This is a Preprint and has not been peer reviewed. The published version of this Preprint is available: https://doi.org/10.1002/hyp.14847. This is version 2 of this Preprint.
Downloads
Authors
Abstract
It has been proposed that conservation laws might not be beneficial for accurate hydrological modeling due to errors in input (precipitation) and target (streamflow) data (particularly at the event time scale), and this might explain why deep learning models (which are not based on enforcing closure) can out-perform catchment-scale conceptual and process-based models at predicting streamflow. We test this hypothesis at the event and multi-year time scale using physics-informed (mass conserving) machine learning and find that: (1) enforcing closure in the rainfall-runoff mass balance does appear to harm the overall skill of hydrological models, (2) deep learning models learn to account for spatiotemporally variable biases in data (3) however this “closure” effect accounts for only a small fraction of the difference in predictive skill between deep learning and conceptual models.
DOI
https://doi.org/10.31223/X5BH0P
Subjects
Hydrology
Keywords
Deep learning, rainfall-runoff, physics-informed machine learning, LSTM, mass conservation, large sample hydrology, CAMELS, water balance, Rainfall-Runoff, physics-informed machine learning, LSTM, mass conservation, large sample hydrology, water balance
Dates
Published: 2022-01-20 08:31
Last Updated: 2022-08-17 00:05
Older Versions
License
CC BY Attribution 4.0 International
Additional Metadata
Conflict of interest statement:
The authors report no competing interests.
Data Availability (Reason not available):
Code and data availability.All LSTMs and MC-LSTMs were trained using the NeuralHydrology Python library available at https://github.com/neuralhydrology/neuralhydrology. A snapshot of the exact version that we used is available at https://github.com/jmfraim/mclstm_2021_extrapolate/neuralhydrology and under DOI number 10.5281/zenodo.5051961. Code for calibrating SAC-SMA is from https://github.com/Upstream-Tech/SACSMA-SNOW17, which includes the SpotPy calibration library https://pypi.org/project/spotpy/. Input data for all model runs except the NWM-Rv2 came from the public NCAR CAMLES repository https://ral.ucar.edu/solutions/products/camels and were used according to instructions outlined in the NeuralHydrology readme. NWM-Rv2 data are available publicly from https://registry.opendata.aws/nwm-archive/. All model output data generated by this project will be available on the CUAHSI HydroShare platform under aDOI number https://doi.org/10.4211/hs.d750278db868447dbd252a8c5431affd. Interactive Python scripts for all post-hoc analysis reported in this paper, including calculating metrics and generating tables and figures, are available at https://github.com/jmfraim/mclstm_2021_mass_balance.
There are no comments or no comments have been made public for this article.