Non-intrusive load feature extraction method based on clustering

Abstract

The invention provides a non-intrusive load feature extraction method based on clustering, and the method optimizes and improves a Gaussian mixture model, carries out the analysis and judgment through the feature information after clustering, selects the optimal number of Gaussian mixture components, improves the selection of an initial value, and finally outputs a feature vector. According to the method, Gaussian fine clustering is firstly carried out, then clustering features are analyzed, optimization is carried out according to the features, an initial value is improved, and typical description and extraction of feature information are realized. The method can realize self-adaption of typical description and extraction of different equipment load characteristic information, is very suitable for being applied to a non-intrusive load monitoring technology, and is also suitable for a signal analysis and intelligent equipment on-line monitoring technology.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grid automated power distribution, and in particular to a clustering-based non-intrusive load feature extraction method. Background Art

[0002] The basic requirement of smart grid is to build a new power system. The important feature of the new power system is the flexible adjustment of power load. The basis is to obtain the small-grained operating status of each power-consuming device in the power system. How to obtain the operating status information of each power-consuming device in the system is the key and difficulty in building a new power system and ubiquitous power Internet of Things.

[0003] The corresponding non-intrusive load monitoring technology (NILM) is the intrusive load monitoring technology (ILM), which is a means of load monitoring technology. Intrusive load monitoring technology (ILM) monitors the operating status of each electrical device and obtains power consumption data by installing sensors or measuring meters on each electrical device; while non-intrusive load monitoring technology (NILM) does not require the installation of sensors or measuring meters on each device, but only requires the installation of sensors or measuring meters on the bus of multiple devices, and obtains the operating status and power consumption data of each electrical device connected to the bus by analyzing the operating status and power consumption data at the bus.

[0004] Non-intrusive load monitoring technology was first proposed in the 1980s. At that time, it was limited by computer technology and computing power and did not make much progress. In recent years, with the development of computer technology and artificial intelligence technology, non-intrusive load monitoring technology has made breakthrough progress and has become a hot spot in electricity monitoring technology.

[0005] Compared with invasive load monitoring technology, non-invasive load monitoring technology does not require the installation of sensors or meters on each device, which greatly reduces the measurement cost and the difficulty of installation and measurement. This makes non-invasive load monitoring technology possible for practical promotion.

[0006] Non-intrusive load monitoring technology generally includes three aspects: data collection, load feature extraction, and load identification. Data collection is to collect the power load data of the bus and digitize the data; load feature extraction is to analyze the power load data of the bus and extract the feature information of the data. Load identification is to identify which device or devices are running and what the running status is through the extracted features.

[0007] Load feature extraction is the key content of non-intrusive load monitoring technology research. The existing non-intrusive load monitoring technology has the following problems in feature extraction:

[0008] 1. Low-frequency data is used as the research object of power load data, and the main methods used are power difference method, load fluctuation statistics method, fast Fourier transform method, etc. The power difference method is to extract features and decompose loads by comparing the differential values ​​of the power of different electrical equipment. This method has a difficult problem to solve, that is, when the load curves or load sequences of different equipment are similar, the accuracy is low. The load fluctuation statistics method counts the duty cycle of the power curve when the equipment is in working state. This method cannot accurately identify the type of power load with similar power curve characteristics but different amplitudes. In addition, the real-time performance is not high, and a longer statistical time of the power load power curve is required to relatively accurately identify the load. The fast Fourier transform method converts the load curve of the electrical equipment into the frequency domain for analysis. This method has poor real-time performance and is easily interfered.

[0009] 2. High-frequency data is used as the research object of power load data, and the main methods used are VI trajectory method, wavelet multi-resolution analysis method, convolutional neural network method, etc. The VI trajectory method has a large amount of computation for feature information extraction and is easily interfered by noise, resulting in poor accuracy. The wavelet multi-resolution analysis method also has a large amount of computation and is easily interfered by noise. The convolutional neural network training method has a large amount of computation and a long training time. Summary of the invention

[0010] In view of the problems existing in the above methods, the scheme of the present invention introduces a Gaussian mixture model of machine learning according to the basic characteristics of the power load, and proposes a method for extracting the characteristics of the power load based on an improved Gaussian mixture model (improved GMM). Based on the Gaussian mixture model, this method proposes to analyze and judge the features after fine clustering in advance, select the optimal number of Gaussian mixture components, improve the selection of initial values, and finally output the feature vector. This method has a small amount of calculation, high real-time performance, and can extract feature information of different types of power loads adaptively.

