Tranformada wavelet e redes neurais artificiais na análise de sinais relacionados à qualidade da energia elétrica

Abstract

This work presents a different method for power quality signal classification using the principal components analysis (PCA) associated to the wavelet transform (WT). The standard deviation of the detail coefficients and the average of the approximation coefficients from WT are combined to extract discriminated characteristics from the disturbances. The PCA was used to condense the information of those characteristics, than a smaller group of characteristics uncorrelated were generated. These were processed by a probabilistic neural network (PNN) to accomplish the classifications. In the application of the algorithm, in the first case, seven classes of signals which represent different types of disturbances were classified, they are as follows: voltage sag and interruption, flicker, oscillatory transients, harmonic distortions, notching and normal sine waveform. In the second case were increased four more situations that usually happen in distributed generation systems connected to distribution grids through converters, they are as follows: connection of the distributed generation, connection of local load, normal operation and islanding occurrence. In this case, the voltage on the point of common coupling between GD and grid were measured by simulations and were analyzed by the proposed algorithm. In both cases, the signals were decomposed in nine resolution levels by the wavelet transformed, being represented by detail and approximation coefficients. The application of the WT generated a lot of variations in the coefficients. Therefore, the application of the standard deviation in different resolution levels can quantify the magnitude of the variations. In order to take into account those features originated from low frequency components contained in the signals, was proposed to calculate the average of the approximation coefficients. The standard deviations of the detail coefficients and the average of the approximation coefficients composed the feature vector containing 10 variables for each signal. Before accomplishing the classification these vectors were processed by the principal component analysis algorithm in order to reduce the dimension of the feature vectors that contained correlated variables. Consequently, the processing time of the neural network were reduced to. The principal components, which are uncorrelated, were ordered so that the first few components account for the most variation that all the original variables acted previously. The first three components were chosen. Like this, a new group of variables was generated through the principal components. Thus, the number of variables on the feature vector was reduced to 3 variables. These 3 variables were inserted in a neural network for the classification of the disturbances. The output of the neural network indicates the type of disturbance.