A signal feature extraction method for intelligent load identification of dedicated transformer users

Abstract

The present invention relates to the field of data processing technology, and specifically to a signal feature extraction method for intelligent load identification of special transformer users, including: calculating the similarity of single-cycle load curves of any two load devices to construct difficult-to-distinguish categories; obtaining a dictionary matrix and a sparse matrix for each load device, obtaining the correspondence between the columns of the dictionary matrices of different load devices according to the dictionary matrix, and further obtaining distinguishing features or corresponding chains; obtaining a feature combination set of each load device according to the sparse matrix and the dictionary matrix, performing a sameness judgment operation on the feature combinations of different load devices according to the corresponding chain, and obtaining a distinguishing feature combination set of each load device in the difficult-to-distinguish category; constructing a distinguishing feature library according to the distinguishing features and the distinguishing feature combination set, and identifying the load curve according to the distinguishing feature library. The present invention can effectively distinguish and identify similar load curves by constructing a distinguishing feature library.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a signal feature extraction method for intelligent load identification of dedicated transformer users. Background Art

[0002] Different loads have different circuit structures and functions, and therefore have different operating characteristics and load features. Load characteristics are the basis for determining load equipment and the key to load decomposition. Different types of load equipment may have multiple characteristics, and the characteristics of different loads may be similar, which will interfere with the process and results of the load identification method based on load characteristics.

[0003] Based on this, the present invention proposes a method for load identification by combining multiple load features, which can reasonably improve the efficiency of load identification. On the other hand, the combination of multiple load features means that it is necessary to establish a load feature database that is as perfect as possible, distinguish different features from each other according to certain rules, and collect and save corresponding feature quantity information. Therefore, the present invention distinguishes similar features between different loads by feature decomposition, and at the same time establishes a distinguishing feature combination. The distinguishing feature combination is used as a distinguishing feature, which can effectively distinguish similar features of different loads, and then establish a distinguishing feature library. When using neural networks to distinguish different loads in existing methods, a large amount of data is required to train the neural network, and a large amount of data labeling and preliminary data preprocessing are required. Compared with the existing method of using neural networks to distinguish different loads, the present invention greatly reduces the workload and improves the detection efficiency. Summary of the invention

[0004] The present invention provides a signal feature extraction method for intelligent load identification of dedicated transformer users to solve the existing problems.

[0005] The signal feature extraction method for intelligent load identification of dedicated transformer users of the present invention adopts the following technical solution:

[0006] An embodiment of the present invention provides a signal feature extraction method for intelligent load identification of a dedicated transformer user, the method comprising the following steps:

[0007] Obtaining a single-cycle load curve of each load device; calculating the similarity of the single-cycle load curves of any two load devices, and classifying the two load devices whose similarity is greater than a first preset threshold into a difficult-to-distinguish category;

[0008] Obtain a dictionary matrix and a sparse matrix for each load device; obtain the similarity between any two columns of the dictionary matrices of two load devices in the difficult-to-distinguish category, and obtain the correspondence between the columns of the dictionary matrices of the two load devices in the difficult-to-distinguish category based on the similarity between any two columns; obtain a similarity sequence and a discrimination coefficient based on all corresponding columns between the dictionary matrices of the two load devices in the difficult-to-distinguish category; when the discrimination coefficient is less than a third preset threshold, use the column in the dictionary matrix corresponding to the element in the similarity sequence that is less than the second preset threshold as a distinguishing feature; when the discrimination coefficient is greater than or equal to the third preset threshold, obtain multiple feature combinations of each load device according to the sparse matrix of each load device to form a feature combination set, wherein the feature combination is a combination of multiple columns in the dictionary matrix;

[0009] According to the correspondence between the columns of the dictionary matrix of the two load devices in all the difficult-to-distinguish categories, a plurality of corresponding chains are obtained;

[0010] Perform the sameness judgment operation on any two feature combinations between the feature combination sets of two load devices in the difficult-to-distinguish category according to the corresponding chain, and obtain the distinguishing feature combination set of each load device in the difficult-to-distinguish category, wherein each feature combination in the distinguishing feature combination set is different from all feature combinations in the distinguishing feature combination set of another load device in the difficult-to-distinguish category;

[0011] A distinguishing feature library is constructed according to the distinguishing features or a combination set of distinguishing features of each load device in each difficult-to-distinguish category; and the load curve to be distinguished is identified according to the distinguishing feature library.

