The role of hydrologic model choice in probable maximum flood (PMF) estimation uncertainty

The probable maximum flood (PMF) is a key factor in dam safety and other hydrologic risk assessment contexts. It is often calculated using the probable maximum precipitation (PMP) in conjunction with streamflow prediction models. However, many such models are available, and structural differences between models and their differing data requirements could lead to significant differences in the PMF estimates they generate. We tested this hypothesis by estimating the PMF for the Alouette River in southwest British Columbia, Canada, using three models selected in part for their diversity: WRF-Hydro (complex spatially distributed process-based model), UBCWM in Raven (intermediate-complexity semi-distributed process-based model) and a machine learning model (lumped empirical model). WRF-Hydro provides the most physically complete representation and is also the most resource-intensive and data-demanding. UBCWM offers an effective balance between realism and efficiency, while the ML approach, though less physically interpretable, is computationally inexpensive. The models were developed while holding the meteorological inputs constant, although all three models handle the meteorological input very differently and have different input requirements. In addition to temperature and PMP sequences, WRF-Hydro required wind speed, radiation and specific humidity inputs which provided an additional challenge of how these inputs might change under a probable maximum storm. All three model approaches produced plausible PMF magnitudes, yet these PMF estimates varied by more than 400 m³/s or 28%, in terms of maximum peak flow and by about 8,000 m³ in terms of event volume. We also utilised IES-PEST++ to quantify parameter estimation uncertainty for the process-based models, which fell in the range of 12% to 16%. Since the PMF estimates produced by all three models appear reasonable for practical adoption—and the true value cannot be empirically verified — these results underscore the importance of explicitly accounting for uncertainty in PMF estimation. Additionally, machine learning models are not commonly applied in PMF studies due to well-known concerns around extrapolation to unseen data, we show that even a relatively simple ML model can extrapolate usefully to values more than 40% greater than the data it was trained on. The outcomes of this study provide a quantitative measure of the implications of hydrologic model and parameter selection uncertainty on PMF estimation and imply that a multi-model ensemble could be an effective pathway toward capturing that uncertainty in flood risk assessment processes.