Feature extraction and load decomposition method based on improved GMM model
By decomposing and extracting load data using an improved GMM model, the problems of accurate feature extraction and load type identification when multiple electrical devices are running simultaneously are solved, achieving efficient and real-time load monitoring.
Patent Information
- Application Number
- CN202511131564.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-13
- Publication Date
- 2025-11-25
AI Technical Summary
Existing non-intrusive load monitoring technologies suffer from low accuracy in feature extraction and load type identification, high computational load, and poor real-time performance when multiple electrical devices are running simultaneously.
An improved Gaussian mixture model (GMM) was used to decompose and extract features from the load data. The load types of various electrical equipment were identified by power combination table, Spearman correlation coefficient test and rank sum test.
It achieves efficient feature extraction and accurate load type identification when multiple electrical devices are running simultaneously, with low computational load and high real-time performance.
Smart Images
Figure CN121009342A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and in particular to a method for feature extraction and load decomposition based on an improved GMM model. Background Technology
[0002] With the vigorous development of smart grids, the core is the construction of smart grids and the energy internet. Smart grids will enable efficient power transmission and distribution, improving grid stability and reliability. The energy internet will interconnect production and consumption ends, achieving optimized energy allocation and efficient utilization, promoting fair and orderly market competition, and reducing the costs of energy production, transmission, distribution, and use. On the consumption end, load monitoring is key to realizing the smart interconnection of the distribution network. There are two technical means of load monitoring: intrusive load monitoring and non-intrusive load monitoring.
[0003] Intrusive load monitoring, due to its numerous monitoring nodes, high communication requirements, and significant data storage costs, is prohibitively expensive, hindering its development towards more refined load monitoring. Non-intrusive load monitoring technology, proposed by Hart GW of MIT in the 1980s, only requires monitoring total electricity consumption data. Analysis of this data reveals detailed load information for each internal electrical device, offering lower costs and easier implementation. However, because non-intrusive load monitoring relies on machine learning algorithms, and computer technology was still in its infancy at the time, its effectiveness was limited, thus it did not attract much attention. Since the 1990s, advancements in computer technology and updates to machine learning and deep learning algorithms have made non-intrusive load monitoring a reality, achieving significant algorithmic breakthroughs and becoming a hot topic in the field of smart grid electricity metering.
[0004] The implementation of non-intrusive load monitoring technology typically involves four steps: data acquisition and preprocessing, load event detection, load data decomposition and feature extraction, and load type identification. Among these, load data decomposition and feature extraction are the core of non-intrusive load monitoring technology.
[0005] The current load data decomposition and feature extraction methods mainly include power difference method, load fluctuation statistics method, Fourier transform method, wavelet analysis method, VI trajectory method, neural network method, etc.
[0006] The power difference method and the load fluctuation statistical method are methods proposed based on the characteristics of the load curve. They mainly use the power difference or the number of fluctuations as feature information for feature extraction. Both methods have a common problem: insufficient feature information extraction, resulting in low recognition accuracy.
[0007] Fourier transform and wavelet analysis are frequency domain analysis methods. They mainly extract features by converting time-domain load data into frequency energy bands as feature information. These two methods are greatly affected by power grid load fluctuations and noise, and their real-time performance is poor due to the large amount of computation required for time-frequency conversion.
[0008] The VI trajectory method uses the closed trajectory pattern formed by current and voltage during equipment operation as feature information for extraction. However, the feature information of this method is affected by power grid load fluctuations and noise, resulting in poor stability and low recognition accuracy.
[0009] Neural network methods have been increasingly used in recent years. However, due to their large computational load, the learning and recognition time of the model is long, resulting in poor real-time performance.
[0010] The methods described above are generally effective for feature extraction and identification of a single load. However, when multiple electrical devices are operating simultaneously, feature extraction and load type identification become significantly more challenging. Summary of the Invention
[0011] To address the problems existing in the prior art, this invention proposes a feature extraction and load decomposition method based on an improved Gaussian Mixture Model (GMM). This method, through improvements to the GMM model, enables the decomposition of load data from multiple electrical devices operating simultaneously, extracts feature information, and accurately identifies the types of electrical devices by comparing the correlation and differences of the feature information. Furthermore, this method has low computational complexity, high real-time performance, and is particularly suitable for situations where multiple electrical devices operate simultaneously.
