Improving Physics-informed, Differentiable Hydrologic Models for Capturing Unseen Extreme Events
- Yalan Song
, - Kamlesh Sawadekar,
- Jonathan M Frame,
- Ming Pan,
- Martyn Clark,
- Wouter J M Knoben,
- Andrew W Wood,
- Trupesh Patel,
- Chaopeng Shen
Kamlesh Sawadekar
Civil and Environmental Engineering, The Pennsylvania State University
Jonathan M Frame
Department of Geological Sciences, University of Alabama
Ming Pan
Center for Western Weather and Water Extremes, Scripps Institution of Oceanography, University of California San Diego
Martyn Clark
Civil Engineering, Schulich School of Engineering, University of Calgary
Wouter J M Knoben
Civil Engineering, Schulich School of Engineering, University of Calgary
Andrew W Wood
U.S. National Science Foundation National Center for Atmospheric Research, Civil and Environmental Engineering, Colorado School of Mines
Trupesh Patel
Department of Computer Science, University of Alabama
Chaopeng Shen
Civil and Environmental Engineering, The Pennsylvania State University
Abstract
• Differentiable models (DMs) better capture extreme events with magnitudes not included in the training data compared to LSTM. • DM optimized for better extreme predictions can still offer good spatial generalization and robust predictions for untrained variables. • We theorize that DM's good extrapolation skill comes from physical constraints like mass conservation and storage-dependent flow.06 Aug 2024Submitted to ESS Open Archive 07 Aug 2024Published in ESS Open Archive
