Uncertainty and driving factor analysis of streamflow forecasting for closed-basin and interval-basin: Based on a probabilistic and interpretable deep learning model

Study region

closed- and interval-basin in the Yangtze River basin, China.

Study focus

Accurate streamflow forecasting and understanding its drivers are essential in hydrology, with deep learning (DL) technologies being increasingly employed. However, challenges persist, including hyperparameter optimization, lack of interpretability, and uncertainty quantification, with most studies focusing on closed-basins and limited research on human-influenced interval-basins. Therefore, this study proposes a hybrid DL model, DTA-CBAS, which combines several techniques for probabilistic and interpretable streamflow forecasting. The model was applied to both closed- and interval-basins to investigate the driving mechanisms of streamflow variation across different basin types.

New hydrological insights for the region

The results demonstrated DTA-CBAS outperformed several state-of-the-art models (i.e., mean NSE: from 0.89 to 0.98 and from 0.87 to 0.98, RE: from 4.55 % to 1.55 % and from 5.48 % to 4.39 % in closed-and interval-basins respectively), with uncertainty analysis revealing greater uncertainty in interval-basin compared to closed-basin (i.e., PINAW:38.78 %, 46.95 %, and 52.98 % higher than interval-basin in 95 %, 75 %, and 50 % confidence intervals), suggesting human regulation increased forecast uncertainty. The driver analysis revealed factors affect streamflow differently across basin types: precipitation, evapotranspiration and temperature were key drivers in closed-basin, while upstream inflow is more significant in interval-basin. Further analysis indicated streamflow variation results from the combined effects of multiple factors rather than a single factor. This study highlighted the role of DTA-CBAS in improving streamflow forecasting.