[0012] A feature extraction and load decomposition method based on an improved GMM model includes the following steps:
[0013] Step S1: Generate a power combination table;
[0014] Step S2: Obtain the feature vector of load data samples when the equipment is running alone;
[0015] Step S3: Load decomposition based on power combination lookup table method;
[0016] Step S4: Perform fitting and clustering based on Gaussian Mixture Model (GMM) on all possible load data sequences obtained after decomposition to obtain feature vectors of all possible decomposed load data.
[0017] Step S5: Perform linear correlation and difference tests on all possible decomposed load data feature vectors obtained in Step S4 and load data sample feature vectors of each device running individually, and obtain the Spearman correlation coefficient test r value and rank sum test p value, until all possibilities are exhausted;
[0018] Step S6: Identify the load type based on linear correlation and differential analysis.
[0019] A feature extraction and load decomposition method based on an improved Gaussian Mixture Model (GMM) is proposed. This method collects load data from all individual electrical devices and calculates the arithmetic mean to obtain a power combination table for all devices. The GMM model is then used to fit and cluster the load data from individual device operation to obtain sample feature vectors and initial values for fitting and clustering multiple loads. Multiple load data from simultaneous operation of multiple devices are collected, and load decomposition is performed according to the power combination table to obtain all possible load decomposition power curves. Furthermore, the GMM model is used to fit and cluster all possible load decomposition power curves to obtain all possible load decomposition power feature vectors. Correlation and difference analyses are performed on all possible load decomposition power feature vectors with the sample feature vectors from individual device operation. The Spearman correlation coefficient test is used for correlation analysis, and the rank-sum test is used for difference analysis. The combination of Spearman correlation coefficient (r-value) and rank-sum test (p-value) values both greater than 0.8 is identified; the resulting load decomposition power curve represents the power curve of that electrical device, and the power combination at this point represents the power combination of the multiple loads operating simultaneously.
[0020] This invention proposes a combined method and principle for load decomposition based on power combination and feature information extraction using a GMM model.
[0021] This invention also proposes a similarity analysis method and principle based on correlation analysis and difference analysis of feature information.
[0022] The load decomposition and feature information extraction method of the present invention includes, but is not limited to, applications in the fields of non-intrusive load monitoring and online monitoring of intelligent devices. Attached Figure Description
[0023] Figure 1 This is a flowchart illustrating the feature extraction and load decomposition method based on the improved GMM model of this invention.
[0024] Figure 2 This is a schematic representation of the power combination when the four electrical devices are operating individually.
[0025] Figure 3 This is a schematic diagram of power data curves when multiple loads are running simultaneously.
[0026] Figure 4 It is P 分解1,1 A schematic diagram of the GMM clustering results of the power curve.
[0027] Figure 5 It is P 分解2,1A schematic diagram of the GMM clustering results of the power curve.
[0028] Figure 6 It is P 分解2,2 A schematic diagram of the GMM clustering results of the power curve.
[0029] Figure 7 It is P 分解3,1 A schematic diagram of the GMM clustering results of the power curve.
[0030] Figure 8 It is P 分解3,2 A schematic diagram of the GMM clustering results of the power curve.
[0031] Figure 9 It is P 分解3,3 A schematic diagram of the GMM clustering results of the power curve. Detailed Implementation
[0032] The technical solution of the present invention will be described in detail below with reference to embodiments:
[0033] like Figure 1 As shown, a feature extraction and load decomposition method based on an improved GMM model includes the following steps:
[0034] Step 1: Create a power combination table. Collect load data of different electrical devices operating individually (assuming there are n types of electrical devices), perform smoothing and filtering processing, and calculate the arithmetic mean of the stable operation of each type of electrical device. Create a power combination table for 1 to n different devices.
