Predicting Weather-Related Failure Risk in Distribution Systems Using Bayesian Neural Network

Abstract

The reliability of distribution systems is often challenged under unfavorable weather conditions, where weather-related failures occur with high probability. Predicting the number of weather-related failures in distribution systems can provide guiding information for operation and maintenance decisions, improving the risk management capability of utility companies. This article proposes a novel Bayesian Neural Network (BNN) based model to predict weather-related failures caused by wind, rain and lightning. Superior prediction performance of the BNN based model is verified by contrast experiments with other advanced prediction models under four different evaluation metrics. BNN based prediction model presents remarkable robustness, especially in the prediction of high failure levels. In addition, compared to most previous used prediction models without any prediction confidence feedback, BNN based prediction model has the capability of uncertainty estimation. The confidence interval of prediction results can be obtained, which provides sufficient information for guiding risk management of utility companies. An effective operation and maintenance guiding scheme based on the analysis of prediction uncertainty is proposed, which fully excavates the interpretability of the proposed model and enrich the application value of the model.