ABSTRACT Forecasting sub‐daily streamflow in regions with limited discharge data, such as newly gauged or sparsely monitored catchments, poses challenges to Deep Learning (DL) models, which require large training datasets. These scenarios, known as Prediction in Newly Gauged Basins (PNB) or Prediction with Limited Data (PLD), require models that remain accurate even when local training data are limited. This study uses the Temporal Fusion Transformer (TFT), a DL model that captures complex dependencies and generates probabilistic forecasts, supporting hydrological decision‐making under uncertainty. A single TFT model trained on 15 diverse gauged catchments was evaluated for a zero‐shot generalization on 10 unseen catchments using hydrological, meteorological, and physical attributes (catchments located across the eastern and south‐eastern United States), simulating real‐world PNB/PLD conditions. Results showed that the TFT model outperformed the National Water Model in continuous hydrograph simulation and flood forecasting. It achieved