Study On Power Fault Recognition Based On Multiwavelet Packet And Artificial Neural Network

With the development of power system, the voltage of power transmission lines is becoming higher and the distance of power transmission lines is becoming longer. They are very important to the safe operation of power system. In the power system of our country there are many single phase reclosing devices in the high voltage transmission system, so an effective phase selector is the base to make reclosing device work correctly. And it is more significant to recognize fault type and select fault phase to ensure relay protection right action when fault occurs on transmission lines. In the thesis a new approach of fault type recognition based on multiwavelet packet theory is first studied and proposed. Whether multiwavelet and multiwavelet packet theory is effective and feasible on power system fault diagnosis and relay protection is the main purpose of the thesis, and trying to solve power fault recognition is also the key point in the thesis.Multiwavelet can simultaneously own symmetry, orthogonality, short support and high order vanish moments, however traditional wavelet cannot possess all these properties at the same time. In the thesis the basic theory of multiwavelet and multiwavelet packet is introduced. The present application situation of power fault recognition and multiwavelet packet is analyzed, and the existent problems are pointed out. A 500 kV transmission line model with PSCAD/EMTDC is built, and different condition short circuit fault signals are simulated based on the model.Multiwavelet packets own better properties than traditional wavelet packets. Compared with traditional wavelet packet decomposition, multiwavelet packet can withdraw more abundant and refined fault features from original fault signals. So in the thesis multiwavelet packet is introduced to the application of fault recognition in power system. A novel power fault recognition method based on multiwavelet packet energy features and artificial neural network is proposed. Firstly, the appropriate multiwavelet packet decomposition of the sampled fault current signal is performed and each frequency band energy of fault currents is calculated. Then eigenvector of multiwavelet packet of the current signal is constructed, and by taking the eigenvector as training samples the back propagation (BP) neural network is trained to implement the fault recognition. The main defect of frequency band energy feature extraction method is just calculating the whole frequency band energy and not thinking the fault signals as time-varying signals. So in order to express comprehensively fault signals the method with extracting frequency band local-energy is proposed and applied to recognize faults.Information entropy is used to describe uncertainty and complexity degree of systems. In recent years entropy has gained some achievements in power system. Through multiwavelet packet decomposition, a sequence of coefficient matrices is obtained. These matrices are thought as a kind of dividing forms of original signals. Then these coefficient matrices are transformed into probability distribution sequences which reflect the sparse degree of the coefficient matrices. According to fundamental principle of information entropy, through combining multiwavelet packet decomposition coefficient probability distribution sequences with information entropy multiwavelet packet coefficient entropy (MPCE) is defined. A novel power fault recognition method based on MPCE and artificial neural network is proposed. Firstly, the appropriate multiwavelet packet decomposition of the sampled fault current signal is performed and each MPCE of fault current is calculated. Then eigenvector of multiwavelet packet of the current signal is constructed, and by taking the eigenvector as training samples the radial basis function (RBF) neural network is trained to implement the fault recognition.Through a lot of simulation results the method based on multiwavelet packet and artificial neural network is proved to be effective and feasible, and the method is better than the one based on traditional wavelet packet and ANN. Another advantage of the method is not sensitive to different operation conditions.