Consumer load prediction and theft classification using intelligent techniques

Abstract

Acquisition and analysis of consumer electricity data are essential for the management of the power system. Several approaches for analyzing these data have been reported in the literature with various degrees of success. However, consumers` aggregated data were applied, which lacks information about individual consumer on the network. It is well known that consumer data could reveal the activities of the consumer as electricity is being used. Hence, there is a need to analyze such data, especially for proper monitoring and decision making, such as routine maintenance, dispatch of generators and detection of electricity theft. The latter has caused a low revenue generation for the power utilities. This thesis investigates the use of advanced signal processing techniques which are based on prediction, correlation analysis and intelligent system for the analysis of power consumption data at low voltage distribution system. Due to unavailability of database for individual consumer on low voltage distribution system, a prototype of consumer load for database has been developed in this research work. The acquired data are then analyzed using parametric modelling techniques which are thereafter referred to as statistical modelling techniques such as autoregressive (AR) and autoregressive moving average (ARMA) techniques. The determination of their model order is very critical to the performance of these techniques, hence, various AR model order selection criteria have been examined in this thesis so as to select the best approach for consumer data analysis. Once the best AR model order is obtained, then an iterative procedure is used for determination of the MA model order for the case of ARMA model. Both AR and ARMA model coefficients are then estimated and later used for prediction. A modified method of intelligent prediction using nonlinear autoregressive with exogenous input network (NARX) which depends on differential evolution (DE) and genetic algorithm (GA) has been developed for the analysis of consumer data in this thesis to alleviate the problem associated with the conventional parametric modelling techniques. In addition, the detection and identification of electricity theft are obtained by using the correlation analysis, parametric models and support vector machines (SVM). The above proposed approaches produce reasonable good results. First the AR prediction requires higher model order in order to produce good results, thus, a model order of 20 is required for daily data, whereas in this thesis, an ARMA (6, 4) was found to be suitable for the analysis of consumer daily load prediction. As expected, the new algorithm developed which comprise the NARX, GA and DE and named (NARX-GA-DE) outperformed the conventional NARX network. The NARX prediction error is with a network structure of two input and output tapped delays and nine hidden neurons while the error of NARX-GA-DE is with a structure of one input and output tapped delay and one hidden neuron. Correlation analysis results for theft detection are found to produce good results when normal data were correlated with theft data. The SVM classification results for theft detection indicates that the training and testing performances range between 97.9% and 61.8% respectively. The methods used in this thesis are novel and the results obtained are unique since no other published work has been seen.