Improving low flow prediction from hydrologic models using alternative model calibration and post-processing techniques

Abstract. Accurate prediction of low flow series and statistics remains a major challenge in hydrologic modeling. This study evaluates the effectiveness of combining model calibration strategies and post-processing approaches to improve low flow simulation from hydrologic models. WRF-Hydro, a fully distributed deterministic watershed model, is calibrated, post-processed, and evaluated using alternative methods that only require observed and simulated streamflows. Model calibration is performed using alternative objective functions that target different flow magnitudes. This study applies two post-processing approaches, quantile mapping bias correction and stochastic ensemble generation using log-streamflow ratios, to three unregulated watersheds in New York State. The skill of the model simulations and post-processing techniques is evaluated by assessing prediction of low flow series and statistics. Calibration alone could not address conditional bias or reduce the variability of low streamflow estimators. While quantile mapping removes conditional bias, estimators of low flow series and design statistics still exhibited large variability. In contrast, ensemble-based methods led to considerable reductions in both bias and variability of low flow series and design statistic estimators. The ensemble methods performed better when statistics were obtained from an average single streamflow trace than as the average of the statistic across all ensembles. In addition, during a forecasting simulation, resampling of errors from the calibration period was shown to improve low flow estimators during forecast periods when observed streamflows are unknown. These findings suggest that improving low flow simulations requires shifting emphasis from calibration and bias correction methods, toward the development of ensemble-based post-processing approaches.