Research On Complex Power Quality Disturbances Classification

The term of Power Quality(PQ)defined in IEEE standard 1159 refers to a wide variety of electromagnetic phenomena that characterize the voltage and current in power system.Typical power qualities disturbances(PQD)include swell,sag,interruption,harmonic,transient oscillation,transient impulse,flicker,notching and noise,etc.With the increasing development of industrial technology,the demand of power users for PQ is also increasing.As the PQD may cause incalculable financial losses,an attempt to improve electric PQ in modern power system has heightened the interest in detecting and classifying PQD.This paper has studied the feature extraction,feature optimization,classifier training method and multi-classifier combination techniques involved in classifying complex PQD with the support of two National Natural Science Foundations,aiming at improving the accuracy and recogonized types of PQD classifiers.The topic is divided into the following issues:(1)Feature extraction: This paper makes further improvements to the existing fast S-transform feature extraction technology.By screening different frequency rows of S matrix under different window parameters,a fast feature extraction scheme for complex PQD is proposed,which reduces the amount of calculations while maintaining PQ characteristics.Several numerical construction methods of features are summarrized,transforming feature parameters in the form of matrices and vectors into real values.Meanwhile,a method that deals the features into the form of probabilities is also proposed,which overcomes the drawbacks of the eigenvalues in the supervised learning classifier and provides a new route for classifying PQD.(2)Classifier design: In order to use as much as possible the observational evidence associated with complex PQD,this paper proposes a multi-label complex PQD classifier based on Bayesian Network(BN).The classification model converts factors such as feature values extracted from the signal,local historical record data,and environmental conditions around the monitoring points into probability correlations.After these probabilistic quantitative models are established,the BN classifier can use mathematical methods to calculate the posterior probabilities of the labels through the states of each observational evidence,and the complex PQD is then classified.At the same time,this paper simplifies the BN model by deeply mining the correlation among labels,which reduces the computational complexity of the algorithm.(3)Feature Optimization: In order to further improve the classification accuracy,this paper proposes four types of optimization models based on the characteristics of BN classifiers to find the optimal feature combinations.Each type of optimization model can avoid the highly correlated features to satisfy the conditional independence hypothesis of the BN classifier.This paper also proposes several methods to reduce the calculation time of the fitness function.The reduction of the calculation time of the fitness function enables the optimization algorithm to set more individual numbers and higher upper limit of iterations,so as to improve the optimization performance.(4)Multi-classifier ensemble technology: This paper applied the multi-classifier ensemble technology into the area of classifying complex PQD for the first time and propose an “upper-level decision-making machine” based on BN.By using the "soft evidence" input characteristics of BN,the output of multiple multi-label classifiers are ensembled together and a more accurate result is obtained by retaining the rank characteristics of each sub-classifier.In order to reduce the computational complexity of the “upper-level decision machine”,a forward-searching based classifier selection method is proposed,which reduces the number of sub-classifiers while guaranteeing the classification performance.