Unsupervised fast load decomposition method and device based on multi-feature fusion

Abstract

The invention discloses an unsupervised fast load decomposition method and device based on multi-feature fusion, and the method comprises the steps: employing REDD data set active power data, extracting a state transition sequence (STS) through adaptive anchor point positioning, and distinguishing a load on state (RSTS) and a load off state (FSTS). A periodic fluctuation sequence is mined by using a multi-window vector, and periodic characteristics are confirmed in combination with delayed embedding and continuous coherence. Through a two-stage STS pairing algorithm, power and time distance are fused to construct a feature set. And on the basis of graph signal processing, auxiliary samples and representative samples are generated to realize unsupervised rapid clustering, and the problems of memory and calculation efficiency caused by sample size increase are solved. The method improves the accuracy and efficiency of load decomposition in a complex scene, and is suitable for non-intrusive power load monitoring.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of non-intrusive power load monitoring, and particularly relates to a multi-feature fusion-based unsupervised fast load decomposition method and device. BACKGROUND

[0002] In the field of power load monitoring, in order to promote the use of renewable energy to achieve the goal of carbon neutrality, demand side management (DSM) is a feasible and flexible method, and its decision relies on the operation state collected by load monitoring. Load monitoring has two ways of intrusive and non-intrusive (NILM). Intrusive needs to install special sensors, which has the problems of high hardware and communication cost, possible invasion of privacy and interference with normal power use of users. NILM was proposed by Hart in the 1980s, which does not interfere with the monitored equipment, has high system reliability, reduces user power consumption, protects privacy, can comprehensively monitor all target loads in the user, and is more suitable for popularization and application.

[0003] At present, the methods applied to NILM are divided into supervised learning, semi-supervised learning and unsupervised learning according to the demand for labeled data. Supervised learning can achieve ideal NILM performance when sufficient labeled data is trained. Semi-supervised learning has a lower dependence on labeled data than supervised methods. By combining a large amount of unlabeled and a small amount of labeled data, it has higher portability and acceptable performance degradation, but collecting labeled data will cause privacy problems and is time-consuming and laborious, which is contrary to the original intention of NILM. Unsupervised learning is concerned because it avoids relying on target scene labeled data. Although the current performance is slightly inferior to supervised and semi-supervised methods, the model portability and scene adaptability are very high, which can greatly improve the deployability of NILM.

[0004] In the field of unsupervised NILM methods, graph signal processing (GSP) was first applied as a key tool, combined with bilateral filter to perform physical signal level filtering and data level feature pairing, but these two methods mainly focus on load power changes, and the NILM performance is poor when distinguishing similar running power loads. Some methods use steady-state power sequences as features and use DTW distance to quantify the similarity of different steady-state power sequences, but they perform poorly when dealing with short running time loads and multiple load simultaneous running scenarios. Some methods propose STS extraction methods, which use STS as features in GSP-based clustering and matching and calculate DTW distance to quantify similarity, but the STS extraction method can only handle simple state transition processes, and the performance is limited. In addition, the above methods all face the problem that the graph size increases when the sample size increases, which further increases the memory requirement and reduces the solving rate of the closed-form solution of the graph signal. SUMMARY

[0005] To this end, the application provides a multi-feature fusion-based unsupervised fast load decomposition method and device, which solves the problems of the existing unsupervised NILM method, such as difficulty in handling complex scenes, large memory requirement and slow calculation due to easy clustering of large sample size, and poor ability of traditional methods to distinguish similar power loads.

[0006] To achieve the above-mentioned purpose, the application provides the following technical scheme: a multi-feature fusion-based unsupervised fast load decomposition method, comprising the following steps:

[0007] Constructing a state transition sequence (STS) extraction algorithm based on adaptive anchor point positioning: by setting an STS detection threshold, a maximum steady-state sequence length before and after a variable point, initializing anchor points and power change amounts, point-by-point calculating power change amounts to determine variable points and update anchor point categories, cutting candidate STS according to anchor point states, and finally normalizing STS and dividing it into opening RSTS and closing FSTS sequences according to power trends;

[0008] Periodic fluctuation sequence mining: using RSTS indexes as starting points, power sequences are intercepted through multi-window vectors and segmented according to mean values, sequences with a subsequence number satisfying a threshold are screened out, statistics of time intervals and power differences are calculated to determine candidate periodic fluctuation sequences, after disturbance is handled by extending a window, qualified periodic fluctuation sequences are screened out by combining delay embedding, persistent homology and k-means clustering, and fluctuation periods are calculated;

[0009] Two-stage STS pairing and feature set construction: in the first stage, based on steady-state power difference or DTW distance, candidate pairs are screened out from RSTS with a time stamp less than the current FSTS; in the second stage, power and time distance are fused and weighted, the best paired FSTS is found forwardly, features such as running time and transient duration are extracted, a feature set is constructed, and paired STSs are removed;

[0010] Unsupervised fast clustering: auxiliary samples are generated to enhance distinguishability, distances are calculated using steady-state power difference or DTW distance for STS features and absolute value difference for non-STS features, a comprehensive distance matrix is generated by normalization and weighting, an undirected graph and a graph Laplacian matrix are constructed, the sample categories are determined by solving the closed-form solution of the graph signal, and representative samples are generated to optimize the clustering process.

[0011] As an optimal scheme of the multi-feature fusion-based unsupervised fast load decomposition method, the STS extraction algorithm based on adaptive anchor point positioning comprises:

[0012] Step 11) For the total active power sequence, set the STS detection threshold, the maximum steady-state sequence length before and after the STS variable point; initialize the adaptive positioning anchor point and the power change amount, the history variable point index;

[0013] Step 12) calling the algorithm for each sampling point of the total active power sequence in order, calculating the power change between the current point and the next point, judging whether the absolute value of the power change is greater than the STS detection threshold, if yes, determining that the current index is a variable point, determining the anchor point category to which the time belongs by using the power change value and sign, updating the corresponding anchor point and its power change, setting the historical variable point index as the current index, storing the steady-state power sequence before the variable point and the variable point in the candidate STS; if the absolute value of the power change is less than the STS detection threshold, and the number of intercepted steady-state points is less than the maximum steady-state sequence interception length after the variable point, then the current power point is stored in the candidate STS;

[0014] Step 13) if the third anchor point is not zero, it is determined that the load is currently in frequent fluctuation, the candidate STS is intercepted by using the first and second anchor points to obtain the real STS, and the second anchor point is set as its corresponding timestamp, which is stored in the initial STS set and the index set respectively, and the corresponding parameters are updated, and the step 12) is returned;

[0015] Step 14) if the candidate STS contains enough steady-state points, the difference sequence is obtained by differentiating the candidate STS state transition part, if the absolute value of the power change in the difference sequence is less than the STS detection threshold, then it is determined that the index corresponding to the point is the fourth anchor point;

[0016] If the second anchor point is not zero and the second anchor point power change is greater than zero or the difference between the end point and the starting point of the candidate STS is greater than the STS detection threshold, the candidate STS is cut by using the second anchor point; the obtained STS and its index are stored in the initial STS and the index set respectively, and the corresponding parameters are updated;

[0017] If the candidate STS contains enough steady-state points and the fourth anchor point is not zero, the candidate STS is cut by using the fourth anchor point, and the obtained STS and its index are stored in the initial STS and the index respectively, and the corresponding parameters are updated; if the candidate STS contains enough steady-state points and only the first anchor point is not zero, the candidate STS and its index are directly stored in the initial STS and the index respectively, and the corresponding parameters are updated;

[0018] Step 15) normalizing each element in the initial STS set; trend judgment is performed on the normalized STS, if the end point is greater than the starting point power value, it is determined as RSTS representing the opening of the load, and the RSTS representing the opening of the load and the corresponding timestamp are stored respectively, otherwise it is determined as FSTS representing the closing of the load, and the FSTS representing the closing of the load and the corresponding timestamp are stored respectively.

[0019] As an optimal solution of the unsupervised fast load decomposition method based on multi-feature fusion, the periodic fluctuation sequence mining comprises:

[0020] Step 21) Given the total active power sequence, the periodic fluctuation sequence mining multi-window vector, the periodic fluctuation detection number threshold, the periodic fluctuation standard deviation threshold, the power standard deviation threshold, the power relative error threshold, the periodic fluctuation relative error threshold, the STS extraction result, the delay embedding dimension and the delay parameter, the one-dimensional coherent group norm threshold;

[0021] Step 22) For the total active power sequence, use the RSTS index as the detection starting point one by one, use the elements in the periodic fluctuation sequence mining multi-window vector to intercept the total active power sequence from large to small, and use the sequence mean to segment the intercepted sequence. If either the number of sub-sequences higher than the mean or the number of sub-sequences lower than the mean satisfies the periodic fluctuation detection number threshold, proceed to the next step of judgment; otherwise, call the next element in the multi-window vector to re-intercept. If it always does not satisfy the above judgment, continue to call step 22);

[0022] Step 23) For the sequence with sub-sequence number not less than the periodic fluctuation detection number threshold, take the sub-sequence higher than the mean as an example, calculate the time interval statistics of the sub-sequence higher than the mean, including the mean and the standard deviation, and calculate the power difference between adjacent segmented sequences and the mean and standard deviation. If the time interval and the standard deviation of the power are both less than the set threshold, it means that the current intercepted is a candidate periodic fluctuation sequence. Otherwise, return to step 22);

[0023] Step 24) Based on the result of step 23), a lengthened window of a given length is used to intercept power data from the total active power sequence. For the possible electrical disturbance in the intercepted power segment, if the relative error with the power mean in step 23) is greater than the set threshold, it is determined as a disturbance, and the electrical disturbance is set to zero. The power segment and the sequence in step 23) are spliced, and the time interval mean and standard deviation after segmentation are calculated. If the relative error of the standard deviation of this round and the last round is less than the set threshold, repeat step 24) until it no longer meets the condition, and store the intercepted candidate sequence and index information into the corresponding set, and return to step 22);

[0024] Step 25) For the candidate sequence, delay embedding is performed using the given delay embedding dimension and delay parameter to obtain the embedded data, and the coherent group is calculated using persistent homology. Then k-means clustering is performed, and the candidate sequence is labeled using the clustering result. If the time interval standard deviation of each class of labeled data and the L2 norm of the one-dimensional coherent group are less than and greater than the corresponding set threshold respectively, the candidate sequence is finally determined as a qualified periodic fluctuation sequence. The time interval mean is taken as the load fluctuation period, and the candidate sequence starting index and the fluctuation period are stored in the periodic fluctuation detection result set.

