Answers are grounded in this paper and cite their sources.
Traditional residual post‐processing methods have demonstrated substantial improvements in simulated hydrological response; however, they typically aggregate multiple sources of uncertainty into a single lumped residual and apply uniform statistical treatment across all hydrologic conditions. This uniform treatment overlooks the fact that error characteristics can vary systematically across different rainfall and flow regimes. Given the fundamentally different processes governing catchment storage accumulation and depletion, mainly driven by the presence or absence of rainfall, there is a strong rationale for employing distinct residual models tailored to specific hydrologic states. We introduce here a novel two‐stage residual modeling framework that represents residuals into their hydrological‐state dependent components. Using hydrologic understanding to define the states and storage‐related information from the pre‐existing hydrologic model to inform residual prediction, we use deep learning to characterize residual behavior in each state and improve streamflow simulations via state‐dependent error correction. We illustrate the usefulness of the proposed model using selected catchments across the varying climatic zones in Australia. In the Warren River Catchment, the proposed model reduces RMSE by 59%–91% overall and by 65%–86% in the upper‐tail flow regimes, while also yielding consistent gains in NSE. These results suggest that hydrologically conditioned residual modeling can improve both bulk predictive performance and the representation of hydrologically important high‐flow events. The framework also retains utility under limited hydrologic model calibration, which may be valuable in operational settings where pre‐existing hydrologic models continue to be used.
No takes yet. Share an insight, caveat, or question.
Hammad et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: