A Bayesian network framework for probabilistic assessment of regional building losses from hurricanes

This paper proposes a Bayesian network (BN) framework for the probabilistic assessment of regional building losses induced by hurricanes. To explicitly capture the spatial heterogeneity of damage, the study area is partitioned into geographic subdivisions, with the model formulated as a BN-based multi-output regression framework. This framework utilises key hurricane parameters, including translation speed, heading angle, location, and central pressure, as inputs to generate probabilistic loss statistics for each subdivision as outputs. This study introduces new schemes for the discretisation of the input and output variables of the BN. For hurricane variables, a supervised discretisation method is developed that integrates weighted principal component analysis with decision tree algorithms, while also considering the relative importance of each subdivision (e.g., building density) to improve accuracy in high-priority areas. For loss variables, a clustering-based discretisation method is applied to capture the characteristics of regional building losses. The proposed framework enables efficient assessment of the spatial distribution of hurricane-induced building losses in a community, accounting for uncertainties and spatial correlation in hurricane hazard predictions and structural performance. Its effectiveness is demonstrated through a numerical example.