Classification Of Power Quality Disturbances Based On Deep Learning

With the rapid development of social economy,there is an increasing demand for better power quality.In addition,the diversified power sources and power load make the power quality disturbances(PQD)increase frequently.In order to better control all kinds of negative influences caused by PQD,lots of researches have been done and a series of results for PQD signals analysis have been achieved.It is known from literature that the key to elimination of PQD is the correct judgment of the disturbances' types,so proper way of elimination and compensation can be utilized.However,most current methods of PQD classification often start from artificial features selection and extraction,this is not appropriate in practical PQD identification application.Therefore,this article aims to do some researches and attempts to use deep learning for classification of PQD signals from the perspective of traditional classification algorithm of PQD.This paper focuses on two domains: stacked denoising autoencoder(SDAE)and deep belief network(DBN)in the analysis of PQD signals.In the SADE domain,the original simulated PQD signals are corrupted by noise first,and then,the features of corrupted signals are automatically extracted by a deep network of two layers of stacked autoencoder,at last,a BP network is added up to fine tune the whole network and classify PQD signals.This method is an improvement on stacked autoencoder.The denoising SAE improves the anti-noise performance compared to simple SAE and hence increases the classification rate of PQD.For DBN,this paper constructs two layers of restricted Boltzmann machine(RBM)to classify PQD signals by using contrastive divergence algorithm to train the two-RBM network layer by layer.At last,a BP network is attached to fine tune the whole network and classify PQD signals.This method is effective to classify up to 7 types of single and mixed PQDs.Both of the two deep learning based algorithms used in this paper avoid the defect of lower classification rate by traditional methods characterized by artificial features selection,and to some extent,overcome the shortcoming of vulnerability to falling into local minimum in training traditional neural network.Experimental results demonstrate that the two proposed approaches achieve not only a high classification rate but a good anti-noise performance.