| This thesis is an investigation into the characterization and classification of power system transients, using advanced signal processing and pattern classification techniques. In the system developed in this thesis, which is intended to act as an artificial consultant to power systems operators, wavelet and multifractal analyses have been used to characterize transients in power systems and to extract features from them. The Daubechies wavelet family used in this thesis decomposes the signal into details and approximations, which contain the high and low frequency content of the signal, respectively. The variance fractal dimension trajectory method characterizes the complexity of the signal using a sequence of fractal dimensions. The thesis considers the usefulness of each of these methods in characterizing the transients as well as their combination. Various classification methods, namely the Bayes rule (based on the method of maximum likelihood, ML), the nearest neighbor (k-NN), and the probabilistic neural networks (PNN) have been used to identify the corresponding class of a transient.; The performance of the system is evaluated both on simulated transients and recorded data obtained from Manitoba Hydro. For the simulated data, the ML-based Bayes rule produced an average accuracy of 82.92% with the VFDT-based features, an average accuracy of 96.25% with the wavelet features and 100% with the combined features. The PNN yielded an average accuracy of 95% with the training set and 88.75% with the testing set of data. Classification of the recorded data produced an average of 82% using the k-NN classifier. The results show superior performance to previous work, both in the accuracy of the classification and significant reduction of the number of features used. |
