A load recognition method based on one-class classification combined with fuzzy width learning

Abstract

The present invention discloses a load recognition method based on single classification combined with fuzzy width learning. The method includes: performing standardization processing on the collected load current data; judging whether there is a switching event by detecting the change of the load current; after detecting the event, extracting the current device current data through the current difference before and after switching, and using Fourier transform to extract the 1st - 9th odd harmonics of the load current to construct a load feature imprint; constructing a load feature library based on the load features, and training a fuzzy width learning system based on the feature library; calculating the feature distance of the feature sequence of the load to be measured through the single classification K-nearest neighbor algorithm to judge whether the sample is an unknown load; classifying the data samples detected as known loads through the fuzzy width learning system. Compared with the prior art, the present invention can improve the detection of unknown loads, improve the robustness of the load classification model, and further improve the recognition rate.

Description

Technical Field

[0001] The present invention relates to the technical field of non-intrusive load monitoring. Background Art

[0002] Electricity is one of the main energy sources for promoting the development of industrial society. With the development of smart power and the enhancement of people's energy conservation awareness, the society's requirements for power quality are increasing day by day. In the construction of smart grid, non-intrusive load monitoring (NILM) has high research value and broad application prospects. Through the excavation of user load information, NILM can effectively alleviate the energy crisis, save energy and reduce consumption, and improve economic benefits. Different from the invasive method, NILM technology installs monitoring equipment at the main power input end to obtain the total power consumption information, thereby identifying the load type and working status of users, improving the safety of measuring equipment, and having the advantages of low cost and convenient maintenance. For power grid companies, the smart meter technology of NILM can predict power demand and provide decision-making basis for decision-makers, strengthen the user's adjustment function for the power grid, and contribute to the construction of smart grid; the use of data analysis and artificial intelligence technology can improve the accuracy of power prediction models, provide reliable basis for power consumption planning, and achieve the purpose of reducing power consumption. For power users, NILM can obtain effective energy-saving measures according to information such as energy consumption, time-of-use electricity price, and power metering. Therefore, NILM is the mainstream trend in the future development of power measurement, and has important significance in the development of power demand side management technology and the construction of smart grid.

[0003] Currently, the more mainstream NILM methods include: load identification based on combinatorial optimization, load identification based on pattern recognition, load identification based on probability model, and load identification based on deep learning. In the above solutions, the load identification algorithms based on combinatorial optimization and probability model have good identification effects on time-varying loads, but as the number of load categories increases, the complexity of model solving increases sharply. Deep learning algorithms have excellent data mining capabilities and can effectively improve the recognition rate of load identification models, but their network structures are complex and a large number of network parameters need to be trained. In the above solutions, most of the current research only considers load identification based on known samples, but the ability to detect unknown loads is the minimum requirement for non-intrusive load devices to operate stably in the actual environment; at the same time, considering the feasibility of non-intrusive load monitoring algorithms, it is necessary to balance the complexity and classification ability of the classification model.

[0004] Aiming at the situation that traditional load identification algorithms cannot cope with the access of unknown loads and the problem of high model complexity, the present invention discloses a load identification method based on one-class classification combined with fuzzy width learning, which can detect unknown loads and is beneficial to improving the robustness of the load classification model. Summary of the Invention

[0005] In view of this, the present invention provides a load identification method based on single classification combined with fuzzy width learning, which has a high recognition rate and recognition speed on the premise of effectively improving the stability of the load identification model in a dynamic environment.

[0006] The technical solution adopted by the present invention is as follows:

[0007] The load identification method based on single classification combined with fuzzy width learning includes the following steps:

[0008] S1: Preprocess the collected load current to reduce noise interference through wavelet denoising;

[0009] S2: Detect the change amount of the load current and judge whether there is a switching event (if there is no switching event, continue to execute S2)

[0010] S3: After detecting the occurrence of a switching event, separate the current event of the system, and extract the odd harmonic amplitudes of the 1st - 9th order of the load current through Fourier transform to construct a feature imprint;

[0011] S4: Construct a load feature library based on the feature imprints of various loads, and train a fuzzy width learning system based on the established feature library;

[0012] S5: Calculate the within-class distance and the distance to be measured of the sample to be measured, and use the single-class K-nearest neighbor algorithm to judge whether the sample is a known-class sample.

[0013] S6: Classify the load detected as a known sample using the fuzzy width learning system.

[0014] Further, the specific step S1 is: Collect the current data of the load operation through an oscillograph, preprocess the collected current data, and use wavelet threshold denoising to filter out the high-frequency environmental noise in the current signal.

