A Model Combination Approach for Improving Streamflow Prediction

Water-related issues are becoming more widespread due to the increased frequency and severity of extreme climatic events. Under such a scenario, accurate streamflow prediction is necessary both during wet and dry periods. Most hydrological models developed so far focus on simulating high flows and thus perform poorly during dry or recession periods. Several studies in the past, therefore, resorted to calibration techniques to enable rainfall-runoff (R-R) models better capture recession flow dynamics. In the present study, a model combination approach is proposed, which utilises the relative strengths of two structurally different models to improve overall streamflow prediction. In particular, the proposed framework combines a conceptual rainfall-runoff model (HBV) with a power-law regression (PLR) model such that the former is used for high flow prediction and the latter for low flow prediction. The proposed framework (HBV-PLR) is evaluated in 108 basins in the United States, and its performance is compared with the original HBV model. The results show that the 25th, 50th, and 75th percentiles of mean absolute error (MAE), which were (0.47, 0.62, and 0.77), respectively, for the HBV, improved to (0.38, 0.50, and 0.67) using the HBV-PLR combination. Similarly, the median Nash-Sutcliffe Efficiency (NSE) during the recession improved from 0.65 to 0.74. In this study, we also demonstrated that the HBV, even being calibrated using an objective function biased towards lower values, may not predict low flows as precisely as the HBV-PLR. Our results suggest that a model combination approach may be a better option than a single model.