Storm Operation Prediction: Modeling the Occurrence of Storm Operations for Fire Stations

Publicly available fire operation data opens new ways to assist fire stations in planning and resource distribution tasks. Previous work has predicted attributes of certain operations, as well as of thunderstorms, but fire department storm operations themselves have not been predicted yet. We present an approach to predict if storm operations will occur for individual fire stations on specific days, based on time, location, and weather data. As days with storm operations are rare, we artificially balance samples with SMOTE, then compare the prediction performance of different machine learning models with an outlier detection on unbalanced data and uninformed prediction models as baseline. To evaluate our approach, we aggregate datasets for 10 fire stations in Upper Austria, Austria, and predict their storm operations. Our approach thereby achieves a median AUC of 0.91 across fire stations, which is an improvement of 0.44 over baseline models.