HydroTFT: a cross-basin attention model for multi-horizon rainfall–runoff prediction

Abstract

Streamflow prediction is essential for water resources management, flood forecasting, and climate resilience. Long short-term memory (LSTM) networks have advanced large-sample hydrology through cross-basin learning, but their recurrent architectures have limited ability to capture long-range temporal dependencies, particularly for medium-range forecasting. Meanwhile, existing applications of attention-based models in hydrology have largely relied on historical streamflow observations as model inputs and have been developed for individual basins, leaving their potential for regional rainfall–runoff prediction largely unexplored. We present HydroTFT, an attention-based rainfall–runoff model based on the Temporal Fusion Transformer that performs multi-horizon streamflow prediction using only antecedent meteorological forcings, static catchment attributes, and hydrologically informed engineered features. Trained jointly across 531 Catchment Attributes and Meteorology for Large-sample Studies basins, HydroTFT consistently outperforms entity-aware LSTM, LSTM, PatchTST, and iTransformer for 1 d, 7 d, and 14 d forecasting, while also exceeding both regionally and basin-wise calibrated process-based hydrologic models for 1 d prediction. Ablation experiments identify the attention mechanism as the primary contributor to the performance gains, and interpretability analyses show that the model learns physically meaningful representations of antecedent wetness and hydrologic memory. These results demonstrate that attention-based architectures provide an accurate, interpretable, and scalable framework for regional rainfall–runoff prediction across diverse hydro-climatic conditions.