[0012] Preferably, obtaining the dictionary matrix and sparse matrix of each load device includes:

[0013] The single-cycle load curve of each load device is converted into a load matrix, and the load matrix is ​​decomposed by K-SVD to obtain a dictionary matrix and a sparse matrix.

[0014] Preferably, the obtaining of the similarity between any two columns of the dictionary matrix of two load devices in the difficult-to-distinguish category, and obtaining the correspondence between the columns of the dictionary matrix of the two load devices in the difficult-to-distinguish category according to the similarity of any two columns comprises:

[0015] Calculate the cosine similarity of any two columns between the dictionary matrices of two load devices in the difficult-to-distinguish category, use the cosine similarity as the weight between the two columns, use the KM matching algorithm to obtain the matching relationship between the columns of the dictionary matrices of the two load devices in the difficult-to-distinguish category, and use the two columns with a matching relationship whose cosine similarity is greater than a second preset threshold as corresponding columns.

[0016] Preferably, the obtaining of similarity sequence and discrimination coefficient according to all corresponding columns between dictionary matrices of two load devices in the difficult-to-distinguish category comprises:

[0017] The cosine similarities of all matching columns between the dictionary matrices of two load devices in the difficult-to-distinguish category are combined into a similarity sequence, the number of elements in the similarity sequence that are greater than a second preset threshold is obtained as the first number, and the ratio of the first number to the number of all elements in the similarity sequence is used as the discrimination coefficient.

[0018] Preferably, the step of acquiring multiple feature combinations of each load device according to the sparse matrix of each load device to form a feature combination set comprises:

[0019] Get the columns corresponding to all the elements with values ​​of 1 in the sparse matrix in the dictionary matrix, take the columns corresponding to all the elements with values ​​of 1 in a column of the sparse matrix in the dictionary matrix as a feature combination, and form a feature combination set with all the feature combinations.

[0020] Preferably, the obtaining of a plurality of corresponding chains according to the correspondence between the columns of the dictionary matrix of two load devices in all the difficult-to-distinguish categories comprises:

[0021] A column in the dictionary matrix of the same load device in multiple difficult-to-distinguish categories and the columns in the dictionary matrices of other load devices corresponding to the column form a corresponding chain.

[0022] Preferably, the operation of performing the sameness judgment operation on any two feature combinations between the feature combination sets of two load devices in the difficult-to-distinguish category according to the corresponding chain includes:

[0023] Obtain any two feature combinations between the feature combination sets of two load devices in the difficult-to-distinguish category as the first feature combination and the second feature combination, respectively. When each column in the first feature combination and the column in the corresponding position in the second feature combination are in the same corresponding chain, the first feature combination and the second feature combination are the same; otherwise, the first feature combination and the second feature combination are different.

[0024] Preferably, the identifying of the load curve to be distinguished according to the distinguishing feature library includes:

[0025] Calculate the similarity between the load curve to be distinguished and the load curve of each load device. When only one of the similarities between the load curve to be distinguished and the load curves of all load devices is greater than a first preset threshold, use the load device corresponding to the similarity greater than the first preset threshold as the load device corresponding to the load curve to be distinguished;

[0026] When two or more of the similarities between the load curve to be distinguished and the load curves of all load devices are greater than a first preset threshold, each load device corresponding to the similarity greater than the first preset threshold is used as a device to be matched; a feature combination set of the load curve to be distinguished is obtained; and the load device corresponding to the load curve to be distinguished is obtained based on the distinguishing features or the distinguishing feature combination set between the devices to be matched in the distinguishing feature library and the feature combination set of the load curve to be distinguished.

[0027] The beneficial effects of the technical solution of the present invention are:

[0028] 1. When the traditional method uses a neural network to distinguish different loads, the amount of data is large, and the same method is used to calculate both easy and difficult to distinguish loads, which easily leads to a large waste of computing resources. The present invention first divides the load curves into easy and difficult to distinguish categories by thresholding through the cosine similarity values ​​of different load curves. Different calculation complexity methods are used to distinguish different categories: for easy to distinguish loads, the cosine similarity between the load curve to be distinguished and the load curve of each load device is directly calculated to determine the load device corresponding to the load curve to be distinguished. The distinction process is simple and efficient; for difficult to distinguish loads, a deeper feature combination method is calculated, and the distinction is made according to the different feature combination methods, and the accuracy of the distinction result is higher. The present invention adopts an adaptive method to distinguish different difficulties of distinction, rather than a single method. On the one hand, the amount of calculation is small, and on the other hand, the distinction accuracy is higher. For different load curves, simple, efficient and high-precision distinction can be achieved.

