Probabilistic physics‐guided deep neural networks with recurrence and attention mechanisms for interpretable daily streamflow simulation

Sadeghi Tabas, Sadegh, Samadi, Vidya, Wilson, Catherine ORCID: https://orcid.org/0000-0002-7128-590X and Bhattacharya, Biswa 2025. Probabilistic physics‐guided deep neural networks with recurrence and attention mechanisms for interpretable daily streamflow simulation. Water Resources Research 61 (9) , e2025WR040173. 10.1029/2025wr040173

PDF - Published Version Available under License Creative Commons Attribution. Download (8MB)

Abstract

Plain Language Summary: Explanations supporting the output of deep neural networks (DNNs) are crucial in rainfall‐runoff modeling, where experts require far more information from the model than a simple classical simulation to support modeling diagnosis. This research delves into exploring interpretable probabilistic DNNs by developing Deep Autoregressive Recurrent (DeepAR) and Temporal Fusion Transformer (TFT) models. These models were rigorously evaluated against traditional hydrologic methods, both conceptual and physics‐based, emphasizing the integration of catchment physical attributes into the models for daily streamflow simulations across the continental United States (CONUS). Leveraging quantile regression to evaluate predictive uncertainty, the physics‐guided TFT model notably outperformed other models by demonstrating superior predictive capabilities, particularly in managing high and low flow fluctuations. Notably, this study showed that physics‐guided TFT model can effectively leverage its interpretable multi‐head attention mechanism to weigh the importance of temporal flow dynamics based on the relationships between forcing data, catchment physical attributes, and streamflow records. The findings of this study show promising results of transformer rainfall‐runoff simulations, thereby highlighting its robustness in effectively utilizing physical attributes and improving model interpretability.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering
Additional Information:	License information from Publisher: LICENSE 1: URL: http://creativecommons.org/licenses/by/4.0/
Publisher:	Wiley
ISSN:	0043-1397
Date of First Compliant Deposit:	29 September 2025
Date of Acceptance:	31 July 2025
Last Modified:	29 Sep 2025 15:30
URI:	https://orca.cardiff.ac.uk/id/eprint/181400

Actions (repository staff only)

Edit Item