[0025] As the preferred solution of the unsupervised fast load decomposition method based on multi-feature fusion, the two-stage STS pairing and feature set construction includes:

[0026] Step 31) Given the periodic fluctuation detection result, the STS detection result, the candidate RSTS screening threshold, and the total feature set is initialized as an empty set;

[0027] Step 32) For each FSTS and its timestamp, in the first stage, the RSTSs with timestamps less than the current FSTS timestamp in the RSTS set are first screened out, and the distance between the screened RSTS and the FSTS is calculated using the steady-state power difference or the DTW distance. If the distance is less than the candidate RSTS screening threshold, the RSTS is determined as a candidate RSTS, and the RSTS and its timestamp are respectively stored in the first candidate RSTS and timestamp set;

[0028] Step 33) In the second stage, for each RSTS and its timestamp in the first candidate RSTS set, all FSTSs with timestamps greater than the RSTS timestamp and not greater than the FSTS timestamp in step 32) are found forwardly, and for each FSTS, the distance calculation is fused with power and time and respectively weighted; the FSTS with the smallest distance, i.e. the best paired FSTS with the current RSTS, is obtained. If the timestamp of the FSTS is equal to the FSTS timestamp in step 32), the RSTS and the corresponding timestamp are stored in the second candidate RSTS and timestamp set, and the distance calculated by the RSTS and the FSTS is stored in the distance vector;

[0029] Step 34) By locating the minimum value index in the distance vector, the optimal paired RSTS and its timestamp are obtained, and feature extraction is performed to construct the paired feature set; the successfully paired STS and its timestamp are removed from the STS detection result.

[0030] As the preferred solution of the unsupervised fast load decomposition method based on multi-feature fusion, the unsupervised fast clustering includes:

[0031] Step 41) Given the same class graph signal screening threshold, the total feature set;

[0032] Step 42) For each element in the total feature set, regarding it as a sample, generating auxiliary samples to enhance the sample distinguishability; first, defining a feature similarity calculation method, calculating the distance of each feature, for STS features, using the steady-state power to take the absolute value after difference or using the DTW distance, and using the method of taking the absolute value after difference for the distance calculation of the remaining features; then, using min-max normalization on each partial distance matrix; weighting and adding each partial distance matrix to obtain a comprehensive distance matrix; screening out the largest distance from the sample to be clustered as the first auxiliary sample; the second auxiliary sample is defined as a zero set; finally, reordering the sample to be classified, the representative sample and the auxiliary sample to generate a second feature set;

[0033] Step 43) For all elements in the second feature set, update the feature distance matrix, and then perform normalization and weighting and summing; adaptively calculate the proportion factor, define the vertex set and the adjacency matrix, and construct the undirected graph; calculate the diagonal matrix and the graph Laplacian matrix according to the adjacency matrix;

[0034] Step 44) Define the graph signal and calculate its global smoothness of the graph Laplacian quadratic form; initialize the graph signal and solve the closed-form solution of the optimization problem;

[0035] Step 45) Screen out the elements in the closed-form solution that are greater than the same type of graph signal screening threshold; take the cluster corresponding to the maximum value of the screened elements as the category to which the sample to be clustered belongs, and integrate the sample to be clustered into the cluster; and generate a representative sample for the cluster; if all elements in the closed-form solution are less than the graph signal screening threshold, a new category cluster is generated and the sample to be clustered is stored in the set, the total feature set is updated, and steps 42)-46) are continued until all samples are clustered.

[0036] The application also provides an unsupervised fast load decomposition device based on multi-feature fusion, comprising:

[0037] The STS extraction module is used for setting an STS detection threshold, a maximum steady-state sequence length before and after a variable point, initializing anchor points and power change amounts, calculating power change amounts point by point to determine variable points and update anchor point categories, cutting and cropping candidate STS according to anchor point states, and finally normalizing STS and dividing it into open RSTS and closed FSTS sequences according to power trends;

[0038] The periodic fluctuation sequence mining module is used for taking the RSTS index as a starting point, cutting power sequences by using a multi-window vector and segmenting them according to mean values, screening out sequences whose subsequence numbers meet a threshold, calculating statistics of time intervals and power differences to determine candidate periodic fluctuation sequences, processing disturbances by extending a window, combining delay embedding, persistent homology and k-means clustering to screen out qualified periodic fluctuation sequences and calculate fluctuation periods;

[0039] STS pairing and feature construction module, for screening candidate pairs from RSTS with time stamp less than current FSTS based on steady-state power difference or DTW distance in the first stage; the second stage fuses power and time distance and gives weight, forward finds the best paired FSTS, extracts running time, transient duration and other features and constructs feature set, and eliminates paired STS;

[0040] Unsupervised fast clustering module, for generating auxiliary sample enhancement discriminability, calculating distance by using steady-state power difference or DTW distance for STS features and absolute value difference for non-STS features, normalizing and weighting to generate comprehensive distance matrix, constructing undirected graph and graph Laplacian matrix, determining the class to which the sample belongs by solving the closed-form solution of the graph signal, and generating representative samples to optimize the clustering process.

[0041] As an unsupervised fast load decomposition device based on multi-feature fusion, the STS extraction module comprises:

[0042] Parameter initialization submodule, for total active power sequence, given STS detection threshold, maximum steady-state sequence length before and after the variable point; initialize adaptive positioning anchor point and power change, historical variable point index;

[0043] Variable point detection and anchor point updating submodule, for sequentially calling algorithm for each sampling point of total active power sequence, calculating the power change of the current point and the next point, judging whether the absolute value of the power change is greater than the STS detection threshold, if yes, determining the current index as a variable point, using the power change value and sign to determine the anchor point category to which the time belongs, and updating the corresponding anchor point and its power change, setting the historical variable point index as the current index, storing the steady-state power sequence before the variable point and the variable point in the candidate STS; if the absolute value of the power change is less than the STS detection threshold, and the number of steady-state points is less than the maximum steady-state sequence length after the variable point, then the current power point is stored in the candidate STS;

[0044] Frequent fluctuation processing submodule, for if the third anchor point is not zero, it is determined that the load is currently in frequent fluctuation, the first and second anchor points are used to cut the candidate STS to obtain the real STS, and the second anchor point is set to its corresponding time stamp, respectively stored in the initial STS set and the index set, and the corresponding parameters are updated;

[0045] a candidate STS cutting submodule, configured to, if the candidate STS contains enough steady-state points, obtain a difference sequence by differentiating the candidate STS state transition part, and determine that the index corresponding to the point is a fourth anchor point if the absolute value of the power variation of the difference sequence is less than an STS detection threshold; if the second anchor point is not zero and the power variation of the second anchor point is greater than zero or the difference between the power values of the candidate STS end point and the start point is greater than the STS detection threshold, the candidate STS is cut using the second anchor point; the obtained STS and index are respectively stored in an initial STS and index set, and corresponding parameters are updated; if the candidate STS contains enough steady-state points and the fourth anchor point is not zero, the candidate STS is cut using the fourth anchor point, and the obtained STS and index are respectively stored in an initial STS and index, and corresponding parameters are updated; if the candidate STS contains enough steady-state points and only the first anchor point is not zero, the candidate STS and index are directly stored in an initial STS and index, and corresponding parameters are updated;

[0046] an STS classification submodule, configured to normalize each element in the initial STS set; perform trend determination on the normalized STS, determine that the STS represents a load opening if the end point is greater than the start point power value, store the STS representing the load opening and the corresponding timestamp, or determine that the STS represents a load closing if the end point is less than the start point power value, and store the STS representing the load closing and the corresponding timestamp.

[0047] As an unsupervised rapid load decomposition device based on multi-feature fusion, the periodic fluctuation sequence mining module comprises:

[0048] a parameter configuration submodule, configured to give a total active power sequence, a periodic fluctuation sequence mining multi-window vector, a periodic fluctuation detection number threshold, a periodic fluctuation standard deviation threshold, a power standard deviation threshold, a power relative error threshold, a periodic fluctuation relative error threshold, an STS extraction result, a delay embedding dimension, and a delay parameter, and a one-dimensional coherent group norm threshold;

[0049] a multi-window sequence cutting submodule, configured to use the RSTS index as a detection start point, cut the total active power sequence using the elements in the periodic fluctuation sequence mining multi-window vector from large to small, and segment the cut sequence using the sequence mean value, and if either the number of sub-sequences higher than the mean value or the number of sub-sequences lower than the mean value satisfies the periodic fluctuation detection number threshold, the next step is determined; otherwise, the next element in the multi-window vector is used to cut again;

[0050] The candidate sequence preliminary screening submodule is configured to screen sequences with a number of sub-sequences not less than a periodic fluctuation detection quantity threshold value, take a sub-sequence higher than the average value as an example, calculate a time interval statistic of the sub-sequence higher than the average value, including the average value and the standard deviation, simultaneously calculate the power difference between adjacent segmented sequences, and calculate the average value and the standard deviation; if both the time interval and the power standard deviation are less than a set threshold value, it is indicated that the current intercepted sequence is a candidate periodic fluctuation sequence;

[0051] The disturbance processing and sequence splicing submodule is configured to intercept power data from the total active power sequence in a given length of an extended window, determine an electrical appliance disturbance in the intercepted power segment, if the relative error of the power average value in the candidate sequence preliminary screening submodule is greater than a set threshold value, determine the electrical appliance disturbance as a disturbance, and set the electrical appliance disturbance to zero; splice the power segment and the sequence in the candidate sequence preliminary screening submodule, continue to calculate the time interval average value and the standard deviation after segmentation, if the relative error of the standard deviation in the current round and the last round is less than a set threshold value, repeat the execution of the disturbance processing and sequence splicing submodule until the condition is no longer met, and store the intercepted candidate sequence and index information into a corresponding set;

[0052] The periodic feature confirmation submodule is configured to perform delay embedding on the candidate sequence by using a given delay embedding dimension and a delay parameter, obtain embedded data, calculate a coherence group of the embedded data by using persistent homology, then perform k-means clustering, label the candidate sequence by using the clustering result, if the time interval standard deviation of each class of labeled data and the L2 norm of the one-dimensional coherence group are respectively less than or higher than a set threshold value, finally determine that the candidate sequence is a qualified periodic fluctuation sequence; take the time interval average value as a load fluctuation period, and store the candidate sequence starting index and the fluctuation period into a periodic fluctuation detection result set.