[0029] 2. Since the columns of the same order corresponding to the dictionary matrices of different load devices are different, directly calculating the maximum cosine similarity will result in large errors when the power is off or charging changes greatly or when the curve is distorted due to other factors. When calculating the correspondence between the features of different load devices, the present invention uses KM matching calculation to ensure that when the curve suddenly changes at a certain moment due to unexpected circumstances and some features that do not belong to the normal curve appear, the overall optimization can be used to avoid the local optimization that causes the feature correspondence error, and further avoid the feature correspondence error causing large errors.

[0030] 3. In actual calculation, the real-time requirement is often high. For example, when judging which situation caused the fire based on the historical curve and taking different firefighting measures, the real-time requirement is high. However, the existing method uses a neural network with a large amount of calculation, which is difficult to meet the real-time requirement. The present invention establishes a distinguishing feature library, directly calculates the features of the load curve to be distinguished, and compares them with the features in the distinguishing feature library to identify the load equipment corresponding to the load curve to be distinguished. The present invention has a small amount of calculation and helps to meet the real-time requirement. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] Figure 1 It is a flow chart of the steps of a signal feature extraction method for intelligent load identification of dedicated transformer users according to the present invention;

[0033] Figure 2 A bipartite graph of the columns of a dictionary matrix of a signal feature extraction method for intelligent load identification of a dedicated transformer user according to the present invention;

[0034] Figure 3 It is a schematic diagram of a feature combination acquisition method of a signal feature extraction method for intelligent load identification of a dedicated transformer user according to the present invention;

[0035] Figure 4 A bipartite graph of the fifth node and the sixth node of a signal feature extraction method for intelligent load identification of a dedicated transformer user according to the present invention;

[0036] Figure 5 It is a bipartite graph of the third node and the fourth node of a signal feature extraction method for intelligent load identification of dedicated transformer users according to the present invention. DETAILED DESCRIPTION

[0037] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the signal feature extraction method for intelligent load identification of special transformer users proposed by the present invention, its specific implementation method, structure, characteristics and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0038] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0039] The following is a detailed description of a specific solution of a signal feature extraction method for intelligent load identification of dedicated transformer users provided by the present invention in conjunction with the accompanying drawings.

[0040] See also Figure 1, which shows a flowchart of a signal feature extraction method for intelligent load identification of a dedicated transformer user provided by an embodiment of the present invention, the method comprising the following steps:

[0041] 101. Obtain the single-cycle load curve of each load device and divide the difficult-to-distinguish categories based on the single-cycle load curve.

[0042] First, the load curves of different load devices are collected as initial data for analysis. Since the purpose of this embodiment is to distinguish devices that are difficult to distinguish by load curves, it is first necessary to calculate which devices need to be distinguished, that is, to obtain devices that cannot be distinguished by the similarity of load curves.

[0043] First, perform spectrum analysis on the load curve of each load device to obtain the single-cycle curve of each load device, which is called the single-cycle load curve. Secondly, calculate the cosine similarity of the single-cycle load curves of any two load devices. The cosine similarity reflects the overall similarity of the single-cycle load curve. It is difficult to identify the local subtle differences in the single-cycle load curve, and the single-cycle load curves that are difficult to distinguish often have differences in local subtle aspects. Therefore, it can be considered that the load devices corresponding to the two single-cycle load curves with larger cosine similarity are difficult to distinguish. Set a first preset threshold k. When the cosine similarity of the single-cycle load curves of two load devices is greater than the first preset threshold, the two load devices are difficult to distinguish, and the two load devices are divided into a difficult-to-distinguish category.

[0044] It should be noted that to calculate the cosine similarity of the single-cycle load curves of any two load devices, the single-cycle load curve of each load device is regarded as a high-dimensional vector, and the cosine similarity between the two high-dimensional vectors is calculated. Furthermore, if the dimensions of the two high-dimensional vectors are different, the high-dimensional vector with high dimension is first reduced to make it the same dimension as the high-dimensional vector with low dimension, and then the cosine similarity is calculated.

[0045] In this embodiment, the first preset threshold k=0.8. In other embodiments, the implementer may set the value of the first preset threshold k according to actual conditions.

[0046] 102. Get the dictionary matrix and sparse matrix of each load device.

[0047] It should be noted that, in order to facilitate subsequent analysis, the single-cycle load curve of each load device needs to be converted into a matrix form with the same number of rows and different numbers of columns (the different number of columns is because the length of the cycle curve of each load device is not necessarily the same). Therefore, it is necessary to obtain the standard number of rows first.

