A data-driven machine learning approach to typhoon evacuation decision-making and system reliability evaluation of stochastic evacuation networks
Abstract
With the increasing frequency and intensity of typhoon disasters, developing effective evacuation decision-making systems is critical for ensuring public safety. This paper proposes a data-driven approach that integrates machine learning (ML) classification models with system reliability evaluation to support typhoon evacuation operations. We address three key challenges: (1) predicting evacuation necessity based on meteorological data, (2) modeling uncertainty in evacuation networks, and (3) quantifying evacuation feasibility under resource constraints. First, the Synthetic Minority Oversampling Technique (SMOTE) is applied to handle class imbalance in historical typhoon evacuation data, and a Random Forest classifier is trained to automatically predict evacuation requirements. Second, the evacuation operation is modeled as a stochastic evacuation network (SEN) with multistate travel times reflecting road damage uncertainty. Third, system reliability is adopted as a quantitative performance indicator defined as the probability of successfully evacuating all affected residents within time and resource constraints. The methodology is demonstrated using Typhoon Gaemi, which struck Taiwan in 2024, in Dongshi district, Chiayi County. Evacuation commanders can dynamically adjust decisions regarding announcement timing, vehicle allocation, and route selection based on real-time meteorological data and reliability assessments. The integrated approach enables more comprehensive and data-informed evacuation planning to enhance response effectiveness.