Interpretable Hourly Flood Forecasting Leveraging Deep Learning Recurrence and Attention Mechanisms

Abstract Timely and precise flood forecasting at subdaily intervals is crucial for mitigating the adverse impact of flood events, particularly through early warnings that enable effective emergency response. This study investigated the efficiency of the Temporal Fusion Transformer (TFT), a state-of-the-art sequence-to-sequence deep learning architecture, for its efficacy in forecasting flood events with lead times from 1 to 24 h across the Atlantic and Gulf coastal regions of the US, extending from Maine to Louisiana. The TFT framework integrated time-dependent meteorological and hydrological variables along with several invariant catchment attributes to effectively model the temporal variability and spatial diversity of hydrological responses across different regions. Model performance was systematically assessed utilizing the Nash–Sutcliffe efficiency (NSE), relative peak bias (RPB), and peak time bias (PTB), benchmarked across multiple forecast horizons and evaluated against both discharge observations and simulations from the physically based National Water Model (NWMv3.0). The TFT demonstrated superior predictive capability, achieving an average NSE of 0.83 for continuous hydrograph simulations across all catchments during the testing phase, significantly outperforming NWMv3.0, which had an NSE of 0.53. The proposed TFT framework incorporates antecedent observed streamflow as an autoregressive input, whereas NWMv3.0 relies primarily on meteorological forcings and physics-based process representations without directly utilizing the same observed discharge information during forecasting. Consequently, part of the superior short-term predictive performance of the TFT framework, particularly at shorter lead times, may be attributed to this additional information advantage associated with streamflow persistence and autocorrelation. An attention mechanism was utilized to perform feature importance analysis, which identified total precipitation, longwave radiation flux, and convective fraction as the dominant dynamic meteorological forcings predictors, and the catchment mean slope was the most important static attribute, underscoring the model’s interpretability for the area of interest. To enhance reliability further, a quantile regression loss function was incorporated into TFT to generate probabilistic discharge forecasts (instead of point forecasts) across seven quantiles. The associated 95% prediction uncertainty (95PPU) was evaluated rigorously using established hydrological uncertainty metrics, including the R -factor, P -factor, and the total uncertainty index (TUI). By leveraging the advanced capabilities of the TFT architecture, such as its capacity for quantifying feature importance and delivering detailed probabilistic forecasts, this study contributes substantive advancements to short-term flood forecasting methodologies. These outcomes support data-informed decision-making and resilience-building strategies in flood-prone regions, reinforcing the applicability of deep learning in operational hydrologic forecasting.