Advanced methods for prediction of animal-related outages in overhead distribution systems

Abstract

Occurrence of outages in overhead distribution systems is a significant factor in determining distribution system reliability. Analysis of animal-related outages has practical value since animals cause a large number of outages in overhead distribution systems. This dissertation presents several different methods to investigate the impact of weather and time of the year on the animal-related outage rate. The animal-related outages from year 1998 to year 2007 for different cities in Kansas are provided by Westar Energy. From examinations of the historical data, two factors which influence the animal-related outages, the month type and the number of fair weather days are taken as inputs along with historical outage data for prediction models. Poisson regression model, neural network model, wavelet based neural network model and Bayesian model combined with Monte Carlo simulations are applied to the weekly data of different cites. Even though Poisson regression models, Bayesian models and neural network models are able to recognize the changing pattern of outage rates under different weather conditions, they are limited in their ability to follow the high peaks in the time series of weekly animal-related outages. The introduction of wavelet transform techniques overcomes this problem. Simulation results indicate that the wavelet based neural network models are able to capture the pattern of fast fluctuations in the weekly outages of different cities in Kansas of various sizes. A hyperpermutation method inspired by artificial immune system algorithm is used to solve the overtraining problem in the application of neural networks. Finally, Monte Carlo simulations based on conditional probability tables from Bayesian models are used to find out the confidence intervals of the predictions. We aggregate the weekly data and carry out the analysis on a monthly and yearly basis too. Simulation results indicate that the models are able to capture the pattern as at least 90% of the observed values are within the upper limits of 95% confidence in the predictions for weekly, monthly and yearly animal-related outages of different cities in Kansas. The results obtained from Monte Carlo simulations are compared with the wavelet based neural network model to indentify years with more than expected level of outages.