Local- and Large-Scale Hydrologic Forecast Merging through Time Series Features–Based Dynamic Weights Estimation Framework

Abstract Hydrologic forecast merging has the potential to enhance forecast accuracy by reducing uncertainties related to model structures and the spatial scale of river basins. This study explores the benefits of merging local- and large-scale forecasts to improve hydrologic predictions. Using the Repositionable Aerial Vane Environmental Network (RAVEN) modeling platform, we applied the Hydrologiska Byrans Vattenbalansavdelning–Environment Canada (HBV-EC) model in a semidistributed manner over the large Moose River basin (MRB), Northern Ontario, Canada, as the large-scale model, while three conceptual models [Génie Rural à 4 Paramètres Journalier (GR4J), the hydrological model (HYMOD), and SAC-SMA] were calibrated for two small local subbasins within the MRB. Model calibration was performed using 10 years (2012–21) of Canadian Precipitation Analysis (CaPA) data and the dynamically dimensioned search algorithm. Streamflow forecasts were generated using the Global Deterministic Prediction System dataset in real-time forecasting mode. To merge forecasts, we implemented a time series feature (TSF)-based dynamic weighting (TSF-W) approach within a Bayesian model averaging (BMA) framework and assessed performance over different lead times. Results showed that while local models performed better overall [Nash–Sutcliffe efficiency (NSE) > 0.65] than the large-scale model (NSE < 0.50), the latter captured certain hydrograph characteristics more effectively. The TSF-W merged forecasts outperformed the best local-scale model, particularly for low-flow (by 10%–80%) and high-flow (by 5%–28%) conditions and for extended lead times. These findings highlight the advantages of merging forecasts from models operating at different spatial scales using the TSF-W approach, providing operational hydrologists with more accurate and reliable forecasts for improved decision-making.