[0048] In this embodiment, the single-cycle load curve of each load device is first converted into a matrix with equal number of rows and columns, for example, a 99-dimensional vector is converted into a 10×10 matrix, where the empty positions in the matrix are filled with 0. Then the sum of the difference between the number of rows of each matrix and the number of rows of all other matrices is calculated as the row number excellence of each matrix, and the number of rows of the matrix corresponding to the minimum row number excellence is taken as the standard number of rows. The single-cycle load curve of each load device is converted into a matrix with the standard number of rows as the load matrix of each load device. At this time, the number of rows of all load matrices is the same. It should be noted that in the conversion process of the single-cycle load curve, the empty positions in the obtained matrix are filled with 0.

[0049] It should be noted that the load curves that are difficult to distinguish must have the same characteristics. It is difficult to distinguish them simply from the characteristics, but they can be distinguished by the connection between different characteristics. The purpose of this embodiment is to solve the problem that two load curves with similar characteristics are difficult to distinguish. By calculating the connection between different characteristics, the load curves are distinguished according to the connection between different characteristics. K-SVD is a method for extracting matrix features.

[0050] In this embodiment, a hyperparameter α is set, and each column of the load matrix of each load device is used as an input vector to obtain all input vectors of each load device. All input vectors of each load device are used as inputs of the K-SVD algorithm, and the load matrix of each load device is decomposed using the K-SVD algorithm to obtain a dictionary matrix and a sparse matrix of each load device. It should be noted that the load matrices of different load devices have the same number of rows and different numbers of columns. By setting the hyperparameter α, the dictionary matrices of all load devices obtained by decomposing the load matrix by K-SVD have the same number of rows and the same number of columns. In this embodiment, the hyperparameter α=8. In other embodiments, the implementer can set the value of the hyperparameter α as needed.

[0051] 103. Obtain the correspondence between the columns of the dictionary matrix of two load devices in the difficult-to-distinguish category.

[0052] It should be noted that each column of the dictionary matrix corresponds to an atomic signal, that is, a feature. The types of atomic signals of two similar load curves are often similar. For example, the load curves of a hair dryer and an electric rice cooker are similar, and both have features a, b, and c. However, since there are two load curves, the combination of features of each load curve is different. For example, the combination of features ab and bc in a hair dryer, and the combination of features ac and bc in an electric rice cooker. Each column of the dictionary matrix represents a feature, and the distribution of each column of the sparse matrix represents different columns of the dictionary matrix, that is, the combination of different features. Since the feature combination of each load device needs to be calculated, the features are first calculated, and then the combination between the features is calculated. When it is difficult to distinguish the dictionary matrices of two load devices A and B in a category, since each column of the dictionary matrix represents a feature, it is necessary to first calculate the correspondence between the columns in the dictionary matrix of load device A and the columns in the dictionary matrix of load device B, that is, the correspondence of the features, so that the distinguishing features or distinguishing feature combinations can be obtained according to the corresponding relationship of the features to construct a distinguishing feature library.

[0053] In this embodiment, the method for obtaining the correspondence between the columns in the dictionary matrix of the load device A in the difficult-to-distinguish category and the columns in the dictionary matrix of the load device B is as follows:

[0054] First, calculate the cosine similarity between each column in the dictionary matrix of load device A and each column in the dictionary matrix of load device B. Take each column in the dictionary matrix of load device A as a node, and each column in the dictionary matrix of load device B as a node, such as Figure 2 As shown, Figure 2 Each node on the left side corresponds to each column in the dictionary matrix of load device A, and each node on the left side of the right side corresponds to each column in the dictionary matrix of load device B. The cosine similarity between each column in the dictionary matrix of load device A and each column in the dictionary matrix of load device B is used as the weight between the corresponding nodes, thus constructing a bipartite graph. The KM matching algorithm is used to perform the optimal maximum matching on the bipartite graph to obtain the matching relationship between nodes in the bipartite graph, that is, to obtain the matching relationship between the columns in the dictionary matrix of load device A and the columns in the dictionary matrix of load device B. The cosine similarity between all the columns in the dictionary matrix of load device A and the matching columns in the dictionary matrix of load device B constitutes a similarity sequence.

