Reinforcement learning for optimizing hurricane evacuation decisions: Hurricane Irma case study

Tropical cyclone risks are expected to intensify with climate change, significantly complicating evacuation management. During Hurricane Irma in 2017, Florida faced its largest evacuation, impacting 6.5 million people and involving 4 million vehicles. Current hurricane simulations are slow and lack probabilistic modeling, making large-scale traffic evacuation management challenging due to incomplete risk profiles. This study presents a novel approach to optimize evacuation orders under uncertain weather conditions using the Pangu weather forecasting model and reinforcement learning. By perturbing Hurricane Irma’s forecast with the Pangu model, we create realistic decision-making scenarios. Reinforcement learning algorithms then optimize evacuation orders, considering factors such as travel time, sheltering risks, and overall safety. Our approach shows an 8% improvement in traffic system performance compared to traditional fixed evacuation orders. This work highlights the potential benefits of enhanced weather forecasting accuracy in improving evacuation strategies, offering a more adaptive and effective response to future tropical cyclones.