Design rainfall framework using multivariate parametric-nonparametric approach

Design rainfall time series are a critical input for urban flood modeling and require depth-duration-frequency curves along with a design temporal pattern. The association between rainfall depth and duration, which depicts the inherent structure of rainfall patterns, plays an important role in flood-causing potential; this association can be revealed through a bivariate rainfall frequency analysis to obtain the depth-duration-frequency curves. Unlike earlier approaches that use either parametric or nonparametric models, the present study proposes a new semiparametric model approach, which can evaluate all possible combinations of the parametric-nonparametric marginals without restrictions. A comparison between the copula-based bivariate frequency analysis and the generalized least-squares regression shows the former to be better in performance. The first three best-fit models-Gaussian kernel, triangle kernel, and Burr type XII combined with the generalized Pareto distribution-yield consistently similar results, indicating the robustness of the approach. Realistic design rainfall temporal pattern could also be derived from the observed set of temporal patterns using their skew, kurtosis, and bimodality measure to quantify the maximum flood-causing potential and also to generate the design rainfall time series. The proposed framework has been demonstrated for a severely flood-prone coastal megacity, Mumbai, in India.