Enhancing runoff simulation by combining superflex with deep learning methods in China's Qinghai Lake Basin, Northeast Tibetan Plateau
Study region: The Qinghai Lake Basin on the Northeast Tibetan Plateau. Study focus: Coupling physical models with deep learning methods offers potential advantages for runoff simulation, optimizing their interaction remains a crucial challenge. This study investigates hybrid models combining the Sup...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-06-01
|
| Series: | Journal of Hydrology: Regional Studies |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2214581825001557 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Study region: The Qinghai Lake Basin on the Northeast Tibetan Plateau. Study focus: Coupling physical models with deep learning methods offers potential advantages for runoff simulation, optimizing their interaction remains a crucial challenge. This study investigates hybrid models combining the Superflex hydrological model with Gated Recurrent Unit (GRU) for runoff modeling and simulation in the Qinghai Lake Basin. Our approach leverages Superflex as a pre-training step for the neural networks and incorporates key process variables from the physical model as inputs to the network, creating a physically-driven deep learning framework. We systematically explore various input-output combinations and selections of deep learning models, and comprehensively evaluated the most effective configuration for this specific basin. Furthermore, we use the SHAP method to reveal how meteorological factors influence runoff and their complex relationships, making the results interpretable. New hydrological insights for the region: Compared to hydrological model, hybrid models significantly improve performance by incorporating internal hydrological variables and meteorological data as input features, reducing the error by over 50 % in Buha River Basin. We further observed that although different deep learning architectures exhibit varying performance outcomes, the GRU-based models consistently demonstrated significantly superior predictive capabilities. In addition, we use SHAP to understand the internal operation of the model, revealing how meteorological factors affect runoff and their complex relationships, successfully unveiling the ''black box'' nature of deep learning models. |
|---|---|
| ISSN: | 2214-5818 |