Causality-Guided Deep learning for streamflow predicting in a mountainous region

Abstract

Accurate streamflow predictions in mountainous regions are crucial for water resource management and flood mitigation. Deep learning (DL) models, which have been widely used for streamflow predicting recently, can simulate the nonlinear hydrological relationships but may not capture the underlying laws of physics. This study proposed a Causality-Guided Deep Learning (CGDL) model to enhance the streamflow predicting for mountainous regions by incorporating physics-based causal inference and improved multivariate Transfer Entropy (IMTE) algorithm. We assessed the CGDL model through a case study in a mountainous catchment using in situ hydrometeorological variables (precipitation, temperature, and humidity, etc.). The results demonstrated that CGDL outperformed DL and process-based (PB) models, achieving a higher NSE (Nash-Sutcliffe Efficiency) of 0.805, compared to 0.701 for DL and 0.716 for PB during the testing period. Furthermore, CGDL significantly reduced the EHF (Error of High Flow) to −9.7%, versus −14.9% for DL and –22.0% for the PB model in the testing period, highlighting its efficiency in high flow predictions. The CGDL also showed superior robustness and generalization when extending forecast lead time and simulating beyond the bounds of the training data. Additionally, the SHAP analysis indicated that CGDL provided greater interpretability than the DL model. This study demonstrates that integrating causality knowledge into deep learning models has the potential to enhance streamflow predicting in mountainous regions. It is helpful for improving our understanding of hydrological processes and decision-making to issue flood warnings.