Open Access

Ningpeng Dong, Haoran Hao, Mingxiang Yang, Jianhui Wei, Shiqin Xu, Harald Kunstmann

• Hydrometeorological forecasting is crucial for managing water resources and mitigating the impacts of hydrological extremes. At sub-seasonal scales, readily available hydrometeorological forecast products often exhibit large uncertainties and insufficient accuracies to support decision-making. We propose a deep-learning-based modelling framework for sub-seasonal joint precipitation and streamflow ensemble forecasts for a lead time of up to 30 d. This is achieved by coupling (1) an ensemble of enhanced convolutional neural network (CNN) models with ResNet blocks and a specialized loss function for statistically downscaling of European Centre for Medium-Range Forecasts (ECMWF) ensemble precipitation forecasts to (2) a hybrid hydrologic model integrating the conceptual Xin'anjiang model (XAJ) and the long short-term memory network (LSTM) for ensemble streamflow forecasting (XAJ-LSTM). Applying the modelling framework to the source region of the Yangtze River Basin, results indicate thatHydrometeorological forecasting is crucial for managing water resources and mitigating the impacts of hydrological extremes. At sub-seasonal scales, readily available hydrometeorological forecast products often exhibit large uncertainties and insufficient accuracies to support decision-making. We propose a deep-learning-based modelling framework for sub-seasonal joint precipitation and streamflow ensemble forecasts for a lead time of up to 30 d. This is achieved by coupling (1) an ensemble of enhanced convolutional neural network (CNN) models with ResNet blocks and a specialized loss function for statistically downscaling of European Centre for Medium-Range Forecasts (ECMWF) ensemble precipitation forecasts to (2) a hybrid hydrologic model integrating the conceptual Xin'anjiang model (XAJ) and the long short-term memory network (LSTM) for ensemble streamflow forecasting (XAJ-LSTM). Applying the modelling framework to the source region of the Yangtze River Basin, results indicate that the CNN-based downscaling model exhibits ∼34 % and ∼26 % less root mean squared error (RMSE) than the raw ECMWF forecasts and the quantile mapping (QM) forecasts, respectively, averaged over the 30 d lead time. Similarly, the CNN achieves approximately 6 % and 10 % lower RMSE than raw forecasts and QM for heavy precipitation events. Using these precipitation forecasts as meteorological forcing for the hybrid XAJ-LSTM hydrologic model, we found that forecasted streamflow and flood peaks driven by CNN-based precipitation forecasts have 16 %–33 % lower relative errors and 20 %–31 % lower RMSE compared to those driven by raw forecasts. However, the standalone XAJ model shows only marginal improvements with the same enhanced precipitation forecasts. This highlights the importance of understanding the effectiveness of the hydrologic model as part of the sub-seasonal hydrometeorological modelling chain. Our study is expected to provide implications for leveraging advanced AI techniques to enhance sub-seasonal hydrometeorological forecasting accuracy and operational efficiency for effective water resources management and disaster preparedness.…