Reliable forecast precipitation can support disaster prevention and mitigation and sustainable socio-economic development. Improving forecast precipitation accuracy remains a challenge. Therefore, a novel method for multi-model forecast precipitation integration considering long lead times was proposed based on deep learning. First, the accuracy of numerical forecast precipitation was evaluated under different lead times. Secondly, an integrated model was built by coupling the attention mechanism and a long short-term memory neural network (LSTM). Finally, integrated forecast precipitation was obtained by taking high-precision numerical forecast precipitation as an input and examining its accuracy and applicability. Considering the example of the Yalong River, the results showed the following: (1) numerical forecast precipitation fails to forecast precipitation of a ≥10 mm/d intensity well, and is less applicable in streamflow forecast; (2) traditional machine learning methods for integrating multi-model forecast precipitation fail to forecast precipitation of a ≥25 mm/d intensity; (3) the LSTM-A integration model formed by attention weighting after the LSTM output can combine the advantages of numerical forecast precipitation under different intensities and improve the forecast precipitation accuracy for 7-day lead times; and (4) the LSTM-A integrated forecast precipitation has the best applicability in streamflow forecast, with an NSE above 0.82 and an MRE below 30% with 7-day lead times. These findings contribute to improving precipitation forecast accuracy at different intensities and enhancing defense against extreme weather events.

Accurate forecasting of the tail water level (TWL) is of great importance for the safe and economic operation of hydropower stations. The prediction accuracy is significantly influenced by the backwater effect of downstream tributaries and the operation of adjacent hydropower stations, but the explicit quantification method of the backwater effect is lacking. In this study, a deep-learning-model-based forecasting method for TWL predictions under the backwater effect is developed and applied in the Xiangjiaba (XJB) hydropower station, which is influenced by the backwater effect of downstream tributaries, including the Hengjiang River (HJR) and the Minjiang River (MJR). Firstly, the random forest algorithm was used to analyze the influence of HJR and MJR flows with different lag times on the TWL prediction error of the XJB hydropower station. The results show that the time lags of the backwater effect of HJR and MJR run offs on the TWL of the XJB are 5~7 h and 1~2 h, respectively. Then, the run off thresholds of the HJR and MJR for impacting the TWL of the XJB station are obtained through scenario comparison, and the results show that the run off thresholds of the HJR and the MJR are 700 m³/s and 7000 m³/s, respectively. Finally, based on the analysis of the time lag and the threshold of the backwater effect, a deep learning model (LSTM)-based TWL forecasting method is established and applied to predict the TWL of the XJB station. The results show that the forecasting model has a good predictive performance, with 98.22% of absolute errors less than 20 cm. The mean absolute error over the validation dataset is 5.27 cm, and the maximum absolute error is 63.35 cm. Compared with the LSTM-based prediction model without considering the backwater effect, the mean absolute error decreased by 31%, and the maximum absolute error decreased by 71%.

Mid–long-term streamflow prediction (MLSP) plays a critical role in water resource planning amid growing hydroclimatic and anthropogenic uncertainties. Although AI-based models have demonstrated strong performance in MLSP, their capacity to quantify predictive uncertainty remains limited. To address this challenge, a DeepAR-based probabilistic modeling framework is developed, enabling direct estimation of streamflow distribution parameters and flexible selection of output distributions. The framework is applied to two case studies with distinct hydrological characteristics, where combinations of recurrent model structures (GRU and LSTM) and output distributions (Normal, Student’s t, and Gamma) are systematically evaluated. The results indicate that the choice of output distribution is the most critical factor for predictive performance. The Gamma distribution consistently outperformed those using Normal and Student’s t distributions, due to its ability to better capture the skewed, non-negative nature of streamflow data. Notably, the magnitude of performance gain from using the Gamma distribution is itself region-dependent, proving more significant in the basin with higher streamflow skewness. For instance, in the more skewed Upper Wudongde Reservoir area, the model using LSTM structure and Gamma distribution reduces RMSE by over 27% compared to its Normal-distribution counterpart (from 1407.77 m³/s to 1016.54 m³/s). Furthermore, the Gamma-based models yield superior probabilistic forecasts, achieving not only lower CRPS values but also a more effective balance between high reliability (PICP) and forecast sharpness (MPIW). In contrast, the relative performance between GRU and LSTM architectures was found to be less significant and inconsistent across the different basins. These findings highlight that the DeepAR-based framework delivers consistent enhancement in forecasting accuracy by prioritizing the selection of a physically plausible output distribution, thereby providing stronger and more reliable support for practical applications.

Accurate forecast of tail water level (TWL) is of great importance for the safe and economic operation and management of hydropower stations. The predictive performance is significantly influenced by the backwater effect of downstream hydropower stations and tributaries, but the explicit quantification method of the backwater effect is lacked. In this study, a deep learning model based forecasting framework for TWL predictions is established and applied to forecast TWL of Xiangjiaba (XJB) hydropower stations, which is influenced by the backwater effect of downstream tributaries including Hengjiang River (HJR) and Minjiang River (MJR). Firstly, the lag time of the backwater effect of HJR and MJR is analyzed based on the permutation importance. The results demonstrate that the lag time of backwater effect on the TWL of XJB is 5-7 hours for the HJR and 1-2 hours for the MJR. Then, the runoff thresholds of the HJR and MJR for impacting the TWL of the XJB station are obtained by scenario comparison, and the results show that the thresholds of HJR and MJR are 700 m3/s and 7000 m3/s respectively. Finally, the deep learning methods based TWL forecasting model is established based on the lag time and threshold analysis. The model is used to forecast the TWL in future 48 hours. The results show that the forecasting model has a good predictive performance with 98.22% of absolute errors less than 20 cm. The mean absolute error over the validation dataset is 5.27 cm and the maximum absolute error is 63.35 cm.

Mid–long term streamflow prediction (MLSP) plays a critical role in water resource planning amid growing hydroclimatic and anthropogenic uncertainties. Although AI-based models have demonstrated strong performance in MLSP, their capacity to quantify predictive uncertainty remains limited. To address this challenge, a DeepAR-based probabilistic modeling framework is developed, enabling direct estimation of streamflow distribution parameters and flexible selection of output distributions. The framework is applied to two case studies with distinct hydrological characteristics, where combinations of recurrent model structures (GRU and LSTM) and output distributions (Normal, Student’s t, and Gamma) are systematically evaluated. Results indicate that models employing the Gamma distribution consistently outperform those using Normal and Student’s t distributions. In the Upper Wudongde Reservoir area, the model using LSTM structure and Gamma distribution reduces RMSE from 1407.77 m³/s to 1016.54 m³/s. As the forecast horizon extends, the Gamma-based models demonstrate more reliable probabilistic predictions, reflected by sharper and better-calibrated prediction intervals. This is evidenced by substantially reduced CRPS values at the 18th forecast horizon (521.4 and 1746.6 m³/s), compared to Normal-based models (747.6 and 1877.5 m³/s). Although the improvements in predictive performance achieved by the proposed modeling framework vary depending on the RNN model architecture used and the specific application region, it generally delivers consistent enhancement in forecasting accuracy, thereby providing stronger support for practical applications.