Research On Feature Reduction And Classification Of Power Quality Disturbances

With the improvement of science and technology, all kinds of complicated, precision andsensitive modern electrical equipments are increasing. People to the electric power qualityrequirements and standards are higher and higher. At the same time all kinds of nonlinear,impact and nonequilibrium electricity characteristics of the power electronic devices are used.That makes the interference of power grid of load increasing, leads to the decline of powerenergy quality, affects the normal work of the industrial enterprise and causes huge economiclosses. In this paper, based on the data mining and machine learning theory, it focus on thefeature reduction and sample reductionof the problem of power quality disturbances.The work of this paper is as follows:1. It is difficult to obtain so many data at real situation. We needs to resort to the simulationdata in the research of the electric power quality disturbance signal recognition. In this thesis,Matlab2009R and power system simulation software PSCAD\EMTDC are used to simulate sixkinds of power quality disturbances. The waveform data simulated requires to be transformed toget important information. This thesis uses multiresolution wavelet transform waveform as datapreprocessing method.2. Proposed a new evaluation function for many kinds of feature selection.After analysised by wavelet transform, wavelet coefficient features with1062d of powerquality disturbances signals are produced. The redundant features should be removed. And theclassification of the power quality disturbancs is a multi-classes classification problem. So amulti-classes evaluation function proposed. The evaluation needs the mean value of Chernoffdistances of every two classes is larger and the variance of t Chernoff distances is smaller. Thebest parameter of β is found by iteration. Most of the features among the13features which areselected by the method are in the frequency band which includes the base frequency.3. A new method of RKM-KNN is proposed.In the recognition of a new sample, the K neighbor algorithm needs to store all of thetraining samples and calculate the distance of the new sample with all the training samples.Thiscan't fit the real time of recognition of power disturbancese. A new method RKM-KNN isproposed to solve the problem. K-means clustering algorithm is proposed to select a trainingsubset of the original training samples and represent them. The training subset is composed ofcluster centers which are generated by K-means during every clustering processes.the result of the t-test shows that only with6.28percent of all the training data, the recognition rate is thesame with using all the training data.