Detection And Recognition Of Power-quality Disturbance Signals

With a diversifying of power sources and electrical loads, the power-quality disturbance signals become increasingly complex. Meanwhile, a large number of electrical equipments of high precision and high sensitivity increasingly require a power supply with good quality. This poses a great challenge to the management of power quality. Correct detection and classification of power-quality disturbance signals are the premise for the power quality management. However, existed methods cannot meet the current requirements well for power quality analysis. Therefore, this paper aims to do some researches and attempts for detection and recognition of power-quality disturbance signals from the perspective of signal processing.This paper focuses on two domains: sparse decomposition and deep learning in the analysis of the power-quality signals.Firstly, in sparse decomposition, two parametric dictionaries based on the structure of disturbance signals of two groups are constructed. Matching pursuit optimized by particle swarm optimization is employed to decompose the signals. Then, reconstruction errors are used to preliminarily classify the disturbances into two groups. At last, the detailed disturbance types can be identified by the values of parameters in the atoms searched out. The main advantage of this method is that there is no need for training set during the process of machine learning, because the dictionary already contains the priori knowledge of disturbance signals. Simulation results show that this method can effectively recognize the classes of disturbance signals in general. Moreover, this sparse decomposition based method can also be applied to detect the disturbance signals with a high accuracy.Secondly, this paper has introduced the deep learning theory into recognition of power-quality disturbance signals. To the best of author’s knowledge, this is the first time in the area of power quality signal analysis. The features of disturbance signals are extracted layer by layer by means of stacked autoencoder. In this way, the hard problem regarding on features selection in existed methods can be effectively resolved. In the procedure of deep learning, the unsupervised sparse autoencoder and the supervised fine-tuning can avoid falling into the local extremum—a shortcoming of traditional artificial neural network. Simulation results prove that the proposed scheme based on deep learning can not only effectively recognize the classes of disturbance signals, but also have a high robustness against noise.