Comparison of MLP network training algorithms for fault classification in transmission lines

Fast and accurate fault classification in transmission lines is important for any protection devices. The present of extra transient signals cause to phase misclassification and malfunction of protection devices. Thus, it is crucial to provide a proper protection scheme that offer a good classification performance when the signals are influenced by noise. This paper presents a comparative study of protection scheme using combination of wavelet transform (WT) with multilayered perceptron (MLP) network classifier using various types of training algorithms for fault classification in extra high voltage (EHV) transmission lines. The performance of the suitable training algorithm in MLP network resulted the highest accuracy for fault classification. The wavelet transform is used as a tool to decompose the input three-phase current signals and extracting the significant features. After extracting all these important features, the MLP network is trained by using eight types of different training algorithms. Classification performance of the MLP network is evaluated using two types of datasets; ideal dataset (without noise) and dataset with Signal-to-noise ratio (SNR) of 30. Simulation results show that the MLP network trained using the Conjugate Gradient backpropogation with Powell Beale Restars (traincgb) algorithm indicated the highest accuracy for both case with or without noise.