Wavelet-based techniques for classification of power quality disturbances

The quality of power supply has become an important issue for electricity utilities and their customers. In recent years there has been a rising incidence of damage attributed to the power quality supplied to the customers of electric utilities. Meanwhile, there has been a rapid increase in the already widespread use of electronic equipment and modem power electronic devices. These trends have both decreased the quality of power on the electric grid and increased the equipment's sensitivity to power quality disturbances. In order to improve the quality of the power supply, identifying the type and source of troublesome disturbances is an essential task. Existing automatic disturbance classification methods have replaced the traditional visual inspection of the disturbance waveforms. However, they are not reliable because those methods rely on the classification capability of large neural networks operating on inputs derived by simply pre-processing the disturbance signals with discrete wavelet transforms [134,135,136,137,138]. Long and redundant feature vectors both take a long time to train the network and result in a reduced classification rate. In this thesis, we aim to develop an efficiency method that automatically classifies power quality disturbances by using wavelet transform techniques to generate short and nonredundant feature vector. Because of the wide range of power quality disturbances and their characteristic waveforms, ranging from very simple stationary and deterministic harmonics to highly transient and stochastic waveforms, different and appropriate analysis techmques are needed to achieve the overall classification objective. It is well known that the traditional Fourier analysis is ideal for analysing steady state signal. Although it is very powerful, Fourier analysis does not have the temporal resolution needed to cope with sharp changes and discontinuities in signals. Recent years have witnessed a proliferation in the applications of wavelet transforms to signal analysis in a wide variety of fields, from geo-physics to telecommunications to bio-medical engineering. This has occurred because wavelet analysis provides dual localisations in both the time and the frequency domains. Moreover, wavelet analysis allows the flexibility of choosing a wavelet that suits a particular application. Especially by using the simple and flexible lifting scheme, we can construct a time-variant or space-variant wavelet - known as second-generation wavelet. The second-generation wavelet analysis makes optimal use of the correlation between neighbouring signal samples and between neighbouring frequency components to construct 'local' wavelets, which adapt to the local characteristics of the signal. Common types of wavelet schemes are the orthonormal or biorthonormal wavelet transforms that are typically used in compression and coding applications. This is due to the fact that those schemes can be implemented with fast algorithms and they are non-redundant representations of a signal. Unfortunately, they suffer the limitation of not being translation invariant; a totally different set of transformed coefficients is obtained when the same signal is shifted. This is the major concern in pattern recognition applications. There exist a number of wavelet schemes that have the shift invariance property in their multiresolution representations. In this thesis, the local maxima and the matching pursuit techniques are presented as the two most appropriate techniques for power quality solutions. This is because the two techniques can efficiently decompose a signal and have the ability to precisely measure power quality disturbance characteristics so that they represent the disturbances by a compact, time-invariance feature vector. The final task of classification is the selection of an appropriate classifier for use with the feature vector. There are two main approaches of pattern recognition: one is parametric and the other is non-parametric [129]. Parametric approaches can be either deterministic or statistical. The statistical parametric approach requires a good assumption about the statistical distribution of the data. On the other hand, the nonparametric approach, known as the neural network approach, does not require any statistical assumption about the data. In our statistical approach, we use a two-layer network structure with locally tuned nodes in the hidden layer, known as Radial Basis Function (RBF) network [106,120,121]. The network has only a local learning capability and a limited learning inference from the training data, but trains quickly as the training of the two layers is decoupled. In an RBF network, the crucial concern is the selection of cluster centres and their widths. However, current techniques give suboptimum positions of cluster centres and their widths, thus limiting the classification rate. To improve the p~rformance of an RBF network, we propose to modify the structure of the RBF network by \' introducing the weight matrix to the input layer (in contrast to the direct connection of the input to the hidden layer of a conventional RBF) so that the training space in the RBF network is adaptively separated by the resultant decision boundaries and class regions. During training iterations cluster centres their widths and the input layer weights are optimally determined together and concurrently adjusted to maximise the discriminant between classes thus minimising the classification error. In this way the network has the ability to deal with complicated problems which have a high degree of interference in the training data and achieves a higher classification rate over the current classifiers using RBF. For the classification of different types of disturbances that may be present on a power supply in this thesis we show that our automatic classification techniques achieve superior recognition rates over current techniques. This improvement is done in two steps. The first improvement is the extraction of disturbance features using appropriate signal processing tools from which we obtain an efficiency and translation invariant feature vector. The second improvement is the designing of an appropriate classifier which maximises the inter-class discriminant function."