Runoff forecasting stops chasing accuracy alone: five storage states keep mass balance, and dry catchments still get recession flows

Rainfall-runoff models now treat mass conservation and interpretable states as goals alongside fitting the flow curve. Most earlier data-driven models remained difficult to interpret and did not explicitly enforce physical conservation laws. Across 50 CAMELS catchments covering snowmelt-dominated, recent-rainfall, historical-rainfall and arid regimes, GCHM reproduced streamflow dynamics, including in arid catchments.

The model uses five storage states for canopy interception, snow, soil moisture, groundwater and streamflow generation, and precipitation is partitioned among them through gates. The gates are constrained and govern precipitation partitioning, evapotranspiration, infiltration, seepage, runoff and baseflow while learning context-dependent process relationships. Ablation experiments showed simplified architectures may perform adequately under wet conditions but struggle to represent flow persistence and recession during dry periods, when groundwater storage sustains low flows.

The learned gates recover hydrologically plausible relationships among snowmelt, runoff, seepage and evapotranspiration. Performance is assessed with the Jawad-Kling-Gupta Efficiency metric, which accounts for stationary and non-stationary hydrologic behavior. The model is forced with NOAA AORC meteorology and remotely sensed vegetation and land-surface data, and compared with ablated variants and recurrent neural network benchmarks.

Intended for researchers and engineers working on hydrologic modeling and interpretable machine learning, the evaluation covers 50 CAMELS catchments spanning snowmelt-dominated, recent-rainfall, historical-rainfall and arid regimes, forced by NOAA AORC meteorology and remotely sensed vegetation and land-surface data. The material positions the structure as a foundation for hydrologic modeling and counterfactual analysis.

The material reports that GCHM reproduces streamflow dynamics across diverse hydroclimatic regimes while maintaining interpretable states and fluxes, and notes that simplified architectures may perform adequately under wet conditions but struggle to represent flow persistence and recession during dry periods. These conclusions rest on summary-level information; specific metric values and per-catchment performance are not given in the material.