The accuracy of long-term runoff models can be increased through the input of local weather variables and global climate indices. However, existing methods do not effectively extract important information from complex input factors across various temporal and spatial dimensions, thereby contributing to inaccurate predictions of long-term runoff. In this study, local–global–temporal attention mechanisms (LGTA) were proposed for capturing crucial information on global climate indices on monthly, annual, and interannual time scales. The graph attention network (GAT) was employed to extract geographical topological information of meteorological stations, based on remotely sensed elevation data. A long-term runoff prediction model was established based on long-short-term memory (LSTM) integrated with GAT and LGTA, referred to as GAT–LGTA–LSTM. The proposed model was compared to five comparative models (LGTA–LSTM, GAT–GTA–LSTM, GTA–LSTM, GAT–GA–LSTM, GA–LSTM). The models were applied to forecast the long-term runoff at Luning and Pingshan stations in China. The results indicated that the GAT–LGTA–LSTM model demonstrated the best forecasting performance among the comparative models. The Nash–Sutcliffe Efficiency (NSE) of GAT–LGTA–LSTM at the Luning and Pingshan stations reached 0.87 and 0.89, respectively. Compared to the GA–LSTM benchmark model, the GAT–LGTA–LSTM model demonstrated an average increase in NSE of 0.07, an average increase in Kling–Gupta Efficiency (KGE) of 0.08, and an average reduction in mean absolute percent error (MAPE) of 0.12. The excellent performance of the proposed model is attributed to the following: (1) local attention mechanism assigns a higher weight to key global climate indices at a monthly scale, enhancing the ability of global and temporal attention mechanisms to capture the critical information at annual and interannual scales and (2) the global attention mechanism integrated with GAT effectively extracts crucial temporal and spatial information from precipitation and remotely-sensed elevation data. Furthermore, attention visualization reveals that various global climate indices contribute differently to runoff predictions across distinct months. The global climate indices corresponding to specific seasons or months should be selected to forecast the respective monthly runoff.

Monthly extreme precipitation (EP) forecasts are of vital importance in water resources management and storage behind dams. Machine learning (ML) is extensively used for forecasting monthly EP, and improvements in model performance have been a popular issue. The innovation of this study is summarized as follows. First, a distance correlation-Pearson correlation (DC-PC) method was proposed to identify the complex nonlinear relationship between global sea surface temperature (SST) and EP and select key input factors from SST. Second, a random forest (RF) model was used for forecasting monthly EP, and the physical mechanism of EP was obtained based on the feature importance (FI) of RF and DC–PC relationship. The middle and lower reaches of the Yangtze River (MLYR) were selected as a case study, and monthly EP in summer (June, July and August) was forecasted. Furthermore, the physical mechanism between key predictors with a large proportion of FI and EP was investigated. Results showed that the proposed model had high accuracy and robustness, in which R² in the test period was above 0.81, and RMSE as well as MAE were below 10 mm. Meanwhile, the key predictors in the high SST years could cause eastward extension of the South Asian High, westward extension of the Western Pacific Subtropical High, water vapor rising motion and an increase in the duration of atmospheric rivers exceeding 66 h, which lead to increasing EP in the MLYR. The results indicated that the DC–PC method could replace Pearson correlation for investigating the nonlinear relationship between SST and EP, as well as for selecting the factors. Further, the key predictors that account for a large proportion of FI can be used for explaining the physical mechanism of EP and directing forecasts.

The partition of surface runoff and infiltration is crucial in hydrologic modeling. To improve the flood prediction, we designed four strategies to explore the influences of the runoff partition method on the flexible hybrid runoff generation model. The runoff partition strategies consist of a hydrological model without the runoff partition module, a two-source runoff partition method, an improved two-source runoff partition method considering the heterogeneity of the subsurface topography and land cover, and a three-source runoff partition method. The Xin’anjiang hydrological model was used as the modeling framework to simulate a six-hourly stream flow for the Xun River watershed in Shaanxi Province, China. And the saturation-excess runoff generation and infiltration-excess runoff generation mechanisms were combined to construct the flexible hybrid runoff generation model. The performances of the four strategies were compared and analyzed based on the continuous flow discharge as well as the flood events. The runoff components analysis method was used to test the model’s conformity with the reality of the watershed. The results showed that the three-source runoff partition method was not applicable to the flexible hybrid runoff generation model because it overestimated the surface runoff and almost ignored the subsurface stormflow runoff. The improved two-source runoff partition method outperformed the others as it considered the heterogeneity of the watershed.