Author(s): Cunwen Niu; Yaobing Sui; Chunfeng Hao
Linked Author(s): Cunwen NIU, Chunfeng Hao
Keywords: Data-driven models; WEP model; Physical interpretability; Surrogate modelling; Semi-arid watershed; Dynamic parameterization
Abstract: Flood disasters, characterized by their extensive impact and severe destructiveness, frequently result in substantial casualties and substantial economic losses. Against the backdrop of global climate change, the increasing frequency of extreme precipitation events, coupled with the continuous expansion of impervious surfaces, has markedly elevated flood risks. Consequently, developing accurate flood forecasting methods for proactive risk mitigation has become an urgent priority. Currently, flood simulation relies primarily on two categories of models: process-based (conceptual or physical) models and data-driven (Artificial Intelligence, AI) models. However, the performance of process-based models is often unstable due to uncertainties in input data, model structure, and parameterization. While data-driven models possess strong nonlinear fitting capabilities, their "black-box" nature limits their physical interpretability and generalizability. Although existing research has attempted to couple these two paradigms to balance physical mechanism representation with simulation accuracy, most coupling frameworks only permit one-way information transfer. Moreover, the mechanistic models integrated in such efforts are typically simplified lumped or conceptual models. While this avoids the complexity of high-dimensional parameter calibration, it sacrifices the realistic representation of spatial heterogeneity and physical processes within a watershed. To address these limitations, this study focused on the semi-arid Dali River Basin. We constructed a distributed hydrological model (Water and Energy transfer Process, WEP) and trained/fine-tuned various data-driven models, including CNN (Convolutional Neural Networks), LSTM (Long Short-Term Memory), ConvLSTM (Convolutional Long Short-Term Memory), Chronos, and Sundial. By systematically comparing the performance of the process-based model versus data-driven models, as well as among different data-driven models in streamflow simulation, we propose a novel deep mutual-feedback coupling framework that integrates WEP with data-driven approaches.
Year: 2026