[0015] Further, the specific step S2 is: Detect the amplitudes of adjacent sampling points of the current signal. When the current difference is greater than the threshold, it is detected that a switching event has occurred, and the current load event needs to be separated and identified.

[0016] Further, the specific step S3 is: When it is detected that a load switching event occurs, based on the additivity of the steady-state current waveform, use the difference between the current signals of the first 5 cycles before switching and the current signals of the first 5 cycles after switching to separate the current load event. Based on the separated current waveform, extract the odd harmonics of the 1st - 9th order through Fourier transform to construct a load feature imprint.

[0017] Further, the specific step S4 is: Train a fuzzy width learning system based on the load feature library.

[0018] (1) The fuzzy width learning system mainly consists of three parts: a fuzzy system, enhancement nodes, and system output. Fuzzy width learning introduces a set of first-order TS fuzzy subsystems to replace feature nodes on the basis of width learning. Different from traditional neural networks that obtain better approximation effects for nonlinear problems by increasing the number of layers, FBLS makes full use of the nonlinear approximation capabilities of the fuzzy system and the nonlinear activation function to achieve this goal. Therefore, FBLS has a simpler network structure under the condition of achieving the same approximation effect, greatly reducing the model complexity and having strong feasibility.

[0019] (2) Fuzzy width learning calculates the weight parameters through pseudoinverse. The system output consists of two parts: the output of the fuzzy system and the output of the enhancement layer.

[0020] Further, the specific step S5 is as follows: Based on the load feature database, for the sample to be measured, calculate the distance to be measured and the within-class distance of the sample to be measured according to the Euclidean distance, and judge whether the sample is a known-class sample by comparing the distance to be measured and the within-class distance. The calculation formula is as follows:

[0021]

[0022]

[0023] In the formula, d 1 , d 2 are the distance to be measured and the within-class distance A respectively (l) , B (l) are the l-th dimensional features of the sample to be measured A and sample B, p is the p-norm. The present invention calculates the feature distance based on the Euclidean distance, that is, p = 2. C is the data set composed of the K NN sample points closest to sample B in the load feature library. The value of K NN is 30, and x is the load sample data in set C.

[0024] Further, the specific step S6 is as follows: Use the fuzzy width learning system for classification to obtain the load recognition result:

[0025] First, obtain the fuzzy system output and the intermediate vector through the TS fuzzy system, then use the intermediate vector as the input of the enhancement layer of the fuzzy width learning system, obtain the enhancement layer output through nonlinear transformation, and finally calculate the weight parameters of the fuzzy width learning system through pseudoinverse according to the known output matrix.

[0026] The beneficial effects of the present invention are as follows: The present invention detects unknown loads through a single-classification algorithm, improving the robustness of the load classification model; and classifies loads through the fuzzy width learning system, improving the recognition rate of the model. Description of the Drawings

[0027] Figure 1 : Flow chart of load identification based on one-class classification combined with fuzzy width learning

[0028] Figure 2 : Schematic diagram of load current and harmonic characteristics in the embodiment

[0029] Figure 3 : Schematic diagram of the structure of the fuzzy width learning system Specific implementation manner

[0030] To make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] The present invention provides a load identification method based on one-class classification combined with fuzzy width learning, and its implementation steps include:

[0032] S1: Collect the current data of the load operation through an oscillograph, preprocess the collected current data, and use wavelet threshold denoising to filter out the high-frequency environmental noise in the current signal. The implementation method of wavelet threshold denoising is as follows:

[0033] (1) Decompose the current signal into different frequency bands at each scale by wavelet decomposition

[0034] (2) Remove the wavelet coefficients belonging to noise in each scale, and retain and enhance the wavelet coefficients belonging to the signal.

[0035] (3) Reconstruct the processed wavelet coefficients by wavelet inverse transform to obtain the denoised load current signal.

[0036] S2: Detect the change amount of the load current, and judge whether there is a switching event (if there is no switching event, continue to execute S2). Set the threshold as H. When the difference between adjacent sampling points of the current signal is greater than the threshold, it is detected that an event has occurred. The determination of the threshold is 1 / 5 of the minimum power load current.

