GNN-Based Deep Reinforcement Learning for Coastal Emergency Response under Extreme Water Levels

When a storm surge approaches a coastal town, emergency response depends on how quickly and effectively an operational plan can be implemented using the available information. This may include real-time water-level forecasts, expected flood extent, road access, the location of vulnerable residents, shelters and available resources. Bringing this information together under time pressure can be slow and depends on the duty officer’s expertise, potentially delaying critical decisions when they matter most. In this research, we ask whether a policy can learn the timing of emergency actions from decades of real events using a spatial representation of the territory. Lido di Venezia is represented as a heterogeneous graph built from public elevation, road, building and census data and driven by hourly tide-gauge records from 1983 to 2024. Seven possible emergency actions are derived from the municipal emergency plan. We train a PPO decision policy using a graph representation that links residential areas, roads, shelters and hazard zones. The trained policy is then tested on storm windows not used during training and compared with a rule-based system simulating the officer’s discretionary decisions. On 18 historical test scenarios spanning low to severe conditions, PPO achieves a similar reduction in population exposure to the rule-based system. In a second test, with changing and uncertain forecasts and limited protection, PPO again performs similarly but does not outperform the tuned rule. In a third test, removing graph message passing does not substantially reduce performance. Together, these tests show that PPO can learn an effective protective response from an outcome-based reward alone: it is given the plan’s actions and the conditions under which each is allowed, but not the thresholds that tell the comparator when to act. Future research will test the learned policy on more complex edge cases and replace the simulated officer with direct assessment by emergency-management stakeholders, supported by large language model (LLM) based interpretation of the recommendations.