Weather‐related power outage model with a growing domain: structure, performance, and generalisability
Peter L. Watson, Diego Cerrai, Marika Koukoula, David W. Wanik, Emmanouil N. Anagnostou
Abstract
The outage prediction model (OPM) is a weather‐related machine learning‐based power outage model, which has been developed at the University of Connecticut for many years and has recently grown to cover three states and five utility service territories. This is a large heterogeneous domain supported by a large dataset of hundreds of storm events. This dataset presents the opportunity to investigate the effect of the spatial organisation and training structure on model performance, identify potential weaknesses in the modelling approach, and evaluate the generalisability of the modelling methodology. By organising and sub‐dividing the modelling system informed by clustering analysis, this study experiments with the structure of the model to identify potential sources of error in the modelling system and evaluate generalisability. The clustering analysis identifies regions of specific climatic, topographic, and environmental characteristics, and performance improvements are observed when the model is subdivided and trained separately for each cluster in certain clustering. Models trained on all available data from all five utility service territories consistently have good performance showing that the OPM modelling approach is generalisable across different service territories and power utilities.