images