Load state identification method based on hierarchical clustering and image feature set
Through the load state recognition method based on hierarchical clustering and image feature sets, the problem of operating state recognition of complex power equipment is solved, and the accuracy of load prediction and demand response potential calculation is improved.
Patent Information
- Application Number
- CN202411589373.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-08
- Publication Date
- 2025-02-18
AI Technical Summary
The prior art is difficult to accurately identify the operating status of complex electrical equipment, especially when the equipment has multiple operating modes, resulting in errors in load prediction and demand response potential calculations.
The load state recognition method based on hierarchical clustering and image feature set is adopted, and the device state recognition is achieved by establishing a typical curve library of equipment, clustering typical operating conditions based on DTW, constructing a load feature set and using a convolutional neural network to identify the device state.
Improve the accuracy of identification of device status combinations in total power data, and enhance the accuracy of load prediction and demand response potential calculation.
Smart Images
Figure CN119474745A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of equipment state recognition, and in particular to a load state recognition method based on hierarchical clustering and image feature sets. Background Art
[0002] Under the new power system, the basis for flexible load forecasting and demand response participation on the multi-user side is accurate load modeling. With the improvement of electrification in various industries and the large-scale application of power electronic equipment, the operating status of power equipment has become more and more complex, and it is difficult to accurately describe it through mathematical methods.
[0003] Most of the existing typical load curves are obtained by selecting certain features from the daily load curves of users over a period of time through clustering methods such as k-means. Load feature selection generally includes transient features and steady-state features. Transient features are parameter features that change when electrical equipment is switched on or off or changes its working state, and steady-state features are parameter features such as power when the equipment is running. Steady-state feature recognition is generally achieved through machine learning, while transient feature recognition is achieved by extracting features from the total power data when a transient event occurs.
[0004] Conventional partition-based and density-based clustering methods have the problem of error propagation. When class B is between class A and class C and the difference between A and C is large, the three will be mistakenly combined into one; when the sample distribution difference is not large and the number of potential classes is large, the error will be propagated. The device state is only defined as one of the two switches, and it cannot accurately represent devices with multiple operating modes. The operating state of electrical equipment is becoming more and more complex, and mathematical methods are not enough to accurately describe it.
[0005] Therefore, a load state recognition method based on hierarchical clustering and image feature sets is needed to identify the equipment operating status. Summary of the invention
[0006] The purpose of the present invention is to propose a load state recognition method based on hierarchical clustering and image feature set, comprising the following steps:
[0007] Establish a typical curve library for equipment;
[0008] Based on DTW, typical working conditions of deterministic logic type equipment and random logic type equipment are clustered;
[0009] Construct load feature sets and perform equipment status identification.
[0010] The specific steps to establish a typical equipment curve library are as follows:
[0011] First, the missing values of the equipment's operating data are filled and outliers are preprocessed. Then, the power segments of each equipment operation are obtained by segmentation. The DTW matrix and energy consumed between all segments are calculated. Finally, based on whether the equipment has obvious operating logic, a hierarchical clustering or probability model is used to generate a device model library.
[0012] The formula for determining the typical curve for a logic type device is:
[0013]
[0014] In the formula, represents the DTW distance between curves i and j belonging to the same cluster, W sum (i) represents the sum of the DTW distances between the i-th curve and all other similar curves.
[0015] The specific steps to construct a load feature set and identify equipment status are as follows:
[0016] Firstly, the total power curve is decomposed by variational mode, and then the load feature set is generated by the Gram angle field and the relative position matrix. Finally, the load state is identified by convolutional neural network.
