A load identification method based on adaptive recursive matrix
By adopting the combination method of adaptive recursive matrix and deep learning network in non-invasive load monitoring technology, the problem of low load recognition accuracy caused by the difficulty of hyperparameter formulation is solved, and higher load recognition accuracy and universality are achieved.
Patent Information
- Application Number
- CN202210537601.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-18
- Publication Date
- 2025-06-06
- Estimated Expiration
- 2042-05-18
AI Technical Summary
In the existing non-invasive load monitoring technology, the formulation of hyperparameters is difficult, resulting in low load recognition accuracy.
The load recognition method based on the adaptive recursive matrix is adopted, and the adaptive recursive matrix is calculated and trained in combination with deep learning networks to achieve adaptive adjustment of hyperparameters.
It effectively improves the accuracy of load identification, reduces the difficulty of hyperparameter formulation, and improves the degree of difference between loads and intra-class similarity.
Smart Images
Figure CN114841450B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of non-intrusive load decomposition and relates to a load identification method based on an adaptive recursive matrix. Background Art
[0002] In smart grids, non-intrusive load monitoring (NILM) technology belongs to the field of advanced measurement system technology and is located at the junction of power distribution and power consumption in the power grid. Its purpose is to obtain the power consumption information of each electrical appliance within the user through smart meters, and then analyze the power consumption habits of regional users, adjust the power supply plan, and at the same time improve the recognition rate of electrical appliance faults and reduce the occurrence of electrical accidents.
[0003] A typical NILM system technology system usually consists of four major modules: data acquisition, event detection, feature extraction, and load decomposition. The main function of the data acquisition module is to collect power main electrical signals, which requires the module to be highly efficient, highly accurate, and highly anti-interference; the event detection module is to accurately identify various load start and stop events and abnormal change events; the feature extraction module is to extract or convert the internal high-dimensional features of the load event electrical signal, filter out interference information, and highlight the degree of difference between different types of load events; the load decomposition module includes two parts: load identification and power consumption decomposition. Load identification uses various traditional, machine learning, deep learning and other methods to classify and identify the extracted load features. Power decomposition is to split the power consumption curve based on the results of load identification to make the load energy consumption situation clear.
[0004] In the above system, load identification is an important part. Many scholars have proposed various identification methods for the acquired electrical signals. In recent years, thanks to the rapid development of image recognition, the extraction and conversion of electrical signals into images and the conversion of load identification into image recognition have become popular research directions in the NILM field. However, when extracting load features, the formulation of various hyperparameters has become a problem that is difficult to avoid in this type of research. Hyperparameters have a great influence on the experimental results and the universality of the methods. The hyperparameters formulated through experiments will bring certain errors in practical applications, resulting in low load identification accuracy. Summary of the invention
[0005] In view of the deficiencies in the prior art, the present invention provides a load identification method based on an adaptive recursive matrix, which can realize adaptive adjustment of hyperparameters, thereby effectively improving the accuracy of load identification.
[0006] In order to solve the above technical problems, the technical solution of the present invention is:
[0007] A load identification method based on an adaptive recursive matrix comprises the following steps:
[0008] S1. Use non-intrusive load monitoring devices to obtain user load information and use high-precision event detection algorithms to extract load events;
[0009] S2, extracting single-load single-cycle electrical signal data from the load event information;
[0010] S3, calculating the adaptive recursive matrix using the original recursive matrix;
[0011] S4, convert the load characteristics into a heat map and put it into a deep learning network for training;
[0012] S5. Build and deploy the actual environment based on the trained deep model, test it, and implement the load identification function.
[0013] Preferably, the acquired user load information includes a current sequence and a voltage sequence.
[0014] Preferably, the extracted load events include electrical signal sequence segments of load start and stop.
[0015] Preferably, the single-load single-cycle electrical signal data includes current characteristics and voltage characteristics.
[0016] Preferably, the method for extracting single-load single-cycle electrical signal data in step S2 is:
[0017] Taking the voltage sequence as the timing standard, the voltage and current timing sequences are obtained by obtaining the length W before the load starts and stops and the length W after the load starts and stops. Let L be the length of the electrical signal cycle, and the calculation formula is:
[0018] L = f s / f
[0019]
[0020] in Indicates rounding down, e is the natural exponent, f s is the sampling frequency, f is the power supply frequency,
[0021] Assume that the current and voltage sequence before start and stop is The current and voltage sequence after start and stop is: Where k∈{1,2,3} represents the electrical signal data of the kth phase; i∈{1,2,3...,W} represents the i-th value in the sequence,
[0022] Afterwards, from Extract sequence segments The sequence obtained by subtracting it is used as the processed current feature I k ; From V k Extract sequence segments As the processed voltage characteristic V k , the calculation method is as follows:
[0023]
[0024] Preferably, step S3 comprises:
[0025] S3-1, calculate the original recursive matrix;
[0026] S3-2. Based on the correlation between the voltage and current sequences, the compression ratio of the recursive matrix is obtained, and the adaptive recursive matrix is calculated.
