Predicting weather-related power outages in large scale distribution grids with deep learning ensembles

Abstract

Weather events are primarily contributors to electrical supply disruptions, prompting the need to accurately forecast these weather-related power outages. This paper focuses on predicting daily reported incidences in electrical grids within specific regions, leveraging weather conditions as specific predictive variables. An optimization-based feature selection approach has been considered for selecting the optimal Reanalysis node locations used as predictors. To overcome the data imbalance challenge and enhance prediction accuracy, we propose a Deep Learning-based (DL) ensemble algorithm. Five distinct DL architectures are considered, generating multiple individual learners with randomly selected hyperparameters. Diversity is ensured by training each model with slightly different randomly sampled data. Three information fusion techniques construct the final ensemble models. The proposed approach has been successfully evaluated in predicting real daily reported incidences in distribution lines across two Spanish provinces, Valencia and Albacete, achieving an accurate prediction of days with extreme incidences while maintaining a good overall performance and a low rate of false alarms. The top-performing models using this methodology achieve detection rates of 38% and 73%, with false alarm rates of 1% and 3%, respectively. This approach not only enhances prediction accuracy compared to individual learners but also improves the generalization ability and robustness of standalone DL models. Additionally, it effectively reduces the inherent overfitting of these methods, removing the necessity for a complex hyperparameter selection process.

Keywords

• Power outage prediction
• Weather-related outages
• Electrical distribution lines
• Deep learning ensembles
• Ensemble learning