\(\delta\) HBV-hydroDL

Code Release

The codes for differentiable hydrologic models are released at this zenodo link

Summary

Differentiable hydrologic modeling provides a generic approach to unify machine learning and physical models. Here we first implemented the hydrologic model HBV in the PyTorch platform with Automatic Differentiation, so that we can freely embed neural network components to parameterize and evolve the structure of the original model. We flexibly evolved the original HBV structure under differentiable modeling by using multi-component computation and introducing dynamic parameters (Feng et al., 2022). These so-called differentiable hydrologic models can achieve state-of-the-art hydrologic simulation accuracy tested in public benchmark dataset while also keep physical process clarity. Different from the purely machine learning methods, differentiable models can provide accurate simulations for a full set of untrained hydrologic variables. We further compared the extrapolation ability of different models (prediction in ungauged regions) in both continental and global scales. Differentiable models stand out showing superior spatial generalization performance compared with traditional parameter regionalization and purely machine learning methods (Feng et al., 2023; Feng et al., 2023). These characteristics show differentiable models can be great candidates as the next-generation global hydrologic model.

Bibtex Citation

@article{feng2022WRR,
author = {Feng, Dapeng and Liu, Jiangtao and Lawson, Kathryn and Shen, Chaopeng},
title = {Differentiable, Learnable, Regionalized Process-Based Models With Multiphysical Outputs can Approach State-Of-The-Art Hydrologic Prediction Accuracy},
journal = {Water Resources Research},
volume = {58},
number = {10},
pages = {e2022WR032404},
keywords = {rainfall runoff, differentiable programming, machine learning, physical model, differentiable hydrology, LSTM},
doi = {https://doi.org/10.1029/2022WR032404},
year = {2022}
}
@Article{feng2023HESS,
AUTHOR = {Feng, D. and Beck, H. and Lawson, K. and Shen, C.},
TITLE = {The suitability of differentiable, physics-informed machine learning
hydrologic models for ungauged regions and climate change impact assessment},
JOURNAL = {Hydrology and Earth System Sciences},
VOLUME = {27},
YEAR = {2023},
NUMBER = {12},
PAGES = {2357--2373},
URL = {https://hess.copernicus.org/articles/27/2357/2023/},
DOI = {10.5194/hess-27-2357-2023}
}
@article{feng2023GMDD,
  title={Deep Dive into Global Hydrologic Simulations: Harnessing the Power of Deep Learning and Physics-informed Differentiable Models ($\delta$HBV-globe1. 0-hydroDL)},
  author={Feng, Dapeng and Beck, Hylke and de Bruijn, Jens and Sahu, Reetik Kumar and Satoh, Yusuke and Wada, Yoshihide and Liu, Jiangtao and Pan, Ming and Lawson, Kathryn and Shen, Chaopeng},
  journal={Geoscientific Model Development Discussions},
  volume={2023},
  pages={1--23},
  year={2023},
  publisher={G{\"o}ttingen, Germany}
}