[0053] As an unsupervised rapid load decomposition device based on multi-feature fusion, the STS pairing and feature construction module comprises:

[0054] The initial parameter and data preparation submodule is configured to initialize the total feature set as an empty set for a given periodic fluctuation detection result, an STS detection result, and a candidate RSTS screening threshold value;

[0055] The first-stage candidate RSTS screening submodule is configured to screen RSTSs with time stamps less than the time stamp of the current FSTS in the RSTS set in the first stage, calculate the distance between the screened RSTS and the FSTS by using the steady-state power difference or the DTW distance, if the distance is less than the candidate RSTS screening threshold value, determine that the RSTS is a candidate RSTS, and store the RSTS and its time stamp into the first candidate RSTS and time stamp set, respectively;

[0056] The second-stage optimal pairing search submodule is configured to, for each RSTS and its timestamp in the first candidate RSTS set, find all FSTSs in the FSTS set whose timestamps are greater than the timestamp of the RSTS and not greater than the FSTS timestamp of step 32), and for each FSTS, perform distance calculation on power and time and assign weights respectively; obtain the FSTS with the minimum distance, that is, the FSTS optimally paired with the current RSTS, and if the timestamp of the FSTS is equal to the FSTS timestamp in the first-stage candidate RSTS screening submodule, store the RSTS and the corresponding timestamp in the second candidate RSTS and timestamp set, and store the distance between the RSTS and the FSTS in the distance vector;

[0057] The feature extraction and set construction submodule is configured to obtain the optimally paired RSTS and its timestamp by locating the minimum value index in the distance vector, perform feature extraction, and construct the pairing feature set; and remove the successfully paired STS and its timestamp from the STS detection result.

[0058] As an optimal scheme of the unsupervised fast load decomposition device based on multi-feature fusion, the unsupervised fast clustering module comprises:

[0059] The clustering initialization submodule is configured to give a same-class signal filtering threshold and a total feature set.

[0060] The auxiliary sample generation and distance calculation submodule is configured to regard each element in the total feature set as a sample, generate auxiliary samples to enhance the sample distinguishability for the sample to be clustered, first define a feature similarity calculation method, calculate the distance of each feature, for the STS feature, use the absolute value after the difference or use the DTW distance, and for the remaining features, use the method of taking the absolute value after the difference to calculate the distance; then normalize each partial distance matrix; add the weighted sum of each partial distance matrix to obtain a comprehensive distance matrix; select the sample with the maximum distance from the sample to be clustered as the first auxiliary sample; the second auxiliary sample is defined as a zero set; finally, reorder the sample to be classified, the representative sample and the auxiliary sample to generate a second feature set.

[0061] The graph model construction submodule is configured to update the feature distance matrix of each element in the second feature set, and perform normalization and weighted summation again; adaptively calculate a proportion factor, define a vertex set and an adjacency matrix, and construct an undirected graph; calculate a diagonal matrix and a graph Laplacian matrix according to the adjacency matrix.

[0062] The graph signal optimization solving submodule is configured to define a graph signal, calculate the global smoothness of the graph Laplacian quadratic form of the graph signal, initialize the graph signal, and solve the closed-form solution of the graph signal of the optimization problem.

[0063] The clustering distribution and optimization submodule is used for screening elements greater than a signal screening threshold of the same type in the closed-form solution, taking the maximum value of the screened elements as a cluster to which the to-be-clustered sample belongs, integrating the to-be-clustered sample into the cluster, and generating a representative sample for the cluster, if all elements in the closed-form solution are less than the signal screening threshold, a new cluster is generated, the to-be-clustered sample is stored in the set, and the total feature set is updated.

[0064] The present application has the following advantages:

[0065] First, the state transition sequence (STS) extraction algorithm based on adaptive anchor point positioning, by analyzing the state transition process of various load opening and closing, according to the power change and its sign of the state transition process, the anchor point is adaptively determined to belong to a category, in addition to simple step type, the present application can effectively intercept various complex state transition processes and their aliasing scenes such as same direction near synchronization, reverse near synchronization and frequent fluctuation.

[0066] Second, for the case that the load operation process is accompanied by obvious fluctuation, a multi-time scale window and a moving standard deviation test are used, first, a candidate periodic fluctuation sequence is extracted, and on this basis, a sustained coherence and a delay embedding topological method are used to accurately screen out a true periodic fluctuation sequence and calculate the corresponding fluctuation period.

[0067] Third, a two-stage RSTS pairing algorithm is adopted, in the first stage, the power information is used to first screen out a candidate RSTS, in the second stage, the time information is fused to realize optimal RSTS positioning, in this process, in order to increase the pairing speed, for the STS with short transient duration, the steady-state power is used to calculate the power distance, and after pairing is completed, a corresponding multi-feature set is constructed.

[0068] Fourth, unsupervised fast clustering is performed, first, two auxiliary samples are generated to enhance the sample discrimination, in order to avoid increasing the complexity of the algorithm and increasing the scene adaptability by manually selecting a proportion factor, an adaptive proportion factor calculation method is proposed, and a graph signal clustering is used by fusing multiple features, thereby significantly improving the load monitoring accuracy. At the same time, in order to solve the problems of increasing memory occupancy and gradually reducing calculation efficiency with the increase of data volume, a representative sample is constructed, in the distance calculation and classification process, the sample is used instead of all samples in the cluster for each classified cluster, which significantly improves the clustering speed while ensuring the clustering accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0069] In order to make the technical solutions in the prior art or the embodiments of the present application clearer, the accompanying drawings needed in the description of the embodiments or the prior art will be briefly described below. Obviously, the accompanying drawings in the following description are only exemplary, and other drawings can be derived from the provided drawings without creative labor for those skilled in the art.

[0070] Figure 1 A flowchart of a multi-feature fusion-based unsupervised fast load decomposition method provided in an embodiment of the present application is shown in the figure.

[0071] Figure 2 A simple flowchart of STS detection provided in an embodiment of the present application is shown in the figure.

[0072] Figure 3 A simple flowchart of periodic fluctuation mining provided in an embodiment of the present application is shown in the figure.

[0073] Figure 4 A schematic diagram of a multi-feature fusion-based unsupervised fast load decomposition device architecture provided in an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0074] The embodiments of the present application will be described below by specific embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the description. Obviously, the described embodiments are part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0075] Embodiment 1

[0076] Referring to Figure 1 Embodiment 1 of the present application provides a multi-feature fusion-based unsupervised fast load decomposition method, which uses the active power data of the commonly used REDD (public data set) in the NILM field in the United States, and specifically includes the following steps:

[0077] Step 1) Construct a state transition sequence (STS) extraction algorithm based on adaptive anchor positioning: set the STS detection threshold, the maximum steady-state sequence length before and after the variable point, initialize the anchor point and the power change, calculate the power change point by point to determine the variable point and update the anchor point category, according to the anchor point state, cut and cut the candidate STS, and finally normalize the STS and divide it into opening RSTS and closing FSTS sequences according to the power trend;

[0078] Step 2) Periodic fluctuation sequence mining: using RSTS index as a starting point, power sequence is intercepted by multi-window vector and segmented by mean value, sequences with subsequence number meeting threshold are screened out, statistics of time interval and power difference are calculated to determine candidate periodic fluctuation sequences, after disturbance by extending window, qualified periodic fluctuation sequences are screened out by combining delay embedding, persistent homology and k-means clustering, and fluctuation period is calculated;

[0079] Step 3) Two-stage STS pairing and feature set construction: in the first stage, candidate pairs are screened out from RSTS with time stamp less than current FSTS based on steady-state power difference or DTW distance; in the second stage, power and time distance are fused and weighted, FSTS with best pairing is found forwardly, features such as running time and transient duration are extracted and feature set is constructed, and paired STS is excluded;

[0080] Step 4) Unsupervised fast clustering: auxiliary sample is generated to enhance discriminability, distance is calculated by steady-state power difference or DTW distance for STS features and absolute value difference for non-STS features, comprehensive distance matrix is generated by normalization and weighting, undirected graph and graph Laplacian matrix are constructed, sample category is determined by solving graph signal closed-form solution, representative sample is generated to optimize clustering process.

[0081] Referring to Figure 2 In the embodiment, in step 1), the state transition sequence STS extraction algorithm based on adaptive anchor positioning includes:

[0082] Step 11) For total active power sequence, given STS detection threshold, maximum steady-state sequence interception length before and after STS turning point; adaptive positioning anchor point and power change, historical turning point index are initialized.

[0083] Specifically, for the active power signal P in a given time range, heuristic parameters are determined, including STS detection threshold T, maximum steady-state sequence interception length l0 and l1 before and after STS turning point, adaptive positioning anchor points a1, a2, a3, a4 and their power changes Δp1, Δp2, Δp3, Δp4 are initialized to 0, and historical turning point index i1 is initialized to negative infinity.

[0084] Step 12) calling the algorithm in sequence for each sampling point of the total active power sequence, calculating the power change between the current point and the next point, judging whether the absolute value of the power change is greater than the STS detection threshold, if yes, determining that the current index is a variable point, determining the anchor point category to which the time belongs by using the power change value and sign, updating the corresponding anchor point and its power change, setting the historical variable point index as the current index, storing the steady-state power sequence before the variable point and the variable point into the candidate STS; if the absolute value of the power change is less than the STS detection threshold, and the number of intercepted steady-state points is less than the maximum steady-state sequence interception length after the variable point, then the current power point is stored into the candidate STS;

[0085] Specifically, in sequence for each sampling point p i in P i , calculating the power change Δp = p i+1 -p i+1 between the current power point p i and p i , judging whether the power change |Δp| is greater than the STS detection threshold T, if yes, determining that it is a variable point, and further determining the anchor point category to which i belongs by using the value of |Δp| and the sign of Δp. Specifically, if c is empty and a2 is 0, then set a1 as i-l0, store the steady-state power sequence before the variable point p and the variable point p i into the candidate STSc, and assign Δp to Δp1; if Δp1*Δp<0 and a2 is zero, then set a2 as i, and Δp2 as Δp; if Δp2*Δp<0 and a3 is zero, then set a3 as i, and Δp3 as Δp. Finally, set the historical variable point index i1 as the current index i; if |Δp| is not greater than T, and the number of intercepted steady-state points is less than l1, then store p i into c.

