Classification of Power-Quality Disturbances Using Deep Belief Network

This paper proposes to utilize an approach of deep belief network (DBN) for the classification of power-quality disturbances (PQDs). DBN is a deep learning algorithm which has been widely used in computer vision, voice recognition, natural language processing and etc., but barely been used in recognizing PQDs. The structure of the DBN consists of several stacked restricted Boltzmann machines (RBMs) for unsupervised learning. The frame of DBN is organized as follows: firstly, the first RBM is fully trained with the original signal by using contrastive divergence (CD) algorithm to obtain desirable features. Secondly, by fixing the weights and bias of the first RBM, the features turn into the next RBM, which is trained similarly as in the first step. Finally, after enough RBM pre-training, the network is fine-tuned with supervised training by back propagation (BP). The PQDs in this paper includes five single disturbance signal such as interruption, sag, swell, harmonic, oscillatory, and two mixed disturbance signals such as sag-harmonic and swell-harmonic. Experimental results demonstrate that the proposed approach achieves a higher classification rate than traditional algorithms.