Improving Streamflow Prediction Using a Model Integration Approach

Accurate streamflow prediction is necessary both during wet and dry periods for efficient management of water resources. However, most hydrological models mainly focus on simulating high flows and perform poorly during low-flow or recession periods. Therefore, past studies have resorted to various calibration techniques to allow rainfall-runoff (R-R) models better capture recession flow dynamics. In the present study, we propose integrating two structurally different models and utilising their relative strengths to improve overall streamflow prediction. The proposed framework integrates a conceptual rainfall-runoff model (HBV) and a simple power-law regression (PLR) such that the former is utilised for high-flow prediction and the latter for low-flow prediction. We compared the performance of this integrated model framework (HBV-PLR) with the original HBV model using data from 108 basins in the United States. It was found that the 25^th, 50^th, and 75^th percentiles of mean absolute error (MAE) for HBV, respectively, improved from (0.47, 0.62, and 0.77) to (0.38, 0.50, and 0.67) using the HBV-PLR integrated framework. Similarly, the median Nash-Sutcliffe Efficiency (NSE) during the recession improved from 0.65 to 0.74. Here, we also argue that forcing HBV model to simulate low-flow dynamics by calibrating it using an objective function biased towards lower values may not lead to a prediction as accurate as HBV-PLR. Therefore, a model integration approach is a better option than using a single model to improve streamflow prediction during different flow regimes.