Authors: Zhennan Shi, Andrew W. Wood –Colorado School of Mines; Leo Lonzarich, Yalan Song, Chaopeng Shen – Pennsylvania State University
Title: Assessing the impact of dynamic versus static parameters on short-range ensemble streamflow forecasts in a hybrid differentiable modeling framework
Presentation Type: Poster Presentation
Abstract: Short-range streamflow forecasting (1–7 days) is critical for flood prediction and water resources operations. Recent advances in hybrid hydrologic modeling enable the integration of differentiable models with deep learning models, including physical (static & dynamic) parameters output from deep learning models; however, their value for short-range ensemble forecasting remains unclear, as most of the research on hybrid modeling has focused less on forecasting (and much less on ensemble forecasting), but on retrospective streamflow simulation. In this presentation, we describe work to assess the impact of using static versus dynamic parameters within a hybrid differentiable modeling framework based on HBV-type representations – namely, the PSU dMG platform – when used for ensemble prediction. To enable this research, we have implemented (and validated) a new restart capability (previously lacking) to preserve pre-forecast model states, including both deep learning model parameters and all physical parameters from the differentiable model, thereby enabling efficient multi-year ensemble hindcast generation without running duplicative spinup sequences. We have also implemented a simple bias-correction for the meteorological forecasts (which are taken from GEFS) and a simple post-processing for the dHBV streamflow predictions. Using trained static and dynamic parameter configurations and focusing on a subset of several dozen well-calibrated basins, we have also evaluated the simulation performance over 5-10 years. Using a series of meteorological hindcasts, we are currently working toward assessing forecast skill across different basins, focusing on error characteristics and ensemble behavior, and showing preliminary results from this analysis.