[0086] Step 13) if the third anchor point is not zero, then it is determined that the load is currently in frequent fluctuation, the candidate STS is intercepted by using the first and second anchor points to obtain the real STS, and the second anchor point is set as its corresponding timestamp, which is respectively stored into the initial STS set and the index set, and the corresponding parameters are updated;

[0087] Specifically, if the anchor point a3 is not zero, it is determined that the load is currently in frequent fluctuation, and c a1:a2 is intercepted as the real STS, and a2 is set as its corresponding timestamp, which is respectively stored into the initial STS set ε pr and the index set T pr , on the basis of which c = c a2:i is set, and a1, a2, a3, Δp1, Δp2, and Δp3 are updated;

[0088] Step 14) If the candidate STS contains enough steady points, then the candidate STS state transition part is differentiated to obtain a difference sequence, and if the absolute value of the power variation of the difference sequence is less than the STS detection threshold, then the point corresponding to the index is determined as the fourth anchor point;

[0089] If the second anchor point is not zero and the second anchor point power variation is greater than zero or the difference between the power values of the candidate STS end point and the start point is greater than the STS detection threshold, then the candidate STS is cut using the second anchor point; the obtained STS and its index are stored in the initial STS and index set respectively, and the corresponding parameters are updated;

[0090] If the candidate STS contains enough steady points and the fourth anchor point is not zero, then the candidate STS is cut using the fourth anchor point, and the obtained STS and its index are stored in the initial STS and index respectively, and the corresponding parameters are updated; if the candidate STS contains enough steady points and only the first anchor point is not zero, then the candidate STS and its index are directly stored in the initial STS and index respectively, and the corresponding parameters are updated;

[0091] Specifically, if the candidate STSc contains l1 steady points, then the difference between is differentiated, and if there is a sampling point in the difference sequence that satisfies |c i+1 -c i |<T, then a4 is set as c i , and Δp4 is set as c i+1 -c i . On this basis, if a2 is not zero, then according to the start point c1 and the end point c end of c, the power values and Δp1, the category to which c belongs is further determined, and if Δp1<0 or c end -c1<T, then it is determined that the candidate STSc contains two STSs. On this basis, c is cut into and two real STSs, and a2 and i are respectively taken as the corresponding time stamps thereof. Otherwise, c is determined as an STS accompanied by a spike in the opening process, and i is set as the corresponding time stamp thereof. Finally, the obtained STS and its index are respectively stored in ε pr and T pr , and a1, a2, Δp1, Δp2 are updated, and c is emptied; if the candidate STSc contains l1 steady points and a4 is not zero, then the candidate STS is cut using a4 into and , and a4 and i are respectively taken as the corresponding time stamps thereof. Finally, the obtained STS and its index are respectively stored in ε pr and , and a4, Δp4 are updated, and c is emptied; if the candidate STSc contains l1 steady points and only a1 is not zero, then c and i are directly stored in ε prWith timestamp set And update a1, Δp1, empty c;

[0092] Step 15) Normalization is performed on each element in the initial STS set; trend determination is performed on the normalized STS, if the end point is greater than the start point power value, it is determined as RSTS representing load opening, and the RSTS representing load opening and the corresponding timestamp are stored respectively, otherwise it is determined as FSTS representing load closing, and the FSTS representing load closing and the corresponding timestamp are stored respectively;

[0093] Specifically, for each element c pr In ε i , its minimum value is used for normalization:

[0094] c i,nor = c i -min(c i )

[0095] Further, trend determination is performed on the normalized STS, if the end point is greater than the start point power value, it is determined as RSTS representing load opening, and it is stored in ε R and , otherwise it is determined as FSTS representing load closing, and it is stored in ε F and :

[0096]

[0097] Referring to Figure 3 , in this embodiment, the periodic fluctuation sequence mining in step 2) includes:

[0098] Step 21) Given the total active power sequence, the periodic fluctuation sequence mining multi-window vector, the periodic fluctuation detection number threshold, the periodic fluctuation standard deviation threshold, the power standard deviation threshold, the power relative error threshold, the periodic fluctuation relative error threshold, the STS extraction result, the delay embedding dimension and the delay parameter, the one-dimensional coherent group norm threshold;

[0099] Specifically, given the total active power sequence P, the periodic fluctuation sequence mining multi-window vector w, the candidate periodic fluctuation detection threshold T num , the periodic fluctuation standard deviation threshold T t,std , the power standard deviation threshold T p,std , the power relative error threshold T p,ir , the periodic fluctuation relative error threshold T t , the STS extraction result ε R , ε F , delay embedding dimension d and delay parameter τ, one-dimensional homology group norm threshold k-means clustering cluster number k n ;

[0100] Step 22) For the total active power sequence, use the RSTS index as the detection starting point one by one, use the periodic fluctuation sequence from large to small to cut the elements in the multi-window vector to intercept the total active power sequence, and segment the intercepted sequence using the sequence mean. If either the number of subsequences higher than the mean or the number of subsequences lower than the mean satisfies the periodic fluctuation detection number threshold, proceed to the next step of judgment; otherwise, call the next element in the multi-window vector to re-cut, and if it always does not satisfy the above judgment, continue to call step 22);

[0101] Specifically, for the active power sequence P, use the elements in w as the detection starting point one by one, use the elements in w from large to small to cut the active power sequence i and segment the using the sequence mean For the number of subsequences n a higher than the mean and the number of subsequences n b lower than the mean, if either n a or n a satisfies the periodic fluctuation detection number threshold T num , proceed to the next step of judgment. Otherwise, call the next element in w to cut, and if none of them satisfies the above judgment, continue to call step 22) until the condition is met.

[0102] Step 23) For the sequence whose number of subsequences is not less than the periodic fluctuation detection number threshold, take the subsequence higher than the mean as an example, calculate the time interval statistics of the subsequence higher than the mean, including the mean and the standard deviation, and calculate the power difference between adjacent segmented sequences and the mean and standard deviation. If both the time interval and the power standard deviation are less than the set threshold, it means that the current intercepted is a candidate periodic fluctuation sequence. Otherwise, return to step 22);

[0103] Specifically, for the subsequence whose number of subsequences is not less than T num , take the subsequence seq a higher than the mean as an example, further calculate its time interval statistics, including the mean and the standard deviation, and the results are represented as and Calculate the power difference between adjacent seq a and seq b , and calculate the mean and standard deviation If both and are less than the set threshold T​​​p,std With T t,std If the current intercepted sequence is a candidate periodic fluctuation sequence, then the process returns to step 22).

[0104] Step 24) Based on the result of step 23), power data is intercepted from the total active power sequence using an extended window of a given length. For any electrical disturbance that may appear in the intercepted power segment, if the relative error with the power mean in step 23) is greater than a set threshold, it is determined to be a disturbance, and the electrical disturbance is set to zero; the power segment is spliced ​​with the sequence in step 23), and the mean and standard deviation of the time interval after segmentation are continuously calculated. If the relative error of the standard deviation between the current round and the previous round is less than the set threshold, step 24) is repeated until the condition is no longer met, and the intercepted candidate sequence and index information are stored in the corresponding set, and the process returns to step 22);

[0105] Specifically, the length The extended window captures power data p from P sub , for p sub Other electrical disturbances that may occur in ir For example), if μ p,seq The relative error is greater than T p,ir It is determined to be a disturbance and set to zero. With p sub For p candidate , continue to calculate p candidate Mean and standard deviation of time intervals after segmentation and like With the previous round The relative error is less than T t , then repeat step 24) until the condition is no longer met, and store the intercepted candidate sequence and index information in C, and return to step 22).

[0106] Step 25) Delay embedding is performed on the candidate sequence using the given delay embedding dimension and delay parameter to obtain the embedded data, and its homology group is calculated using continuous homology. K-means clustering is then performed, and the candidate sequence is labeled using the clustering results. If the time interval standard deviation of each type of labeled data and the L2 norm of the one-dimensional homology group are respectively less than and greater than the corresponding set thresholds, the candidate sequence is finally determined to be a qualified periodic fluctuation sequence; the mean time interval is used as the load fluctuation period, and the candidate sequence starting index and fluctuation period are stored in the periodic fluctuation detection result set;

[0107] Specifically, for the elements in C, that is, each candidate sequence seq candidate, use the given d and τ to perform delayed embedding, get the embedded data E, and use the continuous homology to calculate its homology group, then perform k-means clustering, and use the clustering results to classify seq candidate For labeling, if the standard deviation of the time interval of each type of labeled data is The L2 norm of the one-dimensional homology group is less than the set threshold T t,std , higher than The final judgment seq candidate is a qualified periodic fluctuation sequence. as the fluctuation period, and its starting index and Store the periodic fluctuation detection result set middle.

[0108] In this embodiment, the two-stage STS pairing and feature set construction in step 3) includes:

[0109] Step 31) Given the periodic fluctuation detection results, STS detection results, and candidate RSTS screening threshold, the total feature set is initialized to an empty set;

[0110] Specifically, given the periodic fluctuation detection results STS test results ε R , ε F , Candidate RSTS screening threshold th a , initialize the total feature set Is an empty set.

[0111] Step 32) For each FSTS and its timestamp, in the first stage, first screen out the RSTS in the RSTS set whose timestamp is less than the timestamp of the current FSTS, and calculate the distance between the screened RSTS and the FSTS using the steady-state power difference or DTW distance. If the distance is less than the candidate RSTS screening threshold, the RSTS is determined to be a candidate RSTS, and the RSTS and its timestamp are respectively stored in the first candidate RSTS and timestamp set;

[0112] Specifically, for ε F Each FSTS and its timestamp c F With t F In the first stage, first filter out ε R The timestamp is less than t F RSTS, and use the steady-state power or DTW distance to calculate the filtered RSTS c one by one R,m with c F Distance:

[0113]

[0114] If the distance is less than the given threshold th a , determine that the RSTS is a candidate RSTS and store the candidate RSTS and its timestamp in the first candidate RSTS and timestamp set respectively and

[0115] In the second stage of step 33), for each RSTS and its timestamp in the first candidate RSTS set, a forward search is performed on all FSTSs in the FSTS set whose timestamps are greater than the timestamp of the RSTS and not greater than the timestamp of the FSTS in step 32). For each FSTS, the distance is calculated by fusing power and time and weighted respectively. The FSTS with the smallest distance, i.e., the best pairing with the current RSTS, is obtained. If the timestamp of the FSTS is equal to the timestamp of the FSTS in step 32), the RSTS and the corresponding timestamp are stored in the second candidate RSTS and timestamp set, and the distance calculated between the RSTS and the FSTS is stored in the distance vector.

[0116] Specifically, in the second phase, Each RSTS and its timestamp Forward search for ε F All timestamps greater than and less than or equal to t F FSTS, and for each FSTS and its timestamp Fusion power and time distance and assign weights α and β respectively to represent the importance of the two features:

[0117]

[0118] The smallest distance is obtained The best matching FSTS, if the timestamp of the FSTS is equal to t F , then store the RSTS and its timestamp in the second candidate RSTS and its timestamp set and store the corresponding distance in the distance vector d.

