Hurricane stochastic modeling and efficient regional wind risk assessment

Abstract

As climate change intensifies hurricane activity, coastal communities and infrastructure face growing exposure to hurricane-related hazards. Probabilistic assessment of extreme hurricane winds and associated hazards commonly relies on Monte Carlo simulation of a very large number of synthetic hurricane events, which are then propagated through performance, reliability, and risk analyses of engineered systems. However, the sheer size of these event ensembles can make subsequent analyses computationally prohibitive.To address this challenge, an entropy-regularized Wasserstein distance-based scenario selection (EWDSS) method is proposed to efficiently identify a compact subset of hurricane events from the full synthetic ensemble. The optimally selected subset reproduces the probabilistic characteristics of surface wind speeds at multiple locations with close fidelity to those of the original ensemble, while containing orders of magnitude fewer events. As a result, EWDSS can substantially reduce the computational cost of downstream performance and risk analyses while preserving the statistical fidelity of the hazard representation. Mathematically, optimal hurricane subset selection is formulated as a stochastic dual-optimization problem over integer event inclusion and continuous event weights, with the objective of consistently preserving the probability density functions of a set of wind-speed variables. The problem is solved via entropy-regularized optimal transport, using the Wasserstein distance as the probabilistic metric and enabling fast computation through the Sinkhorn algorithm in a matrix-based implementation. More importantly, the fully differentiable objective function enables recasting the dual-optimization as a single-level, strongly convex optimization problem. This reformulation eliminates the tedious double-loop procedures used in preceding studies and substantially reduces computational cost by allowing the use of standard gradient-based solvers. The proposed EWDSS method is validated by concurrently reproducing annual exceedance probabilities (AEPs) and spatial correlations of 3-s hurricane gust wind speeds across 254 counties in Texas, using a subset of 2,000 events selected from 32,417 synthetic hurricanes. In this example, EWDSS approach reduces computational effort by 95\% relative to existing comparable optimal subset selection methods. In addition, the selected hurricane events satisfactorily replicate the cumulative distribution functions (CDFs) and AEPs of the building portfolio loss for the five representative building types in the same region, paving the way for future applications of EWDSS framework in performance-based wind engineering.