Rethinking Design Storms in the Light of Climate Change

It is becoming increasingly important to plan for climate change. While overall rainfall remains constant and even drops in some places, the frequency of extreme rainfall events is increasing. Design storms consist of extreme-value return periods and unitless hyetographs that describe the temporal distribution of rainfall and are important tools for civil engineers to stress-test infrastructure against extreme rainfall events. This thesis develops a data-driven framework for generating the temporal distribution of design storms. Motivated by an exploratory functional data analysis of the unitless hyetographs, we study whether historical storm data can be used to construct more representative temporal distributions for design storms. Each gridpoint is a pooled matrix of quantile functions, where each quantile function is generated from the annual maxima rainfall. Using these representations, we construct spatially contiguous regions through a scalable divisive hierarchical clustering algorithm based on minimum spanning trees and energy distance. Then, within each region, we aggregate all storm representations and apply hierarchical clustering to generate a set of canonical storms. The resulting methodology provides an alternative to traditional design-storm approaches based on expert-drawn regions and legacy temporal distributions. It offers a more flexible, data-driven way to characterize regional storm behavior.