[0119] Step 34) The optimal paired RSTS and its timestamp are obtained by locating the minimum index in the distance vector, and feature extraction is performed to construct the pairing feature set; the successfully paired STS and its timestamp are removed from the STS detection results;

[0120] Specifically, by locating the minimum index x in d, the optimal pairing RSTS is obtained and its timestamp And perform feature extraction:

[0121]

[0122] On this basis, the pairing feature set is constructed t op Indicates the load running time, l mean represents the average transient duration of RSTS and FSTS, p cy is the fluctuation period corresponding to the pairing, according to t F and From the periodic fluctuation detection results Get, if t F , If both are located in a certain periodic fluctuation segment, the periodic information of the pair is the fluctuation period of the fluctuation segment. Otherwise the feature is zero.

[0123] Further, the successfully paired RSTS FSTS c F and its timestamp t F From ε R , ε F , Eliminate from.

[0124] In this embodiment, the unsupervised fast clustering in step 4) includes:

[0125] Step 41) Given a similar graph signal screening threshold and a total feature set;

[0126] Specifically, given the similar graph signal screening threshold q and the total feature set F;

[0127] Step 42) Each element in the total feature set is considered as a sample, and auxiliary samples are generated for clustering samples to enhance sample discrimination. First, a feature similarity calculation method is defined, and the distance is calculated for each feature. For STS features, the steady-state power is first subtracted and then the absolute value is taken, or the DTW distance is used. For other features, the distance is calculated by first subtracting and then taking the absolute value. Subsequently, min-max normalization is performed on each sub-item distance matrix. Weighted summation is performed on each sub-item distance matrix to obtain a comprehensive distance matrix. The sample with the largest distance to the sample to be clustered is selected as the first auxiliary sample. The second auxiliary sample is defined as an all-zero set. Finally, the sample to be classified, the representative sample, and the auxiliary samples are reordered to generate a second feature set.

[0128] Specifically, yes Each element in is considered as a sample. If J-1 samples have been clustered and clustered into K load category clusters i∈[1,K], each load category cluster has a representative sample, i∈[1,K]. Treat cluster samples Generate auxiliary samples to enhance sample differentiation. First, define the feature similarity calculation method for different samples. and The distance is calculated for each feature. For the first two STS features, the distance is calculated using dist1(·). For the remaining features, the distance is calculated using the following formula:

[0129]

[0130] Then we can get the feature distance matrix D k , k = 1, 2, 3, 4, 5. Then, each item distance matrix is ​​normalized to the interval [0, 1] using min-max:

[0131]

[0132] Furthermore, each item distance matrix is ​​weighted and summed

[0133] D con =α1×D 1,nor +α2×D 2,nor +α3×D 3,nor +α4×D 4,nor +α5×D 5,nor ,

[0134] α1+α2+α3+α4+α5=1,

[0135] On this basis, we screened The largest value, that is, The sample with the lowest similarity is used as the first auxiliary sample Furthermore, the second auxiliary sample is defined as:

[0136]

[0137] Finally, the samples to be classified Reorder the representative samples and auxiliary samples to obtain the second feature set:

[0138]

[0139] Step 43) For all elements in the second feature set, update each feature distance matrix, and then perform normalization and weighting before summing; adaptively calculate the scaling factor, define the vertex set and adjacency matrix, and construct an undirected graph; calculate the diagonal matrix and the graph Laplacian matrix based on the adjacency matrix;

[0140] Specifically, for the second feature set All elements in, update each feature distance matrix D k , k = 1, 2, 3, 4, 5 and perform weighted summation in normalization to obtain D AOn this basis, construct an undirected graph V is the vertex set, A is the adjacency matrix, and its elements are expressed as:

[0141]

[0142] where ρ is the scaling factor and is updated automatically at each classification:

[0143]

[0144] Where ξ is a heuristic parameter close to zero, and ρ is the upper bound of the result of the above formula. Then, the diagonal matrix D and the graph Laplacian matrix L are calculated based on the adjacency matrix A:

[0145]

[0146] Step 44) defining the graph signal and calculating the global smoothness of its graph Laplace quadratic form; initializing the graph signal and finding a closed-form solution to the optimization problem;

[0147] Specifically, define the graph signal s and calculate the global smoothness of its graph Laplace quadratic form:

[0148]

[0149] On this basis, initialize the graph signal s and solve the optimization problem The closed-form solution s*. First, initialize the graph signal:

[0150]

[0151] Finally, the optimization problem can be obtained Closed-form solution of the graph signal:

[0152]

[0153] Step 45) Filter out the elements in the closed-form solution that are greater than the similar graph signal screening threshold; take the cluster corresponding to the maximum value of the screened elements as the category of the sample to be clustered, merge the sample to be clustered into the cluster; and generate a representative sample for the cluster; if all elements in the closed-form solution are less than the graph signal screening threshold, then generate a new category cluster and store the sample to be clustered into the cluster, update the total feature set, and continue to perform steps 42) to 46) until all samples are clustered;

[0154] Specifically, filter out * Elements s greater than q in above . Further, s above The load category cluster represented by the largest element As samples to be clustered The category to which it belongs and incorporate the cluster; on this basis, a representative sample is generated for the cluster its elements is the mean value of the corresponding feature of all samples in the category. By generating a representative sample, the sample is used instead of all samples in the cluster in the subsequent distance calculation and classification process, thereby effectively avoiding the problems of increasing memory occupation and gradually reducing calculation efficiency as the amount of data increases.

[0155] The specific steps of implementing non-intrusive load decomposition by using the method of the application are as follows:

[0156] 1) Obtain total power data of the target scene. The practicability of the application is verified by using the REDD dataset of the United States. House1 and House2 of the REDD dataset are selected as two verification scenes, which are named as Experiment 1 and Experiment 2 respectively. The start and end times of the data are selected as shown in the following table.

[0157] Table 1 Selection of experimental data

[0158]

[0159] 2) Index selection. In the application, F1 score is selected as the classification index; and the evaluation indexes of the load decomposition result include normalized signal total error (SAE), mean absolute error (MAE) and decomposition accuracy (Acc.). The calculation methods of the indexes are as follows:

[0160] PR = TP / (TP+FP),

[0161] RE = TP / (TP+FN),

[0162] F1 = 2 x PR x RE / (PR+RE),

[0163]

[0164] wherein TP is the number of positive samples correctly identified as positive samples, FP is the number of negative samples incorrectly identified as positive samples, and FN is the number of positive samples incorrectly identified as negative samples. PR is the accuracy, and RE is the recall. The harmonic mean F1 balances PR and RE, and reflects the accuracy based on the identification result and the true situation respectively. Therefore, the closer F1 is to 1, the better the identification performance is. is the actual consumption power of the load i at time j, represents the decomposition power of the load i at time j, and N is the length of the power signal. Obviously, the closer Acc. is to 1, the lower MAE is, indicating that the decomposition power is closer to the true value. The lower SAE is, the closer the predicted total energy consumption is to the true total energy consumption.

[0165] 3) Comparative method. Select the method in the prior art as the comparative method. And respectively marked as P-UGSP, SS-UGSP, and STS-UGSP.

[0166] 4) The index results of the target data set load decomposition by the method of the present application are shown in Tables 2-5.

[0167] Table 2 F1 index results (the bold in each row represents the optimal result for the appliance)

[0168]

[0169] Table 3 SAE index results (the bold in each row represents the optimal result for the appliance)

[0170]

[0171] Table 4 MAE index results (the bold in each row represents the optimal result for the appliance)

[0172]

[0173]

[0174] Table 5 Acc. index results (the bold in each row represents the optimal result for the appliance)

[0175]

[0176] From Tables 2-5, it can be seen that the method of the present application can be used to perform non-intrusive load decomposition in the target data set without labels, and the decomposition effect of most of the loads in the table is better than that of other comparative methods.

[0177] 5) The clustering process time consumption of the target data set by the method of the present application and the clustering process time consumption of the comparative method are shown in Table 6.

[0178] Table 6 Clustering process time consumption comparison (the bold in each row represents the shortest clustering time of the experiment)

[0179]

[0180] From Table 6, it can be seen that the present application has obvious advantages in clustering speed compared with other methods. From the above research materials, it can be seen that the method of the present application can perform fast and accurate load decomposition in the target scene without labels through the steps of fusing multiple features and constructing auxiliary samples, representative samples, etc., realize unsupervised NILM for the target data set, and is superior to the current unsupervised NILM algorithm. For real scenes, it has great application significance and high load decomposition potential.

[0181] Example 2

[0182] Referring to Figure 4 The embodiment 2 also provides a device for unsupervised fast load decomposition based on multi-feature fusion, comprising:

[0183] The STS extraction module 100 is configured to set an STS detection threshold, a maximum steady-state sequence length before and after a change point, initialize anchor points and power variation, calculate power variation point by point to determine a change point and update anchor point categories, cut and clip candidate STS according to anchor point states, and finally normalize STS and divide it into open RSTS and closed FSTS sequences according to power trends;

[0184] The periodic fluctuation sequence mining module 200 is configured to use RSTS indexes as starting points, cut power sequences by multi-window vectors and segment them according to mean values, filter sequences with a number of sub-sequences satisfying a threshold, calculate statistics of time intervals and power differences to determine candidate periodic fluctuation sequences, process disturbances by extending windows, filter qualified periodic fluctuation sequences by combining delay embedding, persistent homology and k-means clustering, and calculate fluctuation periods.

[0185] The STS pairing and feature construction module 300 is configured to filter candidate pairs from RSTS with time stamps less than a current FSTS based on steady-state power differences or DTW distances in a first stage, filter FSTS with the best pair in a forward direction by fusing power and time distances and giving weights in a second stage, extract features such as running time and transient duration, construct a feature set, and eliminate paired STS.

[0186] The unsupervised fast clustering module 400 is configured to generate auxiliary samples to enhance distinguishability, calculate distances by using steady-state power differences or DTW distances for STS features and absolute value differences for non-STS features, normalize and give weights to generate a comprehensive distance matrix, construct an undirected graph and a graph Laplacian matrix, determine categories to which samples belong by solving a closed-form solution of a graph signal, and generate representative samples to optimize a clustering process.

[0187] In the embodiment, the STS extraction module 100 comprises:

[0188] The parameter initialization submodule 101 is configured to set an STS detection threshold, a maximum steady-state sequence length before and after a change point for a total active power sequence, initialize adaptive anchor points and power variation, and a historical change point index.