[0017] The beneficial effects of the present invention are:
[0018] The present invention establishes an equipment model library, decomposes the total power curve through variational modal decomposition, and then constructs a load feature set based on a graphical method, which can improve the accuracy of identifying equipment state combinations in total power data and facilitate subsequent load forecasting and calculation of demand response potential. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is a flow chart of the load state recognition method based on hierarchical clustering and image feature set;
[0020] Figure 2 Establish flow chart for typical curve library of equipment;
[0021] Figure 3 This is a schematic diagram of agglomerative hierarchical clustering;
[0022] Figure 4 This is the DTW clustering result diagram;
[0023] Figure 5 The probability distribution diagram of the time a TV is turned on;
[0024] Figure 6 Figure 1 shows the load feature set construction and identification method based on VMD and visualization method;
[0025] Figure 7 This is a comparison chart of GAF of each device at different times;
[0026] Figure 8 It is the relative position matrix diagram of Toaster and Dishwasher;
[0027] Fig. 9 Build a method diagram for a set of load characteristics. DETAILED DESCRIPTION
[0028] The present invention proposes a load state recognition method based on hierarchical clustering and image feature sets, which will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0029] Figure 1 The flowchart of the load state recognition method based on hierarchical clustering and image feature set is as follows:
[0030] Step 1.1 Establish a typical curve library for equipment
[0031] The process of establishing a typical curve library for equipment is as follows: Figure 2 As shown in the figure. First, the operation data of the equipment is preprocessed and segmented to obtain the power segments of each operation of the equipment. The DTW matrix and energy consumed between all segments are calculated and used as the clustering standard. Finally, according to whether the equipment has obvious operation logic, different processing methods are used to generate the equipment model library.
[0032] Step 1.2 Clustering of typical working conditions based on DTW
[0033] Dynamic Time Warping (DTW) is an algorithm for time series data analysis. It is usually used to compare the similarity between two time series data. It uses an event warping function that meets certain conditions to describe the event correspondence between the test template and the reference template, and solves the integer function corresponding to the minimum cumulative distance when matching templates. The DTW algorithm finds the optimal match between two time series by nonlinearly aligning them. DTW can overcome the changes in length and speed in different time series, so that two time series can be better aligned in the time dimension, so as to more accurately calculate the similarity between the sequences.
[0034] Figure 3 The figure shows a schematic diagram of agglomerative hierarchical clustering. Initially, each sample point is regarded as a separate class. Each time, the two closest classes are merged to generate a new class according to certain criteria. After multiple iterations, all objects belong to the same class. It can effectively solve the chain effect problem existing in partition-based and density-based clustering methods.
[0035] (1) Determine the logical type
[0036] For equipment with specific working modes and workflows, each operation will have a relatively certain power curve. Select an appropriate number as the typical operating state number.
[0037] Most of the power segments belong to Cluster 2, so we only need to consider the typical segments belonging to Cluster 2, and take the most representative curve as the "typical curve". The typical curve is determined as shown in the following formula:
[0038]
[0039] In formula (1), represents the DTW distance between curves i and j belonging to the same cluster, W sum (i) represents the sum of the DTW distances between the i-th curve and all other similar curves, and selects W within the cluster. sum The smallest curve is taken as the “typical curve”. Figure 4 This is the DTW clustering result diagram.
[0040] (2) Random logic type
[0041] Compared with the curve with deterministic logic, the power of some equipment is basically unchanged during operation, but the operation time varies greatly. The operation process of such equipment is difficult to describe by one or several deterministic curves. The typical curve of this type of equipment can be described as a function of power and duration, as shown in the following formula:
[0042] X(t)=f(P,T) (2)
[0043] In formula (2), X(t) represents a power sequence, P represents the power value of the power sequence, and T represents the duration of the sequence. For such equipment with distributed operation time, the most likely curve length can be selected through probability distribution as the "typical curve" of the equipment operation.
[0044] Figure 5 It is the probability distribution of the duration of a TV being turned on, and is fitted using Kernel distribution, which estimates the density by placing a kernel function (usually a Gaussian kernel function) window around each data point and calculating the weighted average of these kernel functions at each point.
[0045] The formula for kernel density estimation is as follows:
[0046]
[0047] In formula (3), is the estimated probability density at location x, n is the sample size, and x i is the i-th observation. Kh is the Epanechnikov kernel function.