[0055] When the cosine similarity between a column in the dictionary matrix of load device A and the matching column in the dictionary matrix of load device B is large, the two matching columns have a corresponding relationship. Conversely, when the cosine similarity between all the columns in the dictionary matrix of load device A and the matching columns in the dictionary matrix of load device B is small, the two matching columns do not actually have a corresponding relationship. Therefore, a second preset threshold w is set. When the cosine similarity between a column in the dictionary matrix of load device A and the matching column in the dictionary matrix of load device B is greater than or equal to the second preset threshold w, the column in the dictionary matrix of load device A and the matching column in the dictionary matrix of load device B are used as corresponding columns. In this way, the corresponding relationship between the columns of the dictionary matrices of two load devices in the difficult-to-distinguish category can be obtained. In this embodiment, the second preset threshold w=0.7. In other embodiments, the implementer can set the value of the second preset threshold as needed.

[0056] At this point, the correspondence between the columns of the dictionary matrix of the two load devices in the difficult-to-distinguish category, that is, the correspondence between the features of the two load devices in the difficult-to-distinguish category, is obtained, so that the distinguishing features or distinguishing feature combinations of the two load devices in the difficult-to-distinguish category can be more accurately obtained based on the corresponding relationship of the features, and the distinguishing feature library subsequently constructed based on the distinguishing features and distinguishing feature combinations is further more accurate.

[0057] 104. Obtain distinguishing features or a set of distinguishing feature combinations, and build a distinguishing feature library.

[0058] The number of elements in the similarity sequence that are greater than or equal to the second preset threshold w is obtained as the first number, and the ratio of the first number to the number of all elements in the similarity sequence is used as the discrimination coefficient. When the discrimination coefficient is less than the third preset threshold u, it means that the load curves of load device A and load device B can be distinguished by the feature. At this time, the column in the dictionary matrix corresponding to the elements in the similarity sequence that are less than the second preset threshold is used as the distinguishing feature. In this embodiment, the third preset threshold u=0.8. In other embodiments, the implementer can set the value of the third preset threshold as needed.

[0059] If the distinction coefficient is greater than or equal to the third preset threshold u, it means that the correspondence between the columns in the dictionary matrix of load device A and the dictionary matrix of load device B is strong, which further indicates that it is difficult to distinguish the load curves of load device A and load device B by a single feature. At this time, it is necessary to further distinguish based on the relationship between the features.

[0060] It should be noted that a sparse matrix represents a combination of different features. There are only two numbers in a sparse matrix, 0 and 1. The distribution of 0 and 1 in the columns of a sparse matrix represents a combination of features. 0 means that the feature does not exist, and 1 means that the feature exists. Figure 3 Where D is the dictionary matrix, ri is the i-th column of the sparse matrix, and d1, d2, …, d8 represent the 1st, 2nd, …, 8th columns in the dictionary matrix D respectively. The 1 in the i-th column of the sparse matrix indicates that the features of the 1st and 6th columns of the dictionary matrix D are a feature combination. Each column of the sparse matrix represents a different feature combination, that is, one load device corresponds to multiple feature combinations. For example: two load devices A and B in a difficult-to-distinguish category both have three features a, b, and c. Through the sparse matrix of load device A, we get that load device A is a combination of features ab and bc, and through the sparse matrix of load device B, we get that load device B is a combination of features ac and bc. Then, we can use feature combinations to distinguish load devices A and B.

[0061] In this embodiment, firstly, the corresponding chain is obtained according to the corresponding relationship between the columns of the dictionary matrix of two load devices in all the difficult-to-distinguish categories. The specific process is as follows:

[0062] Assume that the first indistinguishable category includes load device A and load device B, and the second indistinguishable category includes load device B and load device C. The first column in the dictionary matrix of load device A corresponds to the third column in the dictionary matrix of load device B, and the third column in the dictionary matrix of load device B corresponds to the fifth column in the dictionary matrix of load device C. Then the corresponding chain is d A1 -d B3 -d C5 , where d A1 The first column in the dictionary matrix of load device A, d B3 The third column in the dictionary matrix of load device B, d C5 Represents the 5th column in the dictionary matrix of load device C.

[0063] Similarly, all corresponding chains are obtained. For each load device, a characteristic combination can be obtained through the order of the 1 values ​​in each column of its sparse matrix. For example: if the first and third elements of a column of the sparse matrix of load device A are 1, then a characteristic combination of load device A is (d A1 ,d A3 ), where d A1 The first column in the dictionary matrix of load device A; a feature combination of load device B is (d B3 ,d B4 ), where d A3 Represents the third column in the dictionary matrix of load device B.

