Research On Household Load Identification Combining Improved Nearest Neighbour Method And Support Vector Machine

The Future smart grid will have more interaction with users in information andbehavior. So the smart meters need to have more accurate and reliable electric energymanagement capabilities. That means the electric energy can not only measureed bytime, but also measureed by classification. Therefore, the meters are required to have acapacity of load classification firstly. Current studies of Residential electricity loadclassification generally use one-layer classifier to identify the entire load, resulting in acomplex algorithm with large calculating quantity. Also, most studies did not resolvethe problem of mixed multi-load classification with the strong noise and the fluctuatingof high-power and unstable load.For those problems, this paper proposes a non-intrusive identification methodbased on improved nearest neighour algorithm and support vector machines by studingthe pros and cons of current residential load identification methods. Firstly, analyzingthe signal featrue of the common residential electricity load and establish a loadswitching event detection algorithm to extract featrue and establish the featrue databaseof residential electricity load. Secondly, improved nearest neighbor algorithm andcombines it with support vector machine classifier. Using the training data for classifiertraining, on the basis of which, using testing data to proves the validity of the method.Concrete content are as follows:①The voltage and current and power wave of different kinds of residentialelectricity load with different brands and different power are collected and analysed.And the feasibility and feature extraction of load under mixed operation are discussed.Compares the similaritiy and difference of different loads in switching as well as thedifference of waveform changes between the real switching event and false event. Then,a load event detection algorithm is presented and the experimental results show that thealgorithm can eliminate the interference caused by power fluctuations and achieve agood effect of load switching detection. After the load switching detection, the stablestate and transient state area of the load waveform can be divided and the marketingpoints are be located. By using different mathematical transform methods, differentfeature are extracted from current and power data. Then, the original and the normalizedfeature are stored and the20-dimensional feature database of five kinds of loads with100samples is established. In the end, classification capacity of different feature is evaluated according to the ratio between the distance of same class and different class infeature space. After that, an initial screening for the subsequent optimization of featurecombination is accomplished to reduce the range of options and the amount ofcomputation.②an improved method combining the probability of finite nearest neighborelement and the mean center similarity is proposed by analysising the nearest neighbormethod and C-means method. Considering the fewer computation of multi-identify andexpansibility, a multi-classifier based on improved nearest neighbor method is designed.Then, the optimal parameters of classifier and the combination of feature are trained.The actual experiments show that, the three-layers classifier method based on improvenearest neighbor can effectively identify multi-class residential electricity loads withhigh noise interference. Including two recognition conditions of identifing small powerload with unstable and huge-power loads and low sampling rate. Meanwhile, the rate ofrecognition of the proposed method is significantly higher than the traditional nearestneighbor method with single and three-layers and the traditional C-means clusteringmethod. The validity and robustness of the proposed method is proved.③Aiming at the difficulty recognition of different loads with similar waveform,the tatal recognition rate of every classifiers in the proposed method are counted, Andfind the linear unseparable problem of the improve nearest neighbor method whensloving the recognition of similar loads. Therefore, the support vector machine (SVM)classifier, which is fit for solving linear unseparable problem, is introduced to take theplace of one improved nearest neighbor classifier. And the parameters of SVM classifieris optimized. The experimental results show that the recognition rate has beensignificantly improved, and the SVM classifier is proved helpful to solve the problem oflinear unseparable. So, the combination of the two classifiers can improve the overallability of classifiers.