[0027] Preferably, the calculation method of the original recursive matrix in step S3-1 is as follows:
[0028] Get current characteristics I k and voltage characteristics V k , perform characteristic conversion, using the current characteristic I k Compute the primitive recursive matrix OM k , the calculation formula is as follows:
[0029] LK k (i,j)=I k (i)-I k (j),(i,j=1,2,3,...L)
[0030] OM k (i,j)=||LK k (i,j)||
[0031] Among them, ||·|| is a norm, often L 1 Norm, L 2 norm and L∞ norm.
[0032] Preferably, in step S3-2, the method for calculating the adaptive recursive matrix is as follows:
[0033] Using the current characteristic I k and voltage characteristics V k Calculate the compression ratio of the recursive matrix and define the compression ratio of the recursive matrix as λ k , which is expressed as follows:
[0034]
[0035] in c or(X,Y) means calculating the correlation coefficient of X and Y sequences;
[0036] Then, based on the compression ratio of the recursive matrix, the original recursive matrix OM is kPerform exponential compression to obtain load characteristics and form an adaptive recursive matrix ARM k :
[0037] ARM k =OM k .
[0038] Preferably, in step S4, a Swin-transformer deep learning network is used for training and recognition, and the training method is: converting a single-phase load feature into a thermal grayscale map of size L×L×1 as an input feature of the Swin-transformer; converting each phase signal feature in the three-phase load feature into a thermal grayscale map of size L×L×1, and synthesizing these three grayscale maps corresponding to R, G, and B color channels into a single color map, that is, a thermal color map of size L×L×3, as an input feature of the Swin-transformer.
[0039] Preferably, in step S5, the trained swin-transformer is subjected to non-invasive online monitoring of the load, the electrical signal is collected by the non-invasive monitoring host, and it is determined whether a load change event occurs by an event detection algorithm; if it occurs, the electrical signal of the corresponding time period is extracted, the time point of the event occurrence is marked, and the load feature extraction as in steps S2 and S3 is performed to obtain an adaptive recursive matrix feature, which is converted into a thermal grayscale map or a thermal color map based on the phase number of the electrical signal, and input into the trained swin-transformer model for identification, thereby finally realizing the non-invasive load online monitoring function.
[0040] The present invention has the following characteristics and beneficial effects:
[0041] 1. The present invention obtains the original recursive matrix based on the electrical signal information collected by the non-intrusive monitoring host. Since a constant class compression ratio that is too small will result in a large loss of features after the recursive matrix is compressed, which is not conducive to increasing the inter-class distinction; a large compression ratio will reduce the intra-class similarity, both of which will affect the load identification accuracy. Therefore, the present invention introduces the recursive matrix compression ratio λ k As a hyperparameter, the compression ratio is adaptively adjusted based on the correlation between voltage and current, which reduces the difficulty of hyperparameter formulation. At the same time, the natural exponent is used to control the range of compression ratio variation, effectively improving the accuracy of load identification.
[0042] 2. The present invention proposes an adaptive recursive matrix based on the correlation between voltage and current, which solves the problem that hyperparameters in non-invasive load identification have a great influence on the load identification results and are difficult to formulate, and effectively improves the distinction between load classes and the similarity within classes, further improving the accuracy of load identification. Compared with traditional methods, the present invention has the advantages of simple load feature extraction, obvious distinction, and strong universality. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] 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 labor.
[0044] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention.
[0045] Figure 2 The figure is a schematic diagram showing the comparison of the ARM feature heat map presentation effects corresponding to different λ values in the embodiment of the present invention.
[0046] Figure 3 This is a schematic diagram of the effect of applying four datasets (PLAID, WHITED, LILACD, SQL) with different scenarios in an embodiment of the present invention to perform load identification after feature extraction using four methods (ARM, RP, VI, FFT), and taking three types of evaluation indicators (MCC, F1, KIA) for verification. DETAILED DESCRIPTION
[0047] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0048] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.
[0049] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.
[0050] The present invention provides a load identification method based on an adaptive recursive matrix, such as Figure 1 As shown, the following steps are included:
[0051] S1. Obtain user load information using a non-intrusive load monitoring device, wherein the acquired user load information includes a current sequence and a voltage sequence, and extract load events using a high-precision event detection algorithm, wherein the load event includes an electrical signal sequence segment of load start and stop;
[0052] It should be noted that in the field of non-intrusive load identification, event detection algorithm is used to refer to a class of algorithms. The purpose of this class of algorithms is to accurately detect changes in load start and stop and state switching at a certain moment when monitoring electrical signals. It is a conventional method in the field of non-intrusive load identification, so it will not be described in detail in this embodiment.