[0064] Similarly, all feature combinations of each load device are obtained to form a feature combination set of each load device.

[0065] For two load devices in a difficult-to-distinguish category, the identity of any two feature combinations between their feature combination sets is determined. If each element in the feature combination is in the same corresponding chain as the element in the corresponding position in the other feature combination, then the two feature combinations are the same. Otherwise, the two feature combinations are different. For example, a feature combination (d A1 ,d A3 ) and a characteristic combination of load equipment B (d B3 ,d B4 ) to make the same judgment, d A1 With d B3 are all located in a corresponding chain d A1 -d B3 -d C5 Medium, d A3 With d B4 Both are located in another corresponding chain d A3 -d B4 -d C7 Then the feature combination (d A1 ,d A3 ) and feature combination (d B3 ,d B4 )same.

[0066] A feature combination in the feature combination set of each load device in the difficult-to-distinguish category that is different from all feature combinations in the feature combination set of another load device is obtained as the distinguishing feature combination of each load device. All distinguishing feature combinations of each load device in the difficult-to-distinguish category constitute a distinguishing feature combination set of each load device in the difficult-to-distinguish category.

[0067] A distinguishing feature library is constructed using the distinguishing features or a combination of distinguishing features of each load device in all difficult-to-distinguish categories.

[0068] At this point, a distinguishing feature library is obtained. This embodiment obtains a feature combination that can accurately distinguish the two load devices by analyzing the similarities between the feature combinations of each load device in the difficult-to-distinguish category, and constructs a distinguishing feature library based on the distinguishing feature combination set, so that the subsequent recognition result of each load curve based on the distinguishing features in the distinguishing feature library and the distinguishing feature combination set is more accurate.

[0069] 105. Identify the load curve to be differentiated based on the differentiation feature library.

[0070] The cosine similarity between the load curve to be distinguished and the load curve of each load device is calculated. When only one of the cosine similarities between the load curve to be distinguished and the load curves of all load devices is greater than a first preset threshold, the load device corresponding to the cosine similarity greater than the first preset threshold is used as the load device corresponding to the load curve to be distinguished.

[0071] When two or more of the cosine similarities between the load curve to be distinguished and the load curves of all load devices are greater than the first preset threshold, each load device corresponding to the cosine similarity greater than the first preset threshold is used as a device to be matched. The dictionary matrix of the load curve to be distinguished and the feature combination set {s1, s2, …, s n}, where s n represents the bth feature combination in the feature combination set, and n is the number of feature combinations in the feature combination set. All devices to be matched are denoted as R1, R2, …, R z , where R z is the zth device to be matched, and z is the number of devices to be matched. All devices to be matched form a difficult-to-distinguish category, such as R1 and R2. z Form a difficult-to-distinguish category, R2,R z Form a difficult-to-distinguish category, etc. Obtain the distinguishing features or distinguishing feature combination set of each load device in each of the above difficult-to-distinguish categories in the distinguishing feature library.

[0072] When there are distinguishing features between two devices to be matched, each feature of the load curve to be distinguished (i.e., each column in the dictionary matrix of the load curve to be distinguished) is taken as a first node, and each distinguishing feature of a load device in the difficult-to-distinguish category composed of the two devices to be matched is taken as a second node. The cosine similarity between each first node and the second node is calculated as the weight between the two nodes. A bipartite graph is constructed according to the first node, the second node and the weights between the nodes. The KM matching algorithm is used to perform optimal maximum matching on the bipartite graph, and the sum of the weights of the corresponding nodes in the matching result is obtained as the matching degree of the load curve to be distinguished and a device to be matched in the difficult-to-distinguish category composed of the two devices to be matched.

[0073] When there is a distinguishing feature combination set between two devices to be matched, each feature combination in the feature combination set of the load curve to be distinguished is used as a third node, and each feature combination in the distinguishing feature combination set of a load device in the difficult-to-distinguish category composed of the two devices to be matched is used as a fourth node, and the weight between each third node and the fourth node is calculated. The specific method is as follows:

[0074] The third node is the characteristic combination set {s1,s2,…,sn}, each feature combination is composed of multiple columns in the dictionary matrix of the load curve to be distinguished, then each feature combination can be expressed as [d' i ,d' j ,…,d' m ], where d' i is the i-th column in the dictionary matrix of the load curve to be distinguished, d' j is the jth column in the dictionary matrix of the load curve to be distinguished, d' m is the mth column in the dictionary matrix of the load curve to be distinguished. The fourth node is a feature combination of a distinguishing feature combination set of a load device in a difficult-to-distinguish category composed of two devices to be matched. For example, the distinguishing feature combination set of the load device R1 in the difficult-to-distinguish category composed of the devices to be matched R1 and R2 is {p1, p2, …, p t}, where p t is the tth feature combination in the feature combination set of the load device R1 to be distinguished in the difficult-to-distinguish category composed of the devices to be matched R1 and R2, and t is the number of feature combinations in the feature combination set of the load device R1 to be distinguished in the difficult-to-distinguish category composed of the devices to be matched R1 and R2. Then p1, p2, …, p t are respectively a fourth node. The feature combination corresponding to the fourth node is composed of multiple columns in the dictionary matrix of the device to be matched R1, then each feature combination corresponding to the fourth node can be expressed as in is the uth column in the dictionary matrix of the device to be matched R1, is the vth column in the dictionary matrix of the device to be matched R1, is the lth column in the dictionary matrix of the device to be matched R1. Taking the third node s1 and the fourth node p1 as an example, calculate the weight between s1 and p1: replace each column d' corresponding to the third node s1 in the dictionary matrix of the load curve to be distinguished i ,d' j ,…,d' m Each of them is regarded as a node, recorded as the fifth node, and each column corresponding to the fourth node p1 in the dictionary matrix of the device to be matched R1 is Each of them is regarded as a node and recorded as the sixth node. The cosine similarity between each fifth node and each sixth node is calculated as the weight between the two nodes. A bipartite graph is constructed based on the fifth node, the sixth node and the weight between the nodes. Figure 4 The KM matching algorithm is used to perform the optimal maximum matching of the bipartite graph, and the sum of the weights of the corresponding nodes in the matching result is obtained as the weight between the third node s1 and the fourth node p1.

[0075] After obtaining the weight between the third node and the fourth node, a bipartite graph is constructed based on the weights of the third node, the fourth node, and the nodes. Figure 5 The KM matching algorithm is used to perform the optimal maximum matching of the bipartite graph, and the sum of the weights of the corresponding nodes in the matching result is obtained as the matching degree of a device to be matched in the difficult-to-distinguish category consisting of the load curve to be distinguished and the two devices to be matched.

[0076] Similarly, all matching degrees between the load curve to be distinguished and each device to be matched are obtained, and the device to be matched corresponding to the maximum matching degree is selected as the load device corresponding to the load curve to be distinguished.

[0077] Through the above steps, the construction of the distinguishing feature library and the identification of the load curve are completed.

