Everywhere and Locally Relevant Streamflow Simulations in Hydrological Modeling

Contemporary initiatives such as the Digital Twins and paradigms like the Hyper-resolution Modeling are pushing the boundaries of high-performance computing, bringing resolutions of large-scale hydrological models (HM) down to sub-kilometer, in attempts of making "Locally Relevant Hydrological Models Everywhere" a reality. As global water models converge towards this hyper-resolution, we notice the insufficient attention given to model scalability i.e., consistent simulations across different resolutions using the same set of model parameters. Besides, the way distributed HMs have been resolving the stream network with grids requires high resolution model runs to offset the errors, especially at locations with smaller catchment area, hence the calling of the hyper-resolution modeling.

We equip the mesoscale hydrological model (mHM, https://mhm-ufz.org) with a novel stream network upscaling scheme called subgrid catchment conservation or SCC. We hypothesize the conservation of the subgrid catchment area to have threefold effect in distributed HMs: 1) improvement in the consistency of model performance across modeling resolutions, 2) streamflow simulations become both plausible everywhere and locally relevant, and 3) the long-standing conundrum of streamflow estimation at multiple gauging stations within a grid cell will be solved.

The experimental setup is a single modeling domain encompassing 187 GRDC streamflow stations in the Rhine river basin. The wide range of catchment sizes at the gauges (1 km² to more than 150,000 km²), notable presence of proximate clusters of gauges, and good data availability (average availability of 30 years) makes Rhine an apt case for the hypotheses testing. We compare streamflow simulated by the SCC with the D8 stream network upscaling scheme, both with default parameter set. SCC shows remarkable streamflow scalability with nine out of 10 stations exceeding the mean flow benchmark across 1 km to 100 km model resolutions. In comparison, D8 shows poor scalability where the percentage of stations exceeding the benchmark reduces drastically from ≈80 % at 1 km to 50 % at 12 km to <5 % at 100 km. SCC performs significantly better than D8 at smaller catchments (<100 km²) e.g., KGE of 0.34 (±0.02) at Rappengraben (1 km²), at all model resolutions. This demonstrates, for the first time, the ability of a distributed HM to produce locally relevant streamflow everywhere, irrespective of model resolution. The Rhine's 25 km configuration encompasses grid cells featuring as many as 10 gauges within a single grid. The mean annual streamflow at these stations unrealistically exhibit identical values with the D8, whereas the SCC simulations yield values close to the observations at each station.

SCC requires the locations of interest (e.g., gauges) in the model configuration i.e., the catchment area conservation can not be achieved post-process. Still, SCC remains a significant advancement over the older (EAM, DMM) as well as the state-of-the-art (FLOW, IHU) stream network upscaling schemes. SCC finds practical application in switchable systems that require consistent simulations across resolutions demanded by end-users. The opportunity to improve hydrological forecasts using SCC also remains to be explored. But most importantly, the outcome of this research reminds us about "the overlooked hallmark of model reliability i.e., scalability".