[0037] S3: After detecting that a switching event has occurred, separate the current event of the system, and extract the 1st - 9th odd harmonic amplitudes of the load current through Fourier transform to construct a feature imprint. The odd harmonic extraction method is shown as follows:

[0038]

[0039] In the formula, the DC part is A 0 , the even part of the signal is A n cos(nωk), the odd part of the signal is B n cos(nωk)

[0040] S4: Based on the load feature library, train the fuzzy width learning system, and its structure is as Figure 3As shown, the fuzzy width learning system calculates the network weights through pseudoinverse, as shown in the following formula:

[0041] W=(GΩ|E m ) + O

[0042] In the formula, W is the network weight matrix, G is the model parameter matrix subject to uniform distribution in [0,1], and Ω is the activation weighted weight matrix of the fuzzy system E m is the output matrix of the enhancement layer O is the known output category matrix, and the model parameters are quickly calculated through pseudoinverse, which is beneficial to model update.

[0043] S5: Extract harmonic features for the sample to be measured, as Figure 3 shown, and then detect based on the load feature library to determine whether it is a load category known in the feature library. The detection method is:

[0044] (1) Calculate the feature distance between the sample B closest to the sample A in the feature library, denoted as the distance to be measured; then calculate the feature distances of the K NN sample points closest to the sample B in the feature library, and take the average value as the within-class distance. The distance calculation is shown in the following formula:

[0045]

[0046] (2) When the distance to be measured is less than the within-class distance, it is a known sample, otherwise it is an unknown sample.

[0047] S6: For the load samples detected as known categories, use the trained fuzzy width learning system for classification, as shown in the following formula:

[0048] Y=XW

[0049] In the formula, Y is the system output, X is the system input, that is, the load classification result, and W is the network parameter obtained through pseudoinverse.

Claims

1. A load identification method based on single-classification combined with fuzzy width learning, characterized in that, it includes the following steps: S1: Collect the current data at the total power input end, and preprocess the obtained load current data; S2: According to the processed current data, judge whether a load switching event occurs based on the change amount of the monitored current. If no switching event occurs, continue to execute S2; S3: Through the current difference before and after switching, extract the current data of the current device, then use Fourier transform to extract the 1st to 9th odd harmonics to construct a load feature imprint, and standardize the data; S4: Construct a load feature library based on the load features, and train a fuzzy width learning system based on the feature library. The fuzzy width learning system mainly consists of three parts: a fuzzy system, enhanced nodes, and system output; S5: After extracting the features of the sample to be measured, calculate the distance to be measured and the intra-class distance of the sample based on the load feature library through the single-class K-nearest neighbor algorithm, and judge whether the sample is an unknown load according to the size of the distance; S6: Classify the load detected as a known sample using the fuzzy width learning system to identify the load type.

2. A load identification method based on single-classification combined with fuzzy width learning as described in claim 1, characterized in that, in the step S1, the current data at the total power input end is sampled at a high frequency, and the current data is preprocessed, and wavelet threshold denoising is used to reduce noise interference: (1) Decompose the signal into multiple scales through wavelet transform, and denoise the wavelet coefficients of each layer of the signal. The method is threshold quantization; (2) Reconstruct the signal using the quantized wavelet coefficients through inverse wavelet transform, so as to achieve the purpose of denoising.

3. A load identification method based on single-classification combined with fuzzy width learning as described in claim 1, characterized in that, in the step S2, the method for judging the load switching event according to the current change amount is: (1) Calculate the current difference of the load current in adjacent cycles at a certain moment; (2) If the difference is greater than the set threshold, it is detected that a switching event has occurred.

4. A load identification method based on single-classification combined with fuzzy width learning as described in claim 1, characterized in that, after detecting a switching event in the step S3, through the additivity of the steady-state current, extract the current data of the current switching event, extract the 1st to 9th odd harmonics through Fourier transform to construct a load feature imprint, and standardize the data. The standardization formula is as follows: where x is the odd harmonic information of the extracted current for 1 - 9 times, is the harmonic average value, σ is the harmonic standard deviation, and x * is the harmonic feature after standardization.

5. A load identification method based on single-classification combined with fuzzy width learning as described in claim 1, characterized in that, after extracting the features of the sample to be measured in the step S5, calculate the distance to be measured and the intra-class distance of the sample based on the load feature library through the single-class K-nearest neighbor algorithm, and judge whether the sample is an unknown load according to the size of the distance. When the distance to be measured is greater than the intra-class distance, it is judged as an unknown load, otherwise, it is judged as a known sample; the calculation methods of the distance to be measured and the intra-class distance are as follows: where d 1 and d 2 are the distance to be measured and the within-class distance respectively, A (l) , B (l) are the l-th dimensional features of the sample A and sample B to be measured, p is the norm, and the feature distance is calculated based on the Euclidean distance, i.e., p = 2; C is the data set composed of the K NN sample points closest to sample B in the load feature library, the value of K NN is 30, and x is the load sample data in set C.

Images