[0048] Step 1.3 Load feature set construction and equipment status identification
[0049] First, the total power curve of a certain length is subjected to variational mode decomposition (VMD), and then the load feature set is generated through the Gram Angular Fields (GAF) and relative position matrix (RPM) visualization method, and then the load state is recognized through the convolutional neural network. The overall flow chart is as follows Figure 6 The following is an explanation of each part:
[0050] (1) Gram Corner
[0051] Gram Angular Field is an algorithm that converts one-dimensional time series data into two-dimensional images. This method not only increases the sequence dimension, but also effectively obtains the amplitude and position information of each point in the sequence data. It can also mine important information hidden in the time series and eliminate signal redundancy. Gram Angular Summation Fields (GASF) and Gram Angular Difference Fields (GADF) are collectively referred to as Gram Angular Fields.
[0052] (2) Relative position matrix
[0053] The relative position matrix is a method to convert one-dimensional data into a two-dimensional image. The image generated by RPM contains the redundant features of the original time series, making it easier to capture the similarity information between and within classes in the converted image.
[0054] (3) Variational mode decomposition
[0055] The basic idea of VMD is to decompose the signal into multiple frequency components and find the best frequency component through optimization methods. The load curve is a typical nonlinear and non-stationary signal, and the process of equipment state change reflected in the load curve is the signal mutation process. VMD can specify the number of modal decompositions in advance, which is conducive to building a unified scale data set and training a general model.
[0056] (4) Graph-based load feature set
[0057] The purpose of establishing a representative load fingerprint that can accurately identify the load under different total power data is to identify the type of equipment whose operating status has changed in the monitoring area.
[0058] Figure 7 The following are GAF images of several electrical devices of a certain user at different state transition times. Usually, GAF images can well reflect the equipment characteristics contained in the process of total power change. Each change in total power corresponds to a typical state switch. By learning a large number of GAF images at state switching times through neural networks, it is possible to extract load characteristics and identify states. However, in some cases, such as Figure 7 The images of Toaster and Dish Washer are highly similar, which will cause inconvenience in recognition. Figure 8 The figure shows the relative position matrix of Toaster and Dishwasher. It can be seen that while GAF cannot distinguish them, the RPM is quite different.
[0059] (5) Constructing load feature set
[0060] Based on the fusion of GAF and RPM image load fingerprint, the load state is identified through convolutional neural network.
[0061] Fig. 9 The figure shows the load characteristic set construction method. The time series X = (x1, x2, ..., x t ),t=1,2,...N, where the state transition of the target device occurs at (t s1 ,t s2 ) period. In reality, the conversion moment may be at any position in the collected data segment. s1 ,t s2 )∈(t,t+window), it is considered that the device has undergone a state change within the current window length. Each VMD component is converted into a corresponding GAF and RPM image for corresponding annotation, which constitutes the load feature set of the current period.
[0062] The present invention establishes an equipment model library, decomposes the total power curve through variational modal decomposition, and then constructs a load feature set based on a graphical method, which can improve the accuracy of identifying equipment state combinations in total power data and facilitate subsequent load forecasting and calculation of demand response potential.
Claims
1. A load state recognition method based on hierarchical clustering and image feature set, characterized in that: The following steps are involved: Establish a typical curve library for equipment; Based on DTW, typical working conditions of deterministic logic type equipment and random logic type equipment are clustered; Construct load feature sets and perform equipment status identification.
2. The load state recognition method based on hierarchical clustering and image feature set according to claim 1 is characterized in that: The specific steps to establish a typical equipment curve library are as follows: First, the missing values of the equipment's operating data are filled and outliers are preprocessed. Then, the power segments of each equipment operation are obtained by segmentation. The DTW matrix and energy consumed between all segments are calculated. Finally, based on whether the equipment has obvious operating logic, a hierarchical clustering or probability model is used to generate a device model library.
3. The load state recognition method based on hierarchical clustering and image feature set according to claim 1 or 2, characterized in that: The formula for determining the typical curve of the logic type device is: In the formula, represents the DTW distance between curves i and j belonging to the same cluster, W sum (i) represents the sum of the DTW distances between the i-th curve and all other similar curves.
4. The load state recognition method based on hierarchical clustering and image feature set according to claim 3 is characterized in that: The specific steps to construct a load feature set and identify equipment status are as follows: Firstly, the total power curve is decomposed by variational mode, and then the load feature set is generated by the Gram angle field and the relative position matrix. Finally, the load state is identified by convolutional neural network.