[0053] S2. extracting single-load single-cycle electrical signal data from the load event information, wherein the single-load single-cycle electrical signal data includes current characteristics and voltage characteristics;
[0054] Specifically, the method for extracting single-load single-cycle electrical signal data is:
[0055] Taking the voltage sequence as the timing standard, the voltage and current timing sequences are obtained by obtaining the length W before the load starts and stops and the length W after the load starts and stops. Let L be the length of the electrical signal cycle, and the calculation formula is:
[0056] L = f s / f
[0057]
[0058] in Indicates rounding down, e is the natural exponent, f s is the sampling frequency, f is the power supply frequency,
[0059] Assume that the current and voltage sequence before start and stop is The current and voltage sequence after start and stop is: Where k∈{1,2,3} represents the electrical signal data of the kth phase; i∈{1,2,3...,W} represents the i-th value in the sequence,
[0060] Afterwards, from Extract sequence segments The sequence obtained by subtracting it is used as the processed current feature I k ; From V k Extract sequence segments As the processed voltage characteristic V k , the calculation method is as follows:
[0061]
[0062] It should be noted that, since the voltage signal has obvious periodicity in the actual collected signals, the voltage sequence is used as the timing standard to ensure the accuracy of load identification.
[0063] S3. Calculate the adaptive recursive matrix using the original recursive matrix
[0064] Furthermore, the step S3 includes:
[0065] S3-1. Calculate the original recursive matrix
[0066] Specifically, the calculation method of the original recursive matrix is as follows:
[0067] Get current characteristics I k and voltage characteristics V k , perform characteristic conversion, using the current characteristic I k Compute the primitive recursive matrix OM k , the calculation formula is as follows:
[0068] LK k (i,j)=I k (i)-I k (j),(i,j=1,2,3,...L)
[0069] OM k (i,j)=||LK k (i,j)||
[0070] Among them, ||·|| is a norm, often L 1 Norm, L 2 norm and L∞ norm.
[0071] Understandable, L 1 Norm, L 2 The norm is the Euclidean norm, and the L∞ norm is the maximum value norm.
[0072] S3-2. Based on the correlation between the voltage and current sequences, the compression ratio of the recursive matrix is obtained, and the adaptive recursive matrix is calculated.
[0073] Specifically, the method for calculating the adaptive recursive matrix is as follows:
[0074] Using the current characteristic I k and voltage characteristics V k Calculate the compression ratio of the recursive matrix and define the compression ratio of the recursive matrix as λ k , which is expressed as follows:
[0075]
[0076] Where cor(X,Y) means calculating the correlation coefficient of X and Y sequences;
[0077] It is understandable that a reasonable compression ratio can effectively improve the feature differentiation of different types of loads. Figure 2 It can be seen that the technical solution provided in this embodiment ARM feature heat map presents the best effect.
[0078] Furthermore, in this embodiment, the person correlation coefficient is adopted, and the calculation formula is:
[0079]
[0080] Where m is the sequence length of X and Y.
[0081] Then, based on the compression ratio of the recursive matrix, the original recursive matrix OM is k Perform exponential compression to obtain load characteristics and form an adaptive recursive matrix ARM k :
[0082] ARM k =OM k .
[0083] S4, convert the load characteristics into a heat map and put it into a deep learning network for training;
[0084] Specifically, get the load characteristic ARM kAfter that, common deep learning and machine learning algorithms can be used for training and recognition. In this embodiment, the Swin-transformer deep learning network is used to implement this part. Swin-transformer introduces a hierarchical structure to address the shortcomings of the original transformer network architecture, which has a single feature, a quadratic relationship between the computational complexity and the image size, and a long training time. It uses local self-attention in the non-overlapping windows of the segmented image to reduce the computational complexity to a linear relationship with the image size. It not only maintains the deep feature mining of the original architecture, but also effectively reduces the training time. At the same time, it can adapt to multi-scale features and has broad prospects for development.
[0085] In order to adapt to the single-phase and three-phase electrical signal data, the single-phase load characteristics (ie ARM 1 ) is converted into a thermal grayscale image of size L×L×1 as the input feature of Swin-transformer; the three-phase load feature (i.e. ARM 1，2，3 ) is converted into a thermal grayscale map of size L×L×1, and the three grayscale maps are respectively corresponding to the R, G, and B color channels to synthesize a single color map, that is, a thermal color map of size L×L×3, which is used as the input feature of Swin-transformer.
[0086] Furthermore, in order to reduce the deviation of recognition effect caused by unreasonable data division, the cross-validation training method is adopted in this embodiment, and the original data set is randomly and evenly divided into K (K=10) groups, each subset data is used as a validation set, and the remaining K-1 groups of subset data are used as training sets to obtain K models, and the comprehensive judgment of these K models is used to finally determine the load category.