[0011] A non-intrusive load feature extraction method based on clustering includes the following steps:

[0012] Step S1: Preliminarily determine the number of mixed components;

[0013] Step S2: obtaining initial values ​​of mean and variance;

[0014] Step S3: Gaussian mixture model initial value;

[0015] Step S4: Gaussian mixture clustering;

[0016] Step S5: mean, variance and mixing ratio of each component;

[0017] Step S6: Mixture component analysis and optimization.

[0018] Preferably, in step S1 of the present invention, the number of mixed components is initially determined. Specifically, according to the power load characteristics, the number of input mixed components k≥8 is initially determined.

[0019] Preferably, the acquisition of the initial values ​​of the mean and variance in step S2 of the present invention is specifically:

[0020] Step S21, randomly selecting a data from the load data as the cluster center d1;

[0021] Step S22, calculate the distance from all load data to the data, named D(x), where x refers to each sample data point set;

[0022] Step S23, calculating the probability of each sample in all load data being selected as the next cluster center;

[0023]

[0024] Step S24, select the value with the highest probability as the next cluster center;

[0025] Step S25, repeat steps S2-S4 until k cluster centers are selected;

[0026] Step S26: Based on the obtained k cluster centers, calculate the sum of the distances from each point in each cluster to the cluster center.

[0027] Preferably, the initial value of the Gaussian mixture model in step S3 of the present invention is: the data values ​​of the k cluster centers obtained are used as the initial mean of the Gaussian mixture model; the initial covariance of the Gaussian mixture model is calculated by the sum of the distances from each point in each class to the cluster center; and the mixing ratio is set to 1 / k.

[0028] Preferably, the Gaussian mixture clustering in step S4 of the present invention is specifically:

[0029] Step S41, using the initial value obtained previously as a Gaussian probability model for training;

[0030] Step S42, calculating the probability that each data point of the power load data belongs to each Gaussian model, that is, the posterior probability; the closer the data point is to the center of the Gaussian distribution, the greater the probability, that is, the higher the possibility of belonging to the Gaussian model;

[0031] Step S43: Calculate the new (μ i ',σ i ') parameters maximize the probability of the data point; these new parameters are calculated using the weighted probabilities of the data points, where the weight is the probability that the data point belongs to the Gaussian model;

[0032] Step S44: Substitute the mean, variance and mixed ratio model parameters {((μ i ,σ i ,p i )|1≤i≤k)} is updated to {((μ i ',σ i ',p i ')|1≤i≤k)}, repeat the iterative steps S42 and S43 until convergence.

[0033] Preferably, the mean, variance and mixing ratio of each component in step S5 of the present invention are specifically: after Gaussian clustering, characteristic information of each component of the power load data curve is obtained, that is, the mean, variance and mixing ratio ((μ i ,σ i ,p i )|1≤i≤k).

[0034] Preferably, the mixed component analysis and optimization in step S6 of the present invention are specifically: based on the overall characteristics of the power load data curve, the components obtained through the above steps are analyzed and optimized; according to the changes in the power load data curve, the data are divided into four parts: low value area, transition area, high value area, and oscillation area; the low value area reflects the working preparation state or working intermittent state of the power equipment; the transition area represents the transition process of the power equipment from the preparation or intermittent state to the working state; the high value area represents the amplitude of the power consumption curve of the power equipment when it is working; the oscillation area represents the stability of the power consumption curve of the power equipment when it is working.

[0035] The present invention proposes a non-intrusive load feature extraction method based on clustering, which optimizes and improves the Gaussian mixture model, analyzes and judges the characteristic information after clustering in advance, selects the optimal number of Gaussian mixture components, thereby improving the selection of initial values, and finally outputs the characteristic vector. That is, Gaussian fine clustering is first performed, then the clustering characteristics are analyzed, optimized according to the characteristics, the initial value is improved, and the typical characterization and extraction of characteristic information are realized.

[0036] This method has small computational complexity, good real-time performance and low implementation cost. At the same time, this method can realize the typical characterization and extraction of load characteristic information of different equipment, which is very suitable for non-intrusive load monitoring technology. At the same time, this method is also suitable for signal analysis and online monitoring technology of intelligent equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the improved Gaussian mixture model algorithm flow of the present invention.

[0038] Figure 2 It is a schematic diagram of the power load curve of Example 1 of the present invention.

[0039] Figure 3 It is a schematic diagram of curve characteristics after optimized clustering obtained in Example 1 of the present invention. DETAILED DESCRIPTION

[0040] like Figure 1 As shown, a clustering-based non-intrusive load feature extraction method includes the following steps:

[0041] Step S1: Preliminarily determine the number of mixed components; according to the power load characteristics, preliminarily determine the number of input mixed components k≥8.