[0078] The embodiment of the present invention constructs a difficult-to-distinguish category by calculating the similarity of the single-cycle load curves of any two load devices; obtains a dictionary matrix and a sparse matrix for each load device, obtains the correspondence between the columns of the dictionary matrices of different load devices based on the dictionary matrix, and further obtains distinguishing features or corresponding chains; obtains a feature combination set for each load device based on the sparse matrix and the dictionary matrix, performs a sameness judgment operation on the feature combinations of different load devices based on the corresponding chain, and obtains a distinguishing feature combination set for each load device in the difficult-to-distinguish category; constructs a distinguishing feature library based on the distinguishing features and the distinguishing feature combination set, and identifies the load curve based on the distinguishing feature library. The present invention can effectively distinguish and identify similar load curves by constructing a distinguishing feature library. Compared with the existing method of identifying load curves to be distinguished using neural networks, the method of the present invention has a small amount of calculation and helps to meet real-time requirements.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A signal feature extraction method for intelligent load identification of dedicated transformer users, characterized in that: The method comprises the following steps: Obtaining a single-cycle load curve of each load device; calculating the similarity of the single-cycle load curves of any two load devices, and classifying the two load devices whose similarity is greater than a first preset threshold into a difficult-to-distinguish category; Obtain a dictionary matrix and a sparse matrix for each load device; obtain the similarity between any two columns of the dictionary matrices of two load devices in the difficult-to-distinguish category, and obtain the correspondence between the columns of the dictionary matrices of the two load devices in the difficult-to-distinguish category based on the similarity between any two columns; obtain a similarity sequence and a discrimination coefficient based on all corresponding columns between the dictionary matrices of the two load devices in the difficult-to-distinguish category; when the discrimination coefficient is less than a third preset threshold, use the column in the dictionary matrix corresponding to the element in the similarity sequence that is less than the second preset threshold as a distinguishing feature; when the discrimination coefficient is greater than or equal to the third preset threshold, obtain multiple feature combinations of each load device according to the sparse matrix of each load device to form a feature combination set, wherein the feature combination is a combination of multiple columns in the dictionary matrix; According to the correspondence between the columns of the dictionary matrix of the two load devices in all the difficult-to-distinguish categories, a plurality of corresponding chains are obtained; Perform the sameness judgment operation on any two feature combinations between the feature combination sets of two load devices in the difficult-to-distinguish category according to the corresponding chain, and obtain the distinguishing feature combination set of each load device in the difficult-to-distinguish category, wherein each feature combination in the distinguishing feature combination set is different from all feature combinations in the distinguishing feature combination set of another load device in the difficult-to-distinguish category; A distinguishing feature library is constructed according to the distinguishing features or distinguishing feature combination set of each load device in each difficult-to-distinguish category; and the load curve to be distinguished is identified according to the distinguishing feature library; The obtaining of similarity sequence and discrimination coefficient according to all corresponding columns between dictionary matrices of two load devices in the difficult-to-distinguish category includes: The cosine similarities of all matching columns between the dictionary matrices of two load devices in the difficult-to-distinguish category are used to form a similarity sequence, the number of elements in the similarity sequence that are greater than a second preset threshold is obtained as a first number, and the ratio of the first number to the number of all elements in the similarity sequence is used as a discrimination coefficient; The obtaining of a plurality of corresponding chains according to the corresponding relationship between the columns of the dictionary matrix of two load devices in all the difficult-to-distinguish categories comprises: A column in the dictionary matrix of the same load device in multiple difficult-to-distinguish categories and columns in the dictionary matrices of other load devices corresponding to the column form a corresponding chain; The operation of performing the sameness judgment on any two feature combinations between the feature combination sets of two load devices in the difficult-to-distinguish category according to the corresponding chain includes: Obtain any two feature combinations between the feature combination sets of two load devices in the difficult-to-distinguish category, and use them as the first feature combination and the second feature combination, respectively. When each column in the first feature combination and the column at the corresponding position in the second feature combination are in the same corresponding chain, the first feature combination and the second feature combination are the same. Otherwise, the first feature combination and the second feature combination are different. The identifying of the load curve to be distinguished according to the distinguishing feature library includes: Calculate the similarity between the load curve to be distinguished and the load curve of each load device. When only one of the similarities between the load curve to be distinguished and the load curves of all load devices is greater than a first preset threshold, use the load device corresponding to the similarity greater than the first preset threshold as the load device corresponding to the load curve to be distinguished; When two or more of the similarities between the load curve to be distinguished and the load curves of all load devices are greater than a first preset threshold, each load device corresponding to the similarity greater than the first preset threshold is used as a device to be matched; a feature combination set of the load curve to be distinguished is obtained; and the load device corresponding to the load curve to be distinguished is obtained based on the distinguishing features or the distinguishing feature combination set between the devices to be matched in the distinguishing feature library and the feature combination set of the load curve to be distinguished.

2. According to claim 1, a signal feature extraction method for intelligent load identification of dedicated transformer users is characterized in that: The obtaining of the dictionary matrix and the sparse matrix of each load device comprises: The single-cycle load curve of each load device is converted into a load matrix, and the load matrix is ​​decomposed by K-SVD to obtain a dictionary matrix and a sparse matrix.

3. According to claim 1, a signal feature extraction method for intelligent load identification of dedicated transformer users is characterized in that: The obtaining of the similarity between any two columns of the dictionary matrix of two load devices in the difficult-to-distinguish category, and obtaining the correspondence between the columns of the dictionary matrix of the two load devices in the difficult-to-distinguish category according to the similarity between any two columns include: Calculate the cosine similarity of any two columns between the dictionary matrices of two load devices in the difficult-to-distinguish category, use the cosine similarity as the weight between the two columns, use the KM matching algorithm to obtain the matching relationship between the columns of the dictionary matrices of the two load devices in the difficult-to-distinguish category, and use the two columns with a matching relationship whose cosine similarity is greater than a second preset threshold as corresponding columns.

4. According to claim 1, a signal feature extraction method for intelligent load identification of dedicated transformer users is characterized in that: The step of obtaining multiple feature combinations of each load device according to the sparse matrix of each load device to form a feature combination set includes: Get the columns corresponding to all the elements with values ​​of 1 in the sparse matrix in the dictionary matrix, take the columns corresponding to all the elements with values ​​of 1 in a column of the sparse matrix in the dictionary matrix as a feature combination, and form a feature combination set with all the feature combinations.

Images