Enhancing hydrological modeling with transformers: a case study for 24-h streamflow prediction.

Area of Science:

• Hydrology
• Data Science
• Environmental Engineering

Background:

• Accurate 24-hour streamflow forecasting is crucial for water resource management and flood prediction.
• Traditional forecasting methods often struggle to capture complex temporal dynamics in hydrological data.
• The application of advanced deep learning models, particularly transformers, remains underexplored in streamflow forecasting.

Purpose of the Study:

• To evaluate the performance of deep learning models for 24-hour streamflow forecasting.
• To compare the efficacy of the transformer architecture against other models like LSTM, Seq2Seq, and GRU.
• To assess the impact of data extension techniques (zero-padding and persistence) on forecasting accuracy.

Main Methods:

• Comparative analysis of five streamflow forecasting models: persistence, LSTM, Seq2Seq, GRU, and transformer.
• Evaluation across four distinct geographical regions.
• Performance assessment using Nash-Sutcliffe Efficiency (NSE), Pearson's r, and normalized root mean square error (NRMSE).

Main Results:

• The transformer model demonstrated superior performance in capturing temporal dependencies and patterns in streamflow data.
• Transformer models achieved substantial improvements in NSE scores, up to 20% higher than other models.
• The study identified the transformer as the most accurate and reliable model for streamflow forecasting among those tested.

Conclusions:

• Advanced deep learning models, especially the transformer architecture, offer significant advantages for hydrological modeling.
• The transformer's ability to capture complex patterns enhances streamflow forecasting accuracy and reliability.
• Implementing transformer models can lead to more effective water resource management and improved flood prediction strategies.