Failure rate prediction under adverse weather conditions in an electric Distribution System using Negative Binomial Regression

Weather conditions have an important influence on failure rates of Power Distribution Systems. An event with strong winds gust and a large number atmospheric discharge can increase the outages' quantity in a hard to predictable form and, consequently, impact in the System Average Interruption Duration Index (SAIDI). A modelling of such weather impact in failure rate is useful for a better understand of it, allowing utilities to take actions to understatement it or, even, mitigate its consequences in real time. Based on it, this paper analyses the response of Negative Binomial Regression in comparison with the Poisson Regression, using outages data (quantity) from a Brazilian distribution utility and meteorological information as wind gust speed and number of thunders, to estimate the number of expected outages and, consequently, the failure rate per kilometer. The focus is to provide a model that emulates the behaviour of system components and parts during a climatic event of any intensity and has more resilient against the outliers normally presented in this type of data set. The results indicate a significant influence of the wind gust speed, and, based on goodness-of-fit index as Akaike Information Criterion and the analysis of the outliers, the Negative Binomial Regression as the best response. The estimated function can be useful to improve the utilities' response, helping them to improve their contingency plan.