Improving medium-range streamflow forecasts over South Korea with a dual-encoder transformer model.

Area of Science:

• Hydrology
• Artificial Intelligence
• Meteorology

Background:

• Accurate hydrological forecasts are crucial for water security, flood preparedness, and agriculture.
• Medium-range (1-10 days) hydrological forecasting is essential for effective water resource management.

Purpose of the Study:

• To investigate the potential of a novel Transformer neural network, Dualformer, for medium-range hydrological forecasting in South Korea.
• To evaluate Dualformer's performance in predicting daily streamflow using historical and forecast meteorological data.

Main Methods:

• Developed Dualformer, a dual-encoder Transformer model integrating historical and forecast meteorological data.
• Trained and evaluated models using historical runoff, meteorological variables, geographic data, and Global Ensemble Forecast System (GEFSv12) reforecasts.
• Assessed performance across 473 grid cells against a benchmark approach.

Main Results:

• Dualformer demonstrated competitive performance, particularly for short lead times (1-day lead Nash-Sutcliffe efficiency of 0.664 vs. 0.535 for benchmark).
• The model effectively integrates information from two distinct data sources (historical and forecast weather).
• Performance improved with the utilization of larger datasets.

Conclusions:

• Dualformer shows significant potential for improving medium-range hydrological forecasting.
• The model's ability to leverage diverse data sources enhances its predictive accuracy.
• Future work may involve incorporating additional inputs and refining model structures for further enhancements.