Introducing time series features based dynamic weights estimation framework for hydrologic forecast merging

Accurate and reliable hydrologic forecasting through multi-model ensemble averaging is crucial for reducing uncertainty, which aids in effective water resources management and flood risk mitigation. This study addresses the research gap of the limited application of time-varying weights in hydrologic forecast merging, as existing methods rely on weights that do not adapt to changes in model performance over time. We propose a novel framework utilizing time series features (TSFs) of daily streamflow and Bayesian model averaging (BMA) to dynamically adjust merging weights, referred to as TSF-Ws. The methodology involves generating ensemble forecasts, adjusting weights dynamically using TSFs, and comparing the accuracy of these forecasts with traditional streamflow-based weights, referred to as Q-Ws, merging across different forecast horizons. The results demonstrate that TSF-Ws significantly improve forecast performance, particularly for longer lead times, indicating more accurate and reliable deterministic and probabilistic forecasts. Moreover, TSF-Ws based merging achieves higher performance than Q-Ws for deterministic high and low flow forecasts. Furthermore, this newly developed approach reduces the uncertainty bound for probabilistic peak flow predictions. Overall, the proposed TSF-Ws estimation framework can serve as a robust tool for enhancing hydrologic forecast merging, providing significant improvements in accuracy and reliability over traditional methods. These improvements have important implications for water resource management and flood risk assessment.