The paper presents a novel deep learning model, Encoder-Decoder Double-Layer Long Short-Term Memory (ED-DLSTM), for cross-regional streamflow and flood forecasting at a global scale. The model combines static spatial attributes and temporal forcing data to improve the accuracy of predictions in both gauged and ungauged catchments. Key contributions include: 1. **Model Performance**: ED-DLSTM achieved a mean Nash-Sutcliffe Efficiency (NSE) of 0.75 across over 2,000 catchments in the United States, Canada, Central Europe, and the United Kingdom, outperforming traditional hydrologic models. 2. **Transferability**: The model was applied to 160 ungauged catchments in Chile, achieving NSE >0 in 76.9% of cases, demonstrating its ability to generalize to new regions. 3. **Interpretability**: The model's internal parameters were visualized to explain how it recognizes and differentiates between different catchments, showing that it can learn "hydrological general knowledge" from different training sets. The study highlights the potential of deep learning methods to overcome the limitations of traditional hydrologic models, particularly in regions with sparse data or no physical parameters. The ED-DLSTM model's ability to capture and encode spatial attributes significantly enhances its time series forecasting capabilities, making it a promising tool for global-scale hydrological forecasting.The paper presents a novel deep learning model, Encoder-Decoder Double-Layer Long Short-Term Memory (ED-DLSTM), for cross-regional streamflow and flood forecasting at a global scale. The model combines static spatial attributes and temporal forcing data to improve the accuracy of predictions in both gauged and ungauged catchments. Key contributions include: 1. **Model Performance**: ED-DLSTM achieved a mean Nash-Sutcliffe Efficiency (NSE) of 0.75 across over 2,000 catchments in the United States, Canada, Central Europe, and the United Kingdom, outperforming traditional hydrologic models. 2. **Transferability**: The model was applied to 160 ungauged catchments in Chile, achieving NSE >0 in 76.9% of cases, demonstrating its ability to generalize to new regions. 3. **Interpretability**: The model's internal parameters were visualized to explain how it recognizes and differentiates between different catchments, showing that it can learn "hydrological general knowledge" from different training sets. The study highlights the potential of deep learning methods to overcome the limitations of traditional hydrologic models, particularly in regions with sparse data or no physical parameters. The ED-DLSTM model's ability to capture and encode spatial attributes significantly enhances its time series forecasting capabilities, making it a promising tool for global-scale hydrological forecasting.