This paper investigates the hypothesis that the lack of enforced mass conservation is the main reason that deep learning models outperform traditional hydrology models.
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, 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 using physics-informed 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, however (3) this “closure” effect accounts for only a small fraction of the difference in predictive skill between deep learning and conceptual models.
Paper
(EarthArxiv preprint)
Code
All experiments were made with the NeuralHydrology Python library. The exact snapshot for reproducing the results can be found in this GitHub repository.
Citation
@Article{frame2021extreme,
author = {Frame, J. and Kratzert, F. and Gupta, H. V. and Ullrich, P. and Nearing, G. S.},
title = {On Strictly Enforced Mass Conservation Constraints for Modeling the Rainfall-Runoff Process},
journal = {Hydrological Processes},
volume = {in review},
year = {2022},
}