[0189] The variable point detection and anchor point updating submodule 102 is configured to sequentially call an algorithm for each sampling point of the total active power sequence, calculate a power variation between a current point and a next point, determine whether an absolute value of the power variation is greater than an STS detection threshold, if yes, determine that a current index is a variable point, determine an anchor point category to which the time point belongs by using a power variation value and a sign, update a corresponding anchor point and the power variation, set a historical variable point index as the current index, store a steady-state power sequence before the variable point into the candidate STS, and store the variable point into the candidate STS; if the absolute value of the power variation is less than the STS detection threshold, and a number of intercepted steady-state points is less than a maximum steady-state sequence interception length after the variable point, then store the current power point into the candidate STS.

[0190] The frequent fluctuation processing submodule 103 is configured to, if the third anchor point is not zero, determine that the load is currently in frequent fluctuation, intercept the candidate STS by using the first anchor point and the second anchor point to obtain a real STS, set the second anchor point as a corresponding time stamp, store the real STS and the index set into an initial STS set and an index set respectively, and update corresponding parameters.

[0191] The candidate STS cutting submodule 104 is configured to, if the candidate STS contains enough steady-state points, perform difference on a state transition part of the candidate STS to obtain a difference sequence, if an absolute value of a power variation in the difference sequence is less than the STS detection threshold, determine that a corresponding index of the point is the fourth anchor point; if the second anchor point is not zero, and a power variation of the second anchor point is greater than zero or a difference between a power value of a start point and a power value of an end point of the candidate STS is greater than the STS detection threshold, cut the candidate STS by using the second anchor point; store the obtained STS and the index into the initial STS and the index set respectively, and update corresponding parameters; if the candidate STS contains enough steady-state points, and the fourth anchor point is not zero, cut the candidate STS by using the fourth anchor point, store the obtained STS and the index into the initial STS and the index set respectively, and update corresponding parameters; if the candidate STS contains enough steady-state points, and only the first anchor point is not zero, directly store the candidate STS and the index into the initial STS and the index set respectively, and update corresponding parameters.

[0192] The STS classification submodule 105 is configured to normalize each element in the initial STS set; perform trend determination on the normalized STS, if an end point is greater than a start point power value, determine that the STS is an RSTS representing load opening, store the RSTS and a corresponding time stamp respectively, otherwise, determine that the STS is an FSTS representing load closing, and store the FSTS and a corresponding time stamp respectively.

[0193] In the embodiment, the periodic fluctuation sequence mining module 200 includes:

[0194] The parameter configuration sub-module 201 is configured to, given a total active power sequence, mine a multi-window vector from the periodic fluctuation sequence, set a periodic fluctuation detection quantity threshold, set a periodic fluctuation standard deviation threshold, set a power standard deviation threshold, set a power relative error threshold, set a periodic fluctuation relative error threshold, set an STS extraction result, set a delay embedding dimension and a delay parameter, and set a one-dimensional coherent group norm threshold.

[0195] The multi-window sequence intercepting sub-module 202 is configured to, for the total active power sequence, intercept the total active power sequence by using each element in the periodic fluctuation sequence mining multi-window vector as a detection starting point in a descending order, and segment the intercepted sequence by using a sequence mean value. If either the number of sub-sequences higher than the mean value or the number of sub-sequences lower than the mean value satisfies the periodic fluctuation detection quantity threshold, the next step is determined. Otherwise, the next element in the multi-window vector is used to re-intercept.

[0196] The candidate sequence preliminary screening sub-module 203 is configured to, for the sequence with a sub-sequence number not lower than the periodic fluctuation detection quantity threshold, take the sub-sequence higher than the mean value as an example, calculate a time interval statistic of the sub-sequence higher than the mean value, including a mean value and a standard deviation, calculate a power difference between adjacent segmented sequences, and calculate a mean value and a standard deviation. If the time interval and the power standard deviation are both less than a set threshold, it is indicated that the current intercepted sequence is a candidate periodic fluctuation sequence.

[0197] The disturbance processing and sequence splicing sub-module 204 is configured to intercept power data from the total active power sequence by using an extended window with a given length, determine an electrical appliance disturbance in the intercepted power segment, and set the electrical appliance disturbance to zero if the relative error of the power mean value in the candidate sequence preliminary screening sub-module is greater than a set threshold. The power segment and the sequence in the candidate sequence preliminary screening sub-module are spliced, and the time interval mean value and the standard deviation after segmentation are continuously calculated. If the relative error of the standard deviation between this round and the last round is less than a set threshold, the disturbance processing and sequence splicing sub-module is repeatedly executed until the condition is no longer met, and the intercepted candidate sequence and index information are stored in a corresponding set.

[0198] The periodic feature confirmation sub-module 205 is configured to, for the candidate sequence, perform delay embedding by using a given delay embedding dimension and a delay parameter, obtain embedded data, calculate a coherent group of the embedded data by using persistent homology, perform k-means clustering, label the candidate sequence by using the clustering result, and determine the candidate sequence as a qualified periodic fluctuation sequence if the time interval standard deviation of each class of labeled data and the L2 norm of the one-dimensional coherent group are less than and higher than corresponding set thresholds, respectively. The time interval mean value is taken as a load fluctuation period, and the candidate sequence starting index and the fluctuation period are stored in a periodic fluctuation detection result set.

[0199] In this embodiment, the STS pairing and feature construction module 300 comprises:

[0200] An initial parameter and data preparation submodule 301 is configured to initialize the total feature set as an empty set, given the periodic fluctuation detection result, the STS detection result, and the candidate RSTS screening threshold.

[0201] A first-stage candidate RSTS screening submodule 302 is configured to screen out, for each FSTS and its timestamp, the RSTSs with timestamps smaller than the current FSTS timestamp in the RSTS set in the first stage, and calculate the distance between the screened RSTS and the FSTS using the steady-state power difference or the DTW distance. If the distance is smaller than the candidate RSTS screening threshold, the RSTS is determined as a candidate RSTS, and the RSTS and its timestamp are respectively stored in the first candidate RSTS and timestamp set.

[0202] A second-stage optimal pairing search submodule 303 is configured to find, for each RSTS and its timestamp in the first candidate RSTS set, all FSTSs with timestamps greater than the RSTS timestamp and not greater than the FSTS timestamp in step 32), and perform distance calculation and weighting for each FSTS. The FSTS with the minimum distance, i.e., the best pairing FSTS of the current RSTS, is obtained. If the timestamp of the FSTS is equal to the FSTS timestamp in the first-stage candidate RSTS screening submodule, the RSTS and the corresponding timestamp are stored in the second candidate RSTS and timestamp set, and the distance between the RSTS and the FSTS is stored in the distance vector.

[0203] A feature extraction and set construction submodule 304 is configured to obtain the optimal pairing RSTS and its timestamp by locating the minimum value index in the distance vector, perform feature extraction, and construct the pairing feature set. The successfully paired STS and its timestamp are removed from the STS detection result.

[0204] In this embodiment, the unsupervised fast clustering module 400 comprises:

[0205] A clustering initialization submodule 401 is configured to initialize the total feature set, given the same-class signal screening threshold.

[0206] The auxiliary sample generation and distance calculation submodule 402 is configured to generate auxiliary samples to enhance the sample distinguishability for each element in the total feature set, which is regarded as a sample, and for the sample to be clustered. First, a feature similarity calculation method is defined, and the distance of each feature is calculated. For the STS feature, the DTW distance is used after the absolute value is taken after the difference is made. For the rest of the features, the distance is calculated by taking the absolute value after the difference is made. Then, the min-max normalization is performed on each partial distance matrix. The weighted sum of each partial distance matrix is obtained to obtain a comprehensive distance matrix. The sample with the largest distance from the sample to be clustered is selected as the first auxiliary sample. The second auxiliary sample is defined as a zero set. Finally, the sample to be classified, the representative sample and the auxiliary sample are reordered to generate a second feature set.

[0207] The graph model construction submodule 403 is configured to update the feature distance matrix of each element in the second feature set, and perform normalization and weighting after summing. The proportion factor is adaptively calculated, the vertex set and the adjacency matrix are defined, and the undirected graph is constructed. The diagonal matrix and the graph Laplacian matrix are calculated according to the adjacency matrix.

[0208] The graph signal optimization solving submodule 404 is configured to define the graph signal, calculate the global smoothness of the graph Laplacian quadratic form, initialize the graph signal, and obtain the closed-form solution of the optimization problem.

[0209] The clustering assignment and optimization submodule 405 is configured to screen out elements greater than the same type of graph signal screening threshold in the closed-form solution. The cluster corresponding to the maximum value of the screened elements is taken as the category to which the sample to be clustered belongs, and the sample to be clustered is integrated into the cluster. The representative sample of the cluster is generated. If all elements in the closed-form solution are less than the graph signal screening threshold, a new category cluster is generated, and the sample to be clustered is stored in the set. The total feature set is updated.

[0210] It should be noted that the information interaction, execution process and the like between the modules of the above apparatus are based on the same concept as the method embodiment in Embodiment 1 of the present application, and the technical effects brought by the method embodiment are the same as those of the method embodiment. For specific content, refer to the description of the method embodiment in the foregoing method embodiment of the present application, which will not be repeated here.

[0211] Embodiment 3

[0212] Embodiment 3 of the present application provides a non-transitory computer readable storage medium, which stores a program code of an unsupervised fast load decomposition method based on multi-feature fusion. The program code includes instructions for executing the unsupervised fast load decomposition method based on multi-feature fusion of embodiment 1 or any possible implementation manner thereof.

[0213] The computer-readable storage medium can be any available media or a set of one or more available media that is accessible by a computer, data storage device, server, data center, etc. with which the storage medium is integrated. The available media can be a magnetic medium, (e.g., a floppy diskette, a hard disk drive, a magnetic tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a Solid State Disk, SSD), etc.

[0214] Embodiment 4

[0215] Embodiment 4 of the present application provides an electronic device, comprising a memory and a processor;

[0216] The processor and the memory complete the communication between each other through a bus; the memory stores program instructions that can be executed by the processor, and the processor calling the program instructions can execute the multi-feature fusion based unsupervised fast load decomposition method of embodiment 1 or any possible implementation manner thereof.

[0217] Specifically, the processor can be implemented by hardware or software, when implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor, which realizes by reading software codes stored in a memory, the memory can be integrated in the processor, or located outside the processor and exist independently.

[0218] In the above embodiments, all or part of them can be realized by software, hardware, firmware or any combination thereof. When realized by software, all or part of them can be realized in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another, for example, the computer instructions can be transferred from one website, computer, server or data center to another through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode.

[0219] It should be apparent to those skilled in the art that the modules or steps of the application described above can be implemented with a general purpose computing device, which can be centralized on a single computing device or distributed across a network of multiple computing devices, and optionally implemented with program code executable by a computing device, which can be stored in a storage device and executed by a computing device, and in some cases, the steps shown or described can be performed in an order different from that shown, or made into individual integrated circuit modules, or multiple modules or steps made into a single integrated circuit module. Thus, the application is not limited to any particular combination of hardware and software.

