Power quality (PQ) disturbances classification plays an essential role in ensuring high quality power supply of the power grid. Therefore, most of the researchers have aimed to achieve accurate and fast classification. But yet one of the main issues is how to extract the “right” features from massive amount of voltage and current data. The feature extraction should be performed for the aim of not only increasing the classification accuracy, but in the same time reducing the computational complexity of the classification algorithm. Once the optimal feature combination is found, the next challenge is to determine the most appropriate number of input signals for training the classifier, so as the classifier achieves maximum accuracy for the given features. In this paper we investigate the effectiveness of the wavelet based features in order to find less numerous features combination that attains high classification accuracy. Furthermore, using that combination, we investigate the influence of the different number of input voltage signals for training the corresponding classifier, in order to obtain optimal classification method.