DEVELOPMENT OF ADVANCED TECHNIQUES FOR MERGING LOCAL AND LARGE-SCALE HYDROLOGIC FORECASTS

Abstract

Hydrologic modeling and forecasting play a critical role in water resource planning and management, helping to mitigate the socio-economic impacts of droughts and flood hazards. The accuracy and reliability of streamflow forecasting systems largely depend on the structure of hydrologic models and the quality of their inputs. To address the inherent limitations of individual models, multi-model or ensemble forecast merging has emerged as a promising approach to improve forecast accuracy by reducing uncertainty. However, the static nature of traditional merging weights introduces additional uncertainties, as model performance varies dynamically over time. This study aims to develop a dynamic framework for merging forecasts that adjusts weights over forecast lead times. Additionally, the framework explores whether integrating local and large-scale model forecasts can enhance accuracy. The research began with a comprehensive literature review to establish the state-of-the-art in hydrologic forecast merging (HFM) techniques, identifying their applications, advantages, and limitations. Recognizing the unique challenges of large-scale basins containing regulated sub-basins, the study first investigated calibration strategies for hydrologic models in highly regulated basins. Using time series features (TSFs) and Bayesian model averaging (BMA), a dynamic weights (TSF-Ws) estimation framework was then developed. Finally, the method was applied to integrate local and large-scale forecasts from natural and regulated sub-basins. A multi-site, multi-variable, and multi-objective calibration strategy, incorporating targeted water level operational constraints, is effective for highly regulated basins where dam operation and natural streamflow data are unavailable. Dynamic forecast merging significantly improves accuracy by reducing uncertainty, with the TSF-Ws-based method outperforming traditional static weights merging approaches. Integrating large-scale forecasts with locally calibrated ones proved advantageous when applying dynamic weights. The outcomes of this research provide valuable insights and potential for enhanced hydrologic forecasting, especially in real-time flood forecasting. However, the current weight estimation relies solely on BMA. Future research could explore advanced techniques, such as emerging physical system-based learning, for further estimating TSF-Ws to enhance forecast accuracy and adaptability.