[0087] S5. Build and deploy the actual environment based on the trained deep model, test it, and implement the load identification function.
[0088] Specifically, the trained swin-transformer is used for non-invasive online monitoring of the load. The electrical signal is collected by the non-invasive monitoring host, and the event detection algorithm is used to determine whether a load change event occurs. If it occurs, the electrical signal of the corresponding time period is extracted, the time point of the event is marked, and the load feature extraction such as steps S2 and S3 is performed to obtain an adaptive recursive matrix feature, which is converted into a thermal grayscale map or a thermal color map based on the phase number of the electrical signal, and input into the trained swin-transformer model for recognition, thereby finally realizing the non-invasive load online monitoring function.
[0089] pass Figure 3 It can be seen that in the above technical solution, load identification is performed after ARM feature extraction, and the identification effect is the best.
[0090] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.
Claims
1. A load identification method based on adaptive recursive matrix, It is characterized in that The steps include: S1. Using a non-intrusive load monitoring device to obtain user load information, the obtained user load information includes a current sequence and a voltage sequence, and using an event detection algorithm to extract load events, the extracted load events include electrical signal sequence segments of load start and stop; S2. Extracting single-load single-cycle electrical signal data from the load event information, wherein the single-load single-cycle electrical signal data includes current characteristics and voltage characteristics: Taking the voltage sequence as the timing standard, the voltage and current timing sequences of W length before the load start and stop and W length after the load start and stop are obtained. Let L be the length of the electrical signal cycle, and the calculation formula is: L=f s / f in Indicates rounding down, e is the natural exponent, f s is the sampling frequency, f is the power supply frequency, Assume that the current and voltage sequence before start and stop is The current and voltage sequence after start and stop is: Where k∈{1,2,3} represents the electrical signal data of the kth phase; i∈{1,2,3...,W} represents the i-th value in the sequence. Afterwards, from Extract sequence segments The sequence obtained by subtracting it is used as the processed current feature I k ; From V k Extract sequence segments As the processed voltage characteristic V k , the calculation method is as follows: S3. Use the original recursive matrix to calculate the adaptive recursive matrix: S3-1, calculate the original recursive matrix; S3-2, based on the correlation between the voltage and current sequences, the compression ratio of the recursive matrix is obtained, and the adaptive recursive matrix is calculated; S4, convert the load characteristics into a heat map and put it into a deep learning network for training; The Swin-transformer deep learning network is used for training and recognition. The training method is: converting the single-phase load characteristics into a thermal grayscale image of size L×L×1 as the input feature of the Swin-transformer; Each phase signal feature of the three-phase load characteristics is converted into a thermal grayscale image of size L×L×1, and these three grayscale images are respectively corresponding to the R, G, and B color channels to synthesize a single color image, that is, a thermal color image of size L×L×3, which is used as the input feature of Swin-transformer; S5. Build and deploy the actual environment based on the trained deep model, test it, and implement the load identification function. The trained Swin-transformer is used for non-invasive online monitoring of the load. The electrical signal is collected by the non-invasive monitoring host, and the event detection algorithm is used to determine whether a load change event occurs. If it occurs, the electrical signal of the corresponding time period is extracted, the time point of the event is marked, and the load feature extraction as in steps S2 and S3 is performed to obtain an adaptive recursive matrix feature. The feature is converted into a thermal grayscale image or a thermal color image based on the phase number of the electrical signal, and input into the trained Swin-transformer model for recognition, thereby finally realizing the non-invasive load online monitoring function.
2. The load identification method based on the adaptive recursive matrix according to claim 1, It is characterized in that In step S3-1, the calculation method of the original recursive matrix is as follows: Get current characteristics I k and voltage characteristics V k , perform characteristic conversion, using the current characteristic I k Compute the primitive recursive matrix OM k , the calculation formula is as follows: LK k (i,j)=I k (i)-I k (j) i,j=1,2,3,…,L IF k (i,j)=||LK k (i,j)|| Among them, ||·|| is a norm, often L 1 Norm, L 2 norm and L∞ norm.
3. The load identification method based on the adaptive recursive matrix according to claim 2, It is characterized in that In step S3-2, the method for calculating the adaptive recursive matrix is as follows: Using the current characteristic I k and voltage characteristics V k Calculate the compression ratio of the recursive matrix and define the compression ratio of the recursive matrix as λ k , which is expressed as follows: in cor(X,Y) Indicates calculating the correlation coefficient of X and Y series; Then, based on the compression ratio of the recursive matrix, the original recursive matrix OM is k Perform exponential compression to obtain load characteristics and form an adaptive recursive matrix ARM k : ARM k = ABOUT k 。