A hybrid deep learning framework for robust and efficient streamflow prediction under varying data availability

Accurate daily runoff prediction is essential for water resource management and flood control. However, intensifying climate change has amplified the nonlinearity and non-stationarity of runoff series, creating a critical trilemma for deep learning methods: the inability to simultaneously achieve local and long-range feature modeling, computational efficiency, and physical interpretability. To address this challenge, we propose a novel deep learning architecture—the Deep Temporal Gated Residual Network (DTGR)—that synergistically integrates three neural network modules: (i) the Deep Convolutional Residual Module (DCRM), which employs multi-scale causal convolutions to extract hierarchical local temporal features; (ii) the Temporal Encoding Module (TEM), which utilizes linear multi-head attention to model long-range dependencies with O(n) computational complexity; and (iii) the Gate Decoding Module (GDM), which implements adaptive gating mechanisms to dynamically fuse local and global representations. The model was validated across four hydrologically distinct basins—rainfall-driven, snowmelt-dominated, and mixed regimes—achieving Nash-Sutcliffe Efficiency coefficients of 0.826–0.951 for univariate prediction, outperforming seven benchmarks by 6.7%–39.7% in overall accuracy, with NRMSE reductions of 15.7%–57.3%. Under multivariate inputs, NRMSE further drops 20.9%–55.2%, with cross-site NSE variation as low as 0.014. Computationally, DTGR cuts training time by 47% and memory usage by 38%–44% versus Transformers. Mechanistically, SHAP interpretability analysis confirms that DTGR’s learned feature importance aligns with basin-specific physical processes—temperature-driven snowmelt dynamics and precipitation-runoff responses—validating that the architecture achieves interpretability without sacrificing predictive accuracy. These results demonstrate that modular co-design of neural network components can resolve fundamental trade-offs in hydrological deep learning, delivering an o delivering a computationally efficient solution for real-time flood forecasting.