Performance and uncertainty analysis in deep learning frameworks for streamflow forecasting via Monte Carlo dropout technique
Xuan-Hien Le, Đoàn Văn Bình, Giha Lee
Abstract
The research was conducted at the Red River Basin, Vietnam. This study evaluated the performance of six deep learning models (Standard LSTM (S_LSTM), Bidirectional LSTM (Bi_LSTM), LSTM with attention mechanism (At_LSTM), Advanced LSTM (Ad_LSTM), Temporal Convolutional Network (TCN), and Sequence-to-Sequence (S2S)) in forecasting streamflow one to four days ahead at the SonTay station. Utilizing Root Mean Square Error, Mean Absolute Error, Nash-Sutcliffe Efficiency, and Symmetric Mean Absolute Percentage Error as metrics, the research systematically compared model outputs against observed data. This analysis included an examination of the importance of different predictors and an assessment of prediction uncertainty through Monte Carlo Dropout techniques. The study reveals that the S_LSTM model exhibits superior short-term forecasting accuracy, while Ad_LSTM shows potential for medium-range forecasts. Discharge data, especially from the SonTay station, emerges as the most significant predictor, with the importance of rainfall data increasing for longer forecast periods. Uncertainty analysis via Monte Carlo Dropout techniques highlights S_LSTM and At_LSTM as the most reliable models. The inclusion of rainfall data slightly reduces short-term forecast accuracy but improves longer-term predictions. These findings underscore the necessity of selecting appropriate forecasting models based on specific temporal scales and hydrological contexts, offering new insights into optimizing streamflow forecasting methodologies for the region. • Evaluated six deep learning models for streamflow forecasting. • S_LSTM model achieved highest NSE (0.983) in one-day ahead forecasts. • Advanced models (Ad_LSTM) show improved performance for three to four-day forecasts. • Incorporation of rainfall data enhances long-term forecast accuracy. • Utilized Monte Carlo Dropout for uncertainty quantification.