[0220] While the application has been described in detail and with reference to specific embodiments thereof, it will be apparent to one skilled in the art that various modifications or changes can be made therein without departing from the spirit and scope of the application. Accordingly, it is intended that all such modifications and changes be included within the scope of the application as claimed.

Claims

1. Unsupervised fast load decomposition method based on multi-feature fusion, characterized by: The following steps are involved: A state transition sequence (STS) extraction algorithm based on adaptive anchor point positioning was constructed. The algorithm set the STS detection threshold and the maximum steady-state sequence interception length before and after the change point, initialized the anchor point and power change, calculated the power change point by point to determine the change point and update the anchor point category, intercepted and cropped candidate STSs based on the anchor point state, and finally normalized the STSs and classified them into open RSTS and closed FSTS sequences based on power trends. Periodic fluctuation sequence mining: Using the RSTS index as a starting point, the power sequence is intercepted through a multi-window vector and segmented by mean. Sequences with a number of subsequences that meet the threshold are screened out. Time intervals and power difference statistics are calculated to identify candidate periodic fluctuation sequences. After processing disturbances by extending the window, qualified periodic fluctuation sequences are screened out and the fluctuation period is calculated by combining delayed embedding, continuous homology, and k-means clustering. Two-stage STS pairing and feature set construction: The first stage screens candidate pairs from RSTSs whose timestamps are smaller than the current FSTS based on steady-state power difference or DTW distance. The second stage fuses power and time distance and assigns weights to forward search for the best paired FSTS. Features such as run time and transient duration are extracted and constructed into a feature set, eliminating already paired STSs. Unsupervised fast clustering: Generate auxiliary samples to enhance discrimination, use steady-state power difference or DTW distance for STS features, and absolute value difference for non-STS features to calculate distance, normalize and weight them to generate a comprehensive distance matrix, construct an undirected graph and graph Laplacian matrix, determine the sample category by solving the closed-form solution of the graph signal, and generate representative samples to optimize the clustering process.

2. The unsupervised rapid load decomposition method based on multi-feature fusion according to claim 1 is characterized in that: The construction of the state transition sequence STS extraction algorithm based on adaptive anchor point positioning includes: Step 11) For the total active power sequence, the STS detection threshold and the maximum steady-state sequence intercept length before and after the STS change point are given; the adaptive positioning anchor point, power change amount, and historical change point index are initialized; Step 12) Call the algorithm for each sampling point of the total active power sequence in sequence, calculate the power change between the current point and the next point, and determine whether the absolute value of the power change is greater than the STS detection threshold. If so, determine that the current index is a change point, use the power change value and sign to determine the anchor point category at this moment, update the corresponding anchor point and its power change, set the historical change point index as the current index, and store the steady-state power sequence before the change point and the change point in the candidate STS; if the absolute value of the power change is less than the STS detection threshold, and the number of intercepted steady-state points is less than the maximum steady-state sequence interception length after the change point, then store the current power point in the candidate STS; Step 13) If the third anchor point is not zero, it is determined that the load is currently fluctuating frequently. The candidate STS is intercepted using the first and second anchor points to obtain the real STS, and the second anchor point is set as its corresponding timestamp. The STS is stored in the initial STS set and the index set respectively, and the corresponding parameters are updated. Return to step 12); Step 14) If the candidate STS contains enough steady-state points, the state transition portion of the candidate STS is differentiated to obtain a differential sequence. If the absolute value of the power change in the differential sequence is less than the STS detection threshold, the index corresponding to the point is determined to be the fourth anchor point. If the second anchor point is not zero, and the power change of the second anchor point is greater than zero or the power difference between the candidate STS endpoint and the starting point is greater than the STS detection threshold, the candidate STS is trimmed using the second anchor point; the obtained STS and its index are stored in the initial STS and index set respectively, and the corresponding parameters are updated; If the candidate STS already contains enough stable points and the fourth anchor point is not zero, the candidate STS is cropped using the fourth anchor point, the resulting STS and its index are stored in the initial STS and index respectively, and the corresponding parameters are updated; if the candidate STS already contains enough stable points and only the first anchor point is not zero, the candidate STS and its index are directly stored in the initial STS and index respectively, and the corresponding parameters are updated; Step 15) Normalize each element in the initial STS set; perform trend determination on the normalized STS. If the end point power value is greater than the starting point power value, it is determined to be an RSTS indicating that the load is turned on, and the RSTS indicating that the load is turned on and the corresponding timestamp are stored separately. Otherwise, it is determined to be an FSTS indicating that the load is turned off, and the FSTS indicating that the load is turned off and the corresponding timestamp are stored separately.

3. The unsupervised rapid load decomposition method based on multi-feature fusion according to claim 1 is characterized in that: The periodic fluctuation sequence mining includes: Step 21) Given the total active power sequence, the periodic fluctuation sequence mining multi-window vector, the periodic fluctuation detection number threshold, the periodic fluctuation standard deviation threshold, the power standard deviation threshold, the power relative error threshold, the periodic fluctuation relative error threshold, the STS extraction result, the delay embedding dimension and delay parameter, and the one-dimensional homology group norm threshold; Step 22) For the total active power sequence, use the RSTS index as the detection starting point one by one, and use the periodic fluctuation sequence to mine the elements in the multi-window vector from large to small to intercept the total active power sequence, and use the sequence mean to segment the intercepted sequence. If either the number of subsequences above the mean or the number of subsequences below the mean satisfies or exceeds the periodic fluctuation detection number threshold, then proceed to the next step of judgment; otherwise, call the next element in the multi-window vector to re-interpret. If the above judgment is still not met, then continue to call step 22); Step 23) For sequences whose number of subsequences is not less than the periodic fluctuation detection threshold, taking the subsequences above the mean as an example, calculate the time interval statistics of the subsequences above the mean, including the mean and standard deviation. At the same time, calculate the power difference between adjacent segmented sequences, and calculate the mean and standard deviation. If the time interval and the standard deviation of the power are both less than the set threshold, it means that the currently intercepted sequence is a candidate periodic fluctuation sequence; otherwise, return to step 22). Step 24) Based on the result of step 23), power data is intercepted from the total active power sequence using an extended window of a given length. For any electrical disturbance that may appear in the intercepted power segment, if the relative error with the power mean in step 23) is greater than a set threshold, it is determined to be a disturbance, and the electrical disturbance is set to zero; the power segment is spliced ​​with the sequence in step 23), and the mean and standard deviation of the time interval after segmentation are continuously calculated. If the relative error of the standard deviation between the current round and the previous round is less than the set threshold, step 24) is repeated until the condition is no longer met, and the intercepted candidate sequence and index information are stored in the corresponding set, and the process returns to step 22); Step 25) For the candidate sequence, delay embedding is performed using the given delay embedding dimension and delay parameter to obtain the embedded data, and its homology group is calculated using continuous homology. Then, k-means clustering is performed and the candidate sequence is labeled using the clustering results. If the time interval standard deviation of each type of labeled data and the L2 norm of the one-dimensional homology group are respectively less than and greater than the corresponding set thresholds, the candidate sequence is finally determined to be a qualified periodic fluctuation sequence; the mean time interval is used as the load fluctuation period, and the starting index and fluctuation period of the candidate sequence are stored in the periodic fluctuation detection result set.

4. The unsupervised rapid load decomposition method based on multi-feature fusion according to claim 1 is characterized in that: The two-stage STS pairing and feature set construction includes: Step 31) Given the periodic fluctuation detection results, STS detection results, and candidate RSTS screening threshold, the total feature set is initialized to an empty set; Step 32) For each FSTS and its timestamp, in the first stage, first screen out the RSTS in the RSTS set whose timestamp is less than the timestamp of the current FSTS, and calculate the distance between the screened RSTS and the FSTS using the steady-state power difference or DTW distance. If the distance is less than the candidate RSTS screening threshold, the RSTS is determined to be a candidate RSTS, and the RSTS and its timestamp are respectively stored in the first candidate RSTS and timestamp set; In the second stage of step 33), for each RSTS and its timestamp in the first candidate RSTS set, a forward search is performed on all FSTSs in the FSTS set whose timestamps are greater than the timestamp of the RSTS and not greater than the timestamp of the FSTS in step 32). For each FSTS, the distance is calculated by fusing power and time and weighted respectively. The FSTS with the smallest distance, i.e., the best pairing with the current RSTS, is obtained. If the timestamp of the FSTS is equal to the timestamp of the FSTS in step 32), the RSTS and the corresponding timestamp are stored in the second candidate RSTS and timestamp set, and the distance calculated between the RSTS and the FSTS is stored in the distance vector. Step 34) The optimal paired RSTS and its timestamp are obtained by locating the minimum index in the distance vector, and feature extraction is performed to construct the pairing feature set; the successfully paired STS and its timestamp are removed from the STS detection result.

5. The unsupervised rapid load decomposition method based on multi-feature fusion according to claim 1 is characterized in that: The unsupervised fast clustering includes: Step 41) Given a similar graph signal screening threshold and a total feature set; Step 42) Each element in the total feature set is considered as a sample, and auxiliary samples are generated for clustering samples to enhance sample discrimination. First, a feature similarity calculation method is defined, and the distance is calculated for each feature. For STS features, the steady-state power is first subtracted and then the absolute value is taken, or the DTW distance is used. For other features, the distance is calculated by first subtracting and then taking the absolute value. Subsequently, min-max normalization is performed on each sub-item distance matrix. Weighted summation is performed on each sub-item distance matrix to obtain a comprehensive distance matrix. The sample with the largest distance to the sample to be clustered is selected as the first auxiliary sample. The second auxiliary sample is defined as an all-zero set. Finally, the sample to be classified, the representative sample, and the auxiliary samples are reordered to generate a second feature set. Step 43) For all elements in the second feature set, update each feature distance matrix, and then perform normalization and weighting before summing; adaptively calculate the scaling factor, define the vertex set and adjacency matrix, and construct an undirected graph; calculate the diagonal matrix and the graph Laplacian matrix based on the adjacency matrix; Step 44) defining the graph signal and calculating the global smoothness of its graph Laplace quadratic form; initializing the graph signal and finding a closed-form solution to the optimization problem; Step 45) Filter out the elements in the closed-form solution that are greater than the similar graph signal screening threshold; take the cluster corresponding to the maximum value of the screened elements as the category of the sample to be clustered, and merge the sample to be clustered into the cluster; and generate a representative sample for the cluster; if all elements in the closed-form solution are less than the graph signal screening threshold, then generate a new category cluster and store the sample to be clustered into the set, update the total feature set, and continue to execute steps 42)-step 46) until all samples are clustered.

