Enhancing reservoir inflow predictions through dynamic forecast merging

Hydrologic forecast merging (HFM) is critical in enhancing forecast accuracy by addressing uncertainties from model structures and parameters. This study integrates forecasts from spatially large-scale and locally calibrated models to improve reservoir inflow predictions through a dynamic weight estimation approach. The method uses time-series features (TSFs) of streamflow and Bayesian model averaging (BMA) for dynamic weight estimation. The conceptual HBV-EC model is set up on the spatially large Moose River basin in Canada in a semi-distributed fashion, while the GR4J, HYMOD, and SACSMA models are implemented to simulate inflow for the Mesomikenda Lake Dam within the large basin. Both large and local-scale models are calibrated using Canadian Precipitation Analysis (CaPA). Using the Global Deterministic Prediction System (GDPS) dataset, reservoir inflow forecasts are generated up to ten days ahead by applying the calibrated models. Then, the dynamic merging approach is applied to improve inflow forecast accuracy, and the outcomes are compared with the traditional fixed weights  merging method. Results show that while large-scale models generally underperform compared to local-scale models, nonetheless, they provide better fits in specific hydrograph segments. Merging inflow forecasts using the dynamic weight estimation approach shows higher accuracy than the fixed-weight method. Overall, the findings indicate the utility of merging large-scale model forecasts with the local one through the dynamic weight estimation method, offering water resource managers more reliable and precise forecasts for better decision-making.