[0042] Step S2: Obtaining the initial values ​​of mean and variance; specifically:

[0043] Step S21, randomly selecting a data from the load data as the cluster center d1;

[0044] Step S22, calculate the distance from all load data to the data, named D(x), where x refers to each sample data point set;

[0045] Step S23, calculating the probability of each sample in all load data being selected as the next cluster center;

[0046]

[0047] Step S24, select the value with the highest probability as the next cluster center;

[0048] Step S25, repeat steps S2-S4 until k cluster centers are selected;

[0049] Step S26: Based on the obtained k cluster centers, calculate the sum of the distances from each point in each cluster to the cluster center.

[0050] Step S3: Initial value of Gaussian mixture model; the data values ​​of the k cluster centers are taken as the initial mean of the Gaussian mixture model; the initial covariance of the Gaussian mixture model is calculated by the sum of the distances from each point in each class to the cluster center; the mixing ratio is set to 1 / k.

[0051] Step S4: Gaussian mixture clustering; specifically:

[0052] Step S41, using the initial value obtained previously as a Gaussian probability model for training;

[0053] Step S42, calculating the probability that each data point of the power load data belongs to each Gaussian model, that is, the posterior probability; the closer the data point is to the center of the Gaussian distribution, the greater the probability, that is, the higher the possibility of belonging to the Gaussian model;

[0054] Step S43: Calculate the new (μ i ',σi ') parameters maximize the probability of the data point; these new parameters are calculated using the weighted probabilities of the data points, where the weight is the probability that the data point belongs to the Gaussian model;

[0055] Step S44: Substitute the mean, variance and mixed ratio model parameters {((μ i ,σ i ,p i )|1≤i≤k)} is updated to {((μ i ',σ i ',p i ')|1≤i≤k)}, repeat the iterative steps S42 and S43 until convergence.

[0056] Step S5: The mean, variance and mixing ratio of each component; after Gaussian clustering, the characteristic information of each component of the power load data curve is obtained, namely, the mean, variance and mixing ratio ((μ i ,σ i ,p i )|1≤i≤k).

[0057] Step S6: mixed component analysis and optimization; according to the overall characteristics of the power load data curve, analyze and optimize the components obtained through the above steps; according to the changes in the power load data curve, divide the data into four parts: low value area, transition area, high value area, and oscillation area; the low value area reflects the working preparation state or working intermittent state of the power equipment; the transition area represents the transition process of the power equipment from the preparation or intermittent state to the working state; the high value area represents the amplitude of the power consumption curve of the power equipment when it is working; the oscillation area represents the stability of the power consumption curve of the power equipment when it is working.

[0058] Embodiment 1:

[0059] Combination Figure 2 The washing machine power load curve shown in the figure has the initial number of mixed components set to k=8.

[0060] The initial values ​​obtained by k-means++ clustering according to the 8 mixed components are shown in Table 1:

[0061] Table 1 k-means++ clustering initial values

[0062] k <![CDATA[μ i ]]> <![CDATA[σ i ]]> Components 1 68.9517 1.7869 0.125 2 84.2765 1.6078 0.125 3 4.4029 5.8573 0.125 4 31.0370 32.0351 0.125 5 76.3607 3.2797 0.125 6 80.4383 2.6527 0.125 7 56.4883 1.2682 0.125 8 73.9305 29.5791 0.125

[0063] After Gaussian mixture, 8 feature data are obtained as shown in Table 2:

[0064] Table 2 Initial values ​​of Gaussian mixture clustering

[0065] k <![CDATA[μ i ]]> <![CDATA[σ i ]]> Components 1 68.9721 38.9117 0.0022 2 84.1944 1.6805 0.1119 3 4.4029 5.8577 0.1075 4 31.0354 32.0120 0.0559 5 76.6720 4.8580 0.1828 6 80.5787 2.1960 0.2912 7 69.1575 2.1804 0.0321 8 74.6780 30.4261 0.2163

[0066] Gaussian clustering is optimized. The mean represents the size of the data; the variance represents the stability of the data. The larger the variance, the more oscillating the curve, and the smaller the variance, the more stable the data. Therefore, the large variance data in the high value part can be classified as the oscillation area, and the small variance data in the high value part can be classified as the high value area. The median part is classified as the transition area, and the low value part is classified as the low value area. The initial value after optimization is:

[0067] Table 3 Optimized initial values

[0068] k <![CDATA[μ i ]]> <![CDATA[σ i ]]> Components 1 80.1037 10.1097 0.5882 High value area 2 73.9305 29.5791 0.2163 Shock Zone 3 31.0395 32.1692 0.0880 Transition Zone 4 4.4029 5.8573 0.1075 Low value area

