Applying Stochastic Storm Transposition for Large-Scale Flood Inundation Modeling in a Dry Mountainous Basin

Over the past 10 years, HEC-RAS, the flagship hydraulic model developed by the US Army Corps of Engineers (USACE) has benefited from several milestone advancements including the adoption of a 2D overland flow solver, infiltration procedures, and expanded forcing options including gridded precipitation. The Federal Emergency Management Agency (FEMA), tasked with developing regulatory flood maps, relies heavily on HEC-RAS models for hydraulic calculations. While hydrologic methods for FEMA have varied in the past, many studies have relied on USGS regression equations for the development of frequency-based flows. In cases where hydrologic models were developed, simplifying assumptions have often been incorporated to account for traditional methods of lumped modeling, relying on average precipitation developed using NOAA’s Atlas 14 for forcing, and the use of areal reduction factors for larger watersheds. The advent of 2D modeling in HEC-RAS forced by precipitation has led to a large number of studies developed using uniformly distributed precipitation over large areas with no reduction factor applied, changing rather than addressing the need for more complex modeling approaches.

Over these same past 10 years, FEMA has begun the process of evaluating methodologies to incorporate uncertainty into flood hazards. This talk will present a case study for using one of these methods, Stochastic Storm Transposition (SST), to incorporate uncertainty in hydrologic modeling applied in the Upper Green watershed in Southwest Wyoming. This high desert region was a prime location to evaluate the suitability of SST, as there is no published Atlas 14 data available. This work presents a cloud-based stochastic-framework for a complex collection of HEC-HMS hydrologic models developed to provide temporal frequency-based inputs to 2D HEC-RAS models for use supporting a FEMA study. A catalog of historic storms selected from a region that is climatologically homogeneous to the study area was developed using gridded quantitative precipitation estimates (QPE) from the Analysis of Record for Calibration (AORC) dataset, developed by the National Weather Service. Following existing published approaches, a Poisson distribution was sampled to determine the number of storms per synthetic year in the study area. Each of these storms were then normalized by the in-place PRISM 30-year average precipitation, offset (transposed) in the x-y direction by sampling from a uniform distribution, and then de-normalized using the PRISM 30-year average precipitation at the transposed location. This normalization process helped account for topographic differences between in-place and transposed locations, which is especially important in mountainous basins where precipitation has a strong orographic influence.

The SST process was then supplemented with randomly sampled initial baseflow conditions using a gamma distribution and temporally sampling snow water equivalent and temperature from a uniform distribution. The hydrologic simulations were then executed in the cloud and post processed into cloud-native formats for efficient data extraction and analysis. For each synthetic year, the maximum discharge across each simulated event was computed and added to the maximum annual series for each model element. The result is a suite of thousands of possible runoff scenarios which were then statistically analyzed to identify hydrographs with peak discharges corresponding to the flood frequency discharges scoped by the project suitable for forcing the 2D HEC-RAS hydraulic models.

Performed in a complex mountainous region dominated by snowmelt, this method demonstrates the utility of space-for-time substitution for making inferences about infrequent flow quantiles from a period of record whose length is less than the average recurrence interval. This study further demonstrates the application of multi-scale flood frequency where results for flow frequency are obtained at multiple key locations in a watershed, whereas frequency-based storms must be contrived for a specific location due to storm area/area reduction assumptions. Cloud-based infrastructure for handling a multitude of computations in an efficient and manageable way means these kinds of simulations can be expanded to more realizations (to capture uncertainty) and more years per realization (to estimate rarer frequencies). Notably, the methods invoked in this study make no assumption with respect to the frequency of storm spatial patterns, do not require areal reduction factors, and do not rely on published or derived temporal patterns. In their place, the development of a regionalized catalog of observed storm events was used, shifting the statistical analysis from the meteorological forcing to the hydrologic response of the watershed.