Comparison of supervised learning techniques for non-technical loss detection in power utility

Abstract

Non technical losses (NTLs) originating from electricity theft and other customer malfeasances are a problem in the electricity supply industry. In recent times, electricity consumer dishonesty has become a universal problem faced by all power utilities. Previous work carried out for NTL detection resulted in a Support Vector Machine (SVM) based detection framework. The present study performs a comparative study for NTL detection using supervised machine learning techniques such as the: Back-Propagation Neural Network (BPNN) and Online-sequential Extreme Learning Machine (OS-ELM). Model testing is performed using historical customer consumption data for three towns within peninsular Malaysia. The detection hit-rate of all compared models is obtained from TNB Distribution (TNBD) Sdn. Bhd. for onsite customer inspection. Experimental results obtained indicate that the BPNN detection model achieves the lowest average detection hit-rate of 36.07%, while the OS-ELM model obtains a slightly higher average detection hit-rate of 51.38%. The previously proposed SVM-based NTL detection model outperforms the BPNN and OS-ELM by far with the highest average detection hit-rate of 60.75%. This indicates that the use of a SVM-based soft-margin approach results in a better generalization performance for the application of NTL detection as compared to the BPNN and OS-ELM schemes. © 2012 Praise Worthy Prize S.r.l. -All rights reserved. - All rights reserved.