Residual bidirectional gated recurrent unit with spatiotemporal shortcuts and dual-path attention for robust...

While runoff forecasting models based on neural networks demonstrate the superior performance over traditional methods in the hydrological field, they also face challenges such as insufficient depth, rapid decay in accuracy with increasing lead times, and poor interpretability. To address these problems, we innovatively propose an interpretable hourly deep runoff forecasting framework capable of robust multi-lead-time forecasting: Spatio-Temporal Residual Bidirectional Gated Recurrent Unit (STResBiGRU). This framework fundamentally relies on a significantly enhanced BiGRU network, which not only captures the rich hidden relationships in long sequences through bidirectional traversal of the input sequence but also innovatively introduces unique temporal shortcut connections between cells. In the overall network architecture, the novel design of dual residual block pathways and dense spatial shortcut connections considerably deepens the recurrent architecture-based neural network. The integration of spatiotemporal shortcut connections effectively improves the mode of information transmission in different dimensions, mitigates the vanishing gradient problem caused by deep networks and long sequences, and enhances feature reuse. Furthermore, during the fusion of bidirectional information flows in BiGRU, we also propose a lightweight dual-path attention (DPA) mechanism module to explicitly model the interdependencies between features in both temporal and spatial dimensions of the forward and reverse GRU output feature maps, respectively, selectively amplifying key features to enhance the model's temporal robustness in processing long sequences. The training phase of the model employs the snapshot ensemble method, which is able to achieve excellent generalization performance of integrating multiple models through a single training process. Additionally, we use interpretable machine learning (IML) method to visualize the model's forecasting process, thereby enhancing algorithmic transparency. Ablation and comparative experiments on a real-world dataset from the Columbia River in the United States demonstrate that the proposed STResBiGRU framework significantly enhances accuracy and temporal robustness over mainstream models, providing reliable decision support for stakeholders.