Differences in extremes and uncertainties in future runoff simulations using SWAT and LSTM for SSP scenarios

Area of Science:

• Hydrology
• Climate Science
• Machine Learning

Background:

• Accurate hydrological modeling is crucial for water resource management.
• Climate change necessitates reliable projections of future runoff.
• Comparing deep learning (LSTM) and physically-based (SWAT) models offers insights into hydrological simulation capabilities.

Purpose of the Study:

• To compare the performance of LSTM and SWAT in simulating and projecting runoff.
• To evaluate model capabilities under different climate change scenarios (SSPs).
• To assess model uncertainties in historical and future runoff estimations.

Main Methods:

• Utilized 11 CMIP6 Global Climate Models (GCMs) with bias correction via quantile mapping.
• Employed LSTM and SWAT for runoff simulation and projection for SSPs 2-4.5 and 5-8.5.
• Quantified uncertainties using Bayesian Model Averaging (BMA) and reliability ensemble averaging (REA).

Main Results:

• Bias-corrected GCMs significantly improved climate variable replication.
• LSTM demonstrated superior performance over SWAT in reproducing observed runoff.
• SWAT projected a 17.7% increase, while LSTM projected a 13.6% decrease in future runoff.

Conclusions:

• LSTM exhibits greater potential for accurate runoff simulation compared to SWAT.
• Model uncertainties differ between historical and projected periods, and between models.
• Findings provide valuable perspectives on deep learning versus physically-based models in hydrological applications.