6. Unsupervised rapid load decomposition device based on multi-feature fusion, characterized by: include: The STS extraction module is used to set the STS detection threshold and the maximum steady-state sequence interception length before and after the change point, initialize the anchor point and power change, calculate the power change point by point to determine the change point and update the anchor point category, intercept and crop the candidate STS according to the anchor point state, and finally normalize the STS and classify it into open RSTS and closed FSTS sequences according to the power trend; The periodic fluctuation sequence mining module uses the RSTS index as a starting point, intercepts the power sequence through multi-window vectors and segments it by mean, screens out sequences whose number of subsequences meets the threshold, calculates the statistics of time interval and power difference to determine candidate periodic fluctuation sequences, and after processing the disturbance by extending the window, combines delayed embedding, continuous homology and k-means clustering to screen out qualified periodic fluctuation sequences and calculate the fluctuation period. The STS pairing and feature construction module is used in the first phase to screen candidate pairs from RSTSs whose timestamps are smaller than the current FSTS based on the steady-state power difference or DTW distance. In the second phase, the power and time distance are integrated and weighted to forward search for the best paired FSTS. Features such as running time and transient duration are extracted and constructed into a feature set, eliminating already paired STSs. The unsupervised fast clustering module is used to generate auxiliary samples to enhance discrimination. It uses steady-state power difference or DTW distance for STS features and absolute value difference for non-STS features to calculate distance. It normalizes and weights them to generate a comprehensive distance matrix, constructs an undirected graph and graph Laplacian matrix, determines the category of the sample by solving the closed-form solution of the graph signal, and generates representative samples to optimize the clustering process.

7. The unsupervised rapid load decomposition device based on multi-feature fusion according to claim 6 is characterized in that: The STS extraction module includes: The parameter initialization submodule is used to set the STS detection threshold and the maximum steady-state sequence intercept length before and after the STS change point for the total active power sequence; initialize the adaptive positioning anchor point and power change amount, and the historical change point index; The change point detection and anchor point update submodule is used to call the algorithm for each sampling point of the total active power sequence in sequence, calculate the power change between the current point and the next point, and determine whether the absolute value of the power change is greater than the STS detection threshold. If so, the current index is determined to be a change point, and the power change value and sign are used to determine the anchor point category at this moment, and the corresponding anchor point and its power change are updated. The historical change point index is set as the current index, and the steady-state power sequence before the change point and the change point are stored in the candidate STS; if the absolute value of the power change is less than the STS detection threshold, and the number of intercepted steady-state points is less than the maximum steady-state sequence interception length after the change point, the current power point is stored in the candidate STS; The frequent fluctuation processing submodule is used to determine that the load is currently fluctuating frequently if the third anchor point is not zero, intercept the candidate STS using the first and second anchor points to obtain the real STS, set the second anchor point as its corresponding timestamp, store it in the initial STS set and index set respectively, and update the corresponding parameters; The candidate STS cutting submodule is used to, if the candidate STS already contains enough steady-state points, perform differential analysis on the state transition part of the candidate STS to obtain a differential sequence. If the absolute value of the power change in the differential sequence is less than the STS detection threshold, determine the index corresponding to the point as the fourth anchor point; if the second anchor point is not zero, and the power change of the second anchor point is greater than zero or the difference between the power values ​​of the candidate STS end point and the starting point is greater than the STS detection threshold, use the second anchor point to cut the candidate STS; store the obtained STS and its index in the initial STS and index set respectively, and update the corresponding parameters; if the candidate STS already contains enough steady-state points and the fourth anchor point is not zero, use the fourth anchor point to cut the candidate STS, store the obtained STS and its index in the initial STS and index respectively, and update the corresponding parameters; if the candidate STS already contains enough steady-state points and only the first anchor point is not zero, directly store the candidate STS and its index in the initial STS and index respectively, and update the corresponding parameters; The STS classification submodule is used to normalize each element in the initial STS set; the trend of the normalized STS is determined. If the end point is greater than the starting point power value, it is determined to be an RSTS representing load on, and the RSTS representing load on and the corresponding timestamp are stored separately. Otherwise, it is determined to be an FSTS representing load off, and the FSTS representing load off and the corresponding timestamp are stored separately.

8. The unsupervised rapid load decomposition device based on multi-feature fusion according to claim 6 is characterized in that: The periodic fluctuation sequence mining module includes: The parameter configuration submodule is used to determine the total active power sequence, the periodic fluctuation sequence mining multi-window vector, the periodic fluctuation detection number threshold, the periodic fluctuation standard deviation threshold, the power standard deviation threshold, the power relative error threshold, the periodic fluctuation relative error threshold, the STS extraction result, the delay embedding dimension and delay parameter, and the one-dimensional homology group norm threshold. The multi-window sequence interception submodule is used to intercept the total active power sequence by using the RSTS index as the detection starting point one by one, and using the periodic fluctuation sequence to mine the elements in the multi-window vector from large to small to intercept the total active power sequence. The intercepted sequence is segmented using the sequence mean. If either the number of subsequences above the mean or the number of subsequences below the mean meets or exceeds the periodic fluctuation detection threshold, the next step is determined; otherwise, the next element in the multi-window vector is called to intercept again. The candidate sequence initial screening module is used to obtain the time interval statistics of the subsequences above the mean, including the mean and standard deviation, for sequences with subsequences whose number is not less than the periodic fluctuation detection threshold. It also obtains the power difference between adjacent segmented sequences and calculates the mean and standard deviation. If the standard deviation of the time interval and power is less than the set threshold, it means that the currently intercepted sequence is a candidate periodic fluctuation sequence. The disturbance processing and sequence splicing submodule is used to intercept power data from the total active power sequence using an extended window of a given length. If the relative error between the intercepted power segment and the power mean in the candidate sequence primary screening submodule is greater than a set threshold, the electrical disturbance is determined to be a disturbance and the electrical disturbance is set to zero. The power segment is spliced ​​with the sequence of the candidate sequence primary screening submodule, and the mean and standard deviation of the time interval after segmentation are continuously calculated. If the relative error between the standard deviation of the current round and the previous round is less than a set threshold, the disturbance processing and sequence splicing submodule is repeatedly executed until the condition is no longer met, and the intercepted candidate sequence and index information are stored in the corresponding set. The periodic feature confirmation submodule is used to perform delay embedding on the candidate sequence using the given delay embedding dimension and delay parameter to obtain the embedded data, and calculate its homology group using continuous homology. Then, k-means clustering is performed and the clustering results are used to label the candidate sequence. If the standard deviation of the time interval of each type of labeled data and the L2 norm of the one-dimensional homology group are less than and higher than the corresponding set thresholds respectively, the candidate sequence is finally determined to be a qualified periodic fluctuation sequence; the mean time interval is used as the load fluctuation period, and the starting index and fluctuation period of the candidate sequence are stored in the periodic fluctuation detection result set.

9. The unsupervised rapid load decomposition device based on multi-feature fusion according to claim 6 is characterized in that: The STS pairing and feature construction module includes: The initial parameter and data preparation submodule is used to initialize the total feature set to an empty set given the periodic fluctuation detection results, STS detection results, and candidate RSTS screening thresholds; The first-stage candidate RSTS screening submodule is used to screen out RSTSs with timestamps smaller than the current FSTS timestamp in the RSTS set for each FSTS and its timestamp in the first stage, and calculate the distance between the screened RSTS and the FSTS using the steady-state power difference or DTW distance. If the distance is less than the candidate RSTS screening threshold, the RSTS is determined to be a candidate RSTS and the RSTS and its timestamp are stored in the first candidate RSTS and timestamp set respectively; The second-stage optimal pairing search submodule is used to forward search for all FSTSs in the FSTS set whose timestamps are greater than the RSTS timestamp and not greater than the FSTS timestamp in step 32) for each RSTS and its timestamp in the first candidate RSTS set, and for each FSTS, fuse power and time to perform distance calculation and assign weights respectively; obtain the FSTS with the smallest distance, that is, the FSTS that is best paired with the current RSTS; if the timestamp of the FSTS is equal to the FSTS timestamp in the first-stage candidate RSTS screening submodule, then store the RSTS and the corresponding timestamp in the second candidate RSTS and timestamp set, and store the distance calculated between the RSTS and the FSTS in the distance vector; The feature extraction and set construction submodule is used to obtain the optimal paired RSTS and its timestamp by locating the minimum index in the distance vector, perform feature extraction, and construct the paired feature set; the successfully paired STS and its timestamp are removed from the STS detection results.

10. The unsupervised rapid load decomposition device based on multi-feature fusion according to claim 6 is characterized in that: The unsupervised fast clustering module includes: Cluster initialization submodule, used to give the signal screening threshold of similar graphs and the total feature set; The auxiliary sample generation and distance calculation submodule is used to treat each element in the total feature set as a sample, and to generate auxiliary samples for clustering samples to enhance sample discrimination. First, the feature similarity calculation method is defined, and the distance is calculated for each feature. For STS features, the steady-state power is first subtracted and then the absolute value is taken, or the DTW distance is used. For other features, the distance is calculated by first subtracting and then taking the absolute value. Then, the min-max normalization is applied to each sub-item distance matrix. The weighted sum of each sub-item distance matrix is ​​performed to obtain a comprehensive distance matrix. The sample with the largest distance to the sample to be clustered is selected as the first auxiliary sample. The second auxiliary sample is defined as an all-zero set. Finally, the sample to be classified, the representative sample, and the auxiliary samples are reordered to generate the second feature set. The graph model construction submodule is used to update the feature distance matrix for all elements in the second feature set, and then perform normalization and weighting before summing. It also adaptively calculates the scaling factor, defines the vertex set and adjacency matrix, and constructs an undirected graph. It also calculates the diagonal matrix and the graph Laplacian matrix based on the adjacency matrix. The graph signal optimization solution submodule is used to define the graph signal and calculate the global smoothness of its graph Laplace quadratic form; initialize the graph signal and find the closed-form solution of the graph signal optimization problem; The cluster allocation and optimization submodule is used to filter out elements in the closed-form solution that are larger than the similar graph signal screening threshold; the cluster corresponding to the maximum value of the screened elements is used as the category of the sample to be clustered, and the sample to be clustered is merged into the cluster; and a representative sample is generated for the cluster; if all elements in the closed-form solution are smaller than the graph signal screening threshold, a new category cluster is generated and the sample to be clustered is stored in the set, and the total feature set is updated.

Images