[0069] Re-cluster the optimized initial values ​​in Table 3 and obtain Figure 3 The curve characteristics after optimized clustering are shown, and then the characteristic matrix is ​​obtained:

[0070]

[0071] The present invention proposes a non-intrusive load feature extraction method based on clustering, which is a method for obtaining feature vectors by clustering through an improved Gaussian mixture model. The method optimizes and improves the Gaussian mixture model, analyzes and judges the feature information after clustering in advance, selects the optimal number of Gaussian mixture components, thereby improving the selection of initial values, and finally outputs feature vectors. That is, fine clustering is first performed, then the clustering features are analyzed, optimization is performed according to the features, and the initial values ​​are improved, thereby ensuring the typicality of the feature information.

[0072] The present invention proposes a principle for extracting feature information based on the distribution characteristics of the power load time domain curve, and innovatively proposes the regional distribution of the power load time domain curve as an optimization principle. It includes but is not limited to applications in the fields of non-intrusive load monitoring, online monitoring technology of intelligent devices, and signal analysis.

Claims

1. A non-intrusive load feature extraction method based on clustering, characterized in that: The steps include: Step S1: Preliminarily determine the number of mixed components; Step S2: obtaining initial values ​​of mean and variance; Step S3: Gaussian mixture model initial value; Step S4: Gaussian mixture clustering; Step S5: mean, variance and mixing ratio of each component; Step S6: Mixture component analysis and optimization.

2. The clustering-based non-intrusive load feature extraction method according to claim 1, characterized in that: In the above step S1, the number of mixed components is initially determined. Specifically, according to the power load characteristics, the number of input mixed components k≥8 is initially determined.

3. The clustering-based non-intrusive load feature extraction method according to claim 2, characterized in that: The acquisition of the initial values ​​of the mean and variance in the above step S2 is specifically: Step S21, randomly selecting a data from the load data as the cluster center d1; Step S22, calculate the distance from all load data to the data, named D(x), where x refers to each sample data point set; Step S23, calculating the probability of each sample in all load data being selected as the next cluster center; Step S24, select the value with the highest probability as the next cluster center; Step S25, repeat steps S2-S4 until k cluster centers are selected; Step S26: Based on the obtained k cluster centers, calculate the sum of the distances from each point in each cluster to the cluster center.

4. The clustering-based non-intrusive load feature extraction method according to claim 3, characterized in that: The initial value of the Gaussian mixture model in the above step S3 is specifically: the data values ​​of the k cluster centers are obtained as the initial mean of the Gaussian mixture model; the initial covariance of the Gaussian mixture model is calculated by the sum of the distances from each point in each class to the cluster center; and the mixing ratio is set to 1 / k.

5. The clustering-based non-intrusive load feature extraction method according to claim 4, characterized in that: The Gaussian mixture clustering in step S4 above is as follows: Step S41, using the initial value obtained previously as a Gaussian probability model for training; Step S42, calculating the probability that each data point of the power load data belongs to each Gaussian model, that is, the posterior probability; The closer the data point is to the center of the Gaussian distribution, the greater the probability, that is, the higher the possibility of belonging to the Gaussian model; Step S43: Calculate the new (μ i ',σ i ') parameters maximize the probability of the data point; these new parameters are calculated using the weighted probabilities of the data points, where the weight is the probability that the data point belongs to the Gaussian model; Step S44: Substitute the mean, variance and mixed ratio model parameters {((μ i ,σ i ,p i )|1≤i≤k)} is updated to {((μ i ',σ i ',p i ')|1≤i≤k)}, repeat the iterative steps S42 and S43 until convergence; the mean, variance and mixing ratio are obtained by training the Gaussian probability model using the data values ​​of the k cluster centers as the initial values ​​of the Gaussian mixture model.

6. The clustering-based non-intrusive load feature extraction method according to claim 5, characterized in that: The mean, variance and mixing ratio of each component in the above step S5 are specifically: after Gaussian clustering, characteristic information of each component of the power load data curve, namely, the mean, variance and mixing ratio, is obtained.

7. The clustering-based non-intrusive load feature extraction method according to claim 6, characterized in that: The mixed component analysis and optimization in the above step S6 is specifically: according to the overall characteristics of the power load data curve, the components obtained through the above steps are analyzed and optimized; according to the change of the power load data curve, the data is divided into four parts: low value area, transition area, high value area, and oscillation area; The low value area reflects the working preparation state or working intermittent state of the power equipment; the transition area represents the transition process of the power equipment from the preparation or intermittent state to the working state; The high value area represents the amplitude of the power consumption curve when the power equipment is working; the oscillation area represents the stability of the power consumption curve when the power equipment is working.

Images