Probabilistic runoff forecasting by integrating improved conceptual hydrological model with interpretable deep learning approach in a typical karst basin, Southwest China

Abstract

Nonhomogeneous geological features and changing environmental conditions bring great uncertainty and modeling difficulty to karst runoff prediction, probabilistic runoff forecasting is thus of great significance to flood control and water resource management in karst basins. To enhance the accuracy, reliability, and interpretability of runoff prediction, this paper presents an interpretable conceptual-based-data-driven hybrid modeling framework (Interpretable Karst_HyMod-LSTM-KDE), which integrates an improved conceptual hydrological model, a deep learning model, a kernel density estimation model, and the Shapley additive explanation method for runoff point and interval prediction in a typical karst watershed (Chengbi River basin) in Southwest China. The results indicate that the improved conceptual model is superior to the original model, particularly in predicting peak flow. Compared to the standalone and HyMod-LSTM models, the Karst_HyMod-LSTM model improves the NSE, LogNSE, and KGE values in point prediction by 0.6%-11.7%, 0.5%-41.6%, and 5.8%-11.2%, respectively. The Karst_HyMod-LSTM model performs better in predicting extreme flows than a single model. The proposed modeling framework has significant advantages in probabilistic interval prediction with higher coverage and narrower interval width. Additionally, interpretability analyses reveal that recent hydrologic characteristics and subsurface conduit flow contribute the most to runoff variability, and subsurface fracture flow makes important contributions to maintaining perennial base runoff. The proposed hybrid modeling framework can improve the runoff prediction performance in karst basins.