[0035] Step 2: Obtain the feature vector of load data samples when the equipment is running individually. Perform Gaussian Mixture Model (GMM) fitting and clustering on the smoothed and filtered load data of different electrical devices running individually. The number of Gaussian mixture components is m, and the mean and variance are randomly initialized to obtain the mean of the m Gaussian mixture components as the sample feature vector. Save the mean, variance, and number of mixture components for each Gaussian mixture component as the initial values for the next step of the GMM model.
[0036] Step 3: Load decomposition based on power combination lookup table method. Collect load data of different electrical devices running simultaneously. Based on the values in the power combination table in Step 1, decompose the load data of different electrical devices running simultaneously to obtain all possible load data sequences after decomposition.
[0037] Step 4: Perform Gaussian Mixture Model (GMM)-based fitting and clustering on all possible load data sequences obtained after decomposition to obtain feature vectors of all possible decomposed load data.
[0038] Step 5: Perform linear correlation (Spearman correlation coefficient test) and difference (rank sum test) tests on all possible decomposed load data feature vectors obtained in the previous step and load data sample feature vectors of each device running individually, to obtain the Spearman correlation coefficient test r value and the rank sum test p value, until all possibilities are obtained.
[0039] Step 6: Identify load types based on linear correlation and difference analysis. Using the Spearman correlation coefficient test (r-value) and rank-sum test (p-value) obtained in the previous step, if both r-value and p-value are greater than 0.8, the decomposed load data feature vector is considered similar to the sample feature vector. In this case, the load type put into operation is the electrical equipment corresponding to the sample feature vector.
[0040] An application for feature extraction and load decomposition of load data when three electrical devices (air conditioner, desk lamp, and monitor) are running simultaneously:
[0041] Step 1: Collect load data for four types of electrical equipment operating individually: air conditioner fan, desk lamp, monitor, and television. This application uses power data as the research object to generate a power combination table as follows: Figure 2 As shown.
[0042] Step 2: Perform GMM-based clustering on the power data of the four types of electrical devices when they are running individually. Apply Gaussian mixture components with 8 components, and use the mean of each Gaussian mixture component as the feature value for each device. The sample feature vectors of the four devices are shown in Table 1.
[0043] Table 1. Feature Vector Table of Four Types of Electrical Equipment
[0044]
[0045]
[0046] Step 3: Collect multi-load data of three electrical devices operating simultaneously: air conditioner fan, desk lamp, and monitor. The power data curve is shown below. Figure 3 As shown:
[0047] Based on the power data of 53W at the load decomposition point, and according to the power combination table formed in the first step, there are three possible combinations of multiple load data:
[0048] Scenario 1: The television is operating independently;
[0049] Scenario 2: Both the air cooler and the monitor are running simultaneously;
[0050] Scenario 3: Power combination of three devices: air cooler, desk lamp, and monitor.
[0051] The following are all possible load breakdown power sequences for these three scenarios:
[0052] Scenario 1: P 分解1,1 =P Σ
[0053] Scenario 2: P 分解2,1 =P Σ -P1; P 分解2,2 =P Σ -P3.
[0054] Scenario 3: P 分解3,1 =P Σ -P2-P3; P 分解3,2 =P Σ -P1-P3; P 分解3,3 =P Σ -P1-P2.
[0055] Step 4: Perform GMM-based cluster fitting on the power sequence curves of the six load decompositions respectively. The number of mixture components is 8. The initial values are the mean and variance of the GMM cluster fitting of the power curves of the four devices operating individually. The results are as follows: Figure 4-9 As shown. Figure 4 For scenario one, when the device is running alone, the GMM model is designed for... Figure 3 Clustering fitting results of the power curves; Figure 5 For scenario two, where both devices are operating simultaneously, the GMM model is designed for... Figure 3 The power curve decomposes the cluster fitting results after removing the mean power of the air conditioner fan; Figure 6 For scenario two, where both devices are operating simultaneously, the GMM model is designed for... Figure 3 The power curve is decomposed to remove the cluster fitting results of the display power mean; Figure 7 For scenario three, where all three devices are operating simultaneously, the GMM model is designed for... Figure 3 The power curves are decomposed to remove the cluster fitting results of the mean power values of the desk lamp and monitor; Figure 8 For scenario three, where all three devices are operating simultaneously, the GMM model is designed for... Figure 3 The power curve decomposition removes the cluster fitting results of the average power of the air conditioner fan and the monitor. Figure 9 For scenario three, where all three devices are operating simultaneously, the GMM model is designed for... Figure 3 The power curves were decomposed to remove the cluster fitting results of the average power of the air conditioner fan and the table lamp.
[0056] The characteristic vectors of these six possible load decomposition power curves are shown in Table 2, with the mean of each Gaussian component as the characteristic value.
[0057] Table 2. Characteristic Vectors of Each Decomposition Power Curve
[0058]
[0059]
[0060] Step 5: Similarity and Difference Tests. The Spearman correlation coefficient tests were performed on the feature vectors of the power curves of the six possible load decompositions and the feature vectors of the power curves of the four types of electrical equipment operating individually. The results are shown in Table 3.
[0061] Table 3. Results of Spearman's Correlation Coefficient Test
[0062]
[0063] The rank-sum test was performed on the eigenvectors of the power curves of the six possible load decompositions and the eigenvectors of the power curves of the four types of electrical equipment operating alone. The results are shown in Table 4.
[0064] Table 4. Results of the Rank-Sum Test
[0065]
[0066] Step 6: Find Spearman correlation coefficient (r-value) and rank-sum test (p-value) values both greater than 0.8. This confirms that the fourth load decomposition power curve and the air conditioner fan power curve meet the conditions. Therefore, the fourth load decomposition power curve represents the air conditioner fan, and the power combination of the multiple loads is the air conditioner fan, desk lamp, and monitor. This aligns with actual conditions.
[0067] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.
Claims
1. A feature extraction and load decomposition method based on an improved GMM model, characterized in that, Includes the following steps: Step S1: Generate a power combination table; Step S2: Obtain the feature vector of load data samples when the equipment is running alone; Step S3: Load decomposition based on power combination lookup table method; Step S4: Perform Gaussian mixture model-based fitting and clustering on all possible load data sequences obtained after decomposition to obtain feature vectors of all possible decomposed load data. Step S5: Perform linear correlation and difference tests on all possible decomposed load data feature vectors obtained in Step S4 and load data sample feature vectors of each device running individually, and obtain the Spearman correlation coefficient test r value and rank sum test p value, until all possibilities are exhausted; Step S6: Identify the load type based on linear correlation and differential analysis.
2. The feature extraction and load decomposition method based on the improved GMM model according to claim 1, characterized in that, The above step S1 generates a power combination table; the specific process is as follows: collect load data of different electrical devices when they are running alone, perform smoothing and filtering processing, calculate the arithmetic average value of different electrical devices when they are running stably alone, and form a power combination table of different devices.
3. The feature extraction and load decomposition method based on the improved GMM model according to claim 1, characterized in that, In step S2 above, the feature vector of load data samples when the equipment is running alone is obtained. The specific process is as follows: Gaussian mixture model is used to fit and cluster the load data of different electrical equipment running alone after smoothing and filtering. The number of Gaussian mixture components is m. The mean and variance are randomly initialized to obtain the mean of m Gaussian mixture components as the sample feature vector. The mean, variance and number of mixture components of each Gaussian mixture component are saved as the initial values of the Gaussian mixture model in step S3.
4. The feature extraction and load decomposition method based on the improved GMM model according to claim 1, characterized in that, The load decomposition in step S3 above is based on the power combination lookup table method. The specific process is as follows: collect load data of different electrical devices running at the same time, and decompose the load data of different electrical devices running at the same time according to the values in the power combination table in step S1 to obtain all possible load data sequences after decomposition.
5. The feature extraction and load decomposition method based on the improved GMM model according to claim 1, characterized in that, In step S6 above, the load type is identified based on linear correlation and differential analysis; The specific process is as follows: Based on the Spearman correlation coefficient test r value and rank sum test p value obtained in step S5, if both the r value and p value are greater than 0.8, it is considered that the decomposed load data feature vector is similar to the sample feature vector. At this time, the load type put into operation is the electrical equipment corresponding to the sample feature vector.