Residual error modelling for process-based flood frequency analysis

Flood frequency analysis (FFA) is essential for flood risk management and infrastructure design, yet estimates of flood quantiles remain strongly influenced by the length and representativeness of available streamflow record. We introduce UQ-flood, a process-based flood-frequency framework that integrates a stochastic weather generator, a process-based hydrologic model, and a residual error model developed at the annual-maximum scale. The framework explicitly addresses the presence of errors in simulated annual flood peaks generated by models calibrated to continuous streamflow data. We apply UQ-flood to three Canadian watersheds with contrasting hydroclimatic conditions using 30 years of meteorological and streamflow observations. Long-record empirical annual-maximum estimates provide the primary benchmark, while Bayesian Generalized Extreme Value estimates based on short and long records provide secondary statistical benchmarks for the central flood-quantile estimates. Across the three watersheds, the central UQ-flood estimates remained within the confidence intervals for the long-record discharge quantiles, although their relative proximity to the empirical estimates varied by watershed and return period. Process-based simulations without residual-error correction systematically underestimated flood magnitudes, whereas applying the residual error model shifted the estimated flood-frequency curves toward the long-record observational benchmarks. The framework also propagates residual predictive variability directly at the annual-maximum scale. These results demonstrate that annual-maximum residual error modelling can correct for systematic discrepancies in process-based flood estimates and support central flood-quantile estimation even at sites with short data records.