Machine Learning-Based Weather Impact Forecasting

Natural disasters influenced over 4 billion people, required 1.23 million lives, and caused almost US$ 3 trillion economic losses between 2000 and 2019. The picture becomes even more deplorable when hazards, smaller-scale severe weather events not requiring casualties, are considered. For example, 78 percent of power outages in Finland were inflicted by extreme weather in 2017, and train delays, often caused by adverse weather, have been estimated to cost 1 billion pounds during 2006 and 2007 in the UK. To mitigate the effects of the adverse weather and increase the resilience of the societies, the World Meteorological Organisation (WMO) raised the consciousness of impact-based warnings along with impact forecasts. Such warnings and predictions can be used in various domains to prepare, alleviate and recuperate from adverse weather conditions.  This thesis studies how to preprocess data and use machine learning to create valuable impact forecasts for power grid and rail traffic operators. The thesis introduces a novel object-oriented method to predict power outages caused by convective storms. The method combines state-of-the-art storm identification, tracking, and nowcasting algorithms with modern machine learning methods. The proposed object-oriented method is also adapted to predict power outages caused by large-scale extratropical storms days ahead. In addition, the thesis studies the task of predicting weather-inflicted train delays. The method presented in the thesis hinges weather parameters on train delays to anticipate the delays days ahead. The thesis shows that the object-oriented approach is a vindicable method to predict power outages caused by convective storms and that a similar approach is feasible also in the context of extratropical storms. The introduced methods provide power grid operators increasingly accurate outage predictions. The thesis also demonstrates that the train delays related to adverse weather can be predicted with good quality training data. Such predictions offer cardinal information for rail traffic operators in preparing the challenging conditions. Presumably, similar approaches can be applied to any other domain with quantitative impacts produced by identifiable weather events, if sufficient impact data are available. Several advanced machine learning methods were evaluated in the tasks. The results corroborate with existing research: random forests provided a robust performance in all tasks, but also gradient boosting trees, Gaussian processes, and support vector machines proved useful.