Improving streamflow predictions in data-scarce nested basins through residual transfer and post-processing

Reliable streamflow prediction in data-scarce and partially gauged river basins remains a major challenge for hydrological modeling. Even after calibration, conceptual rainfall-runoff models often retain systematic errors, particularly in low- and intermediate-flow regimes, due to structural limitations and parameter equifinality. This study proposes a residual-based post-processing framework that leverages temporally correlated simulation errors and enables residual transfer across nested catchments to improve streamflow predictions where downstream gauges are absent. Using daily flow and meteorological data from the River Don basin (UK), we calibrated a HYMOD model for four sub-catchments and applied a first-order autoregressive (AR(1)) model to residuals between observed and simulated flows. All four post-processing schemes consistently outperformed the baseline calibration, improving the Nash–Sutcliffe Efficiency (NSE) from 0.749 to 0.826 and reducing the Root Mean Square Error (RMSE) from 8.67 to 7.23 m3/s. For a downstream ungauged site, residuals from upstream models were scaled by catchment area and transferred, raising NSE from 0.744 to 0.804. Flow duration and regime-specific analyses confirmed the largest improvements during intermediate and low flows, with RMSE reductions of 25–38%. These emphasize the value of explicitly modeling residuals, capturing temporal error correlations more effectively than conventional methods. In future research, integrating multi-objective calibrations that balance high- and low-flow priorities, exploring higher-order AR models, and considering the spatial heterogeneity of catchment characteristics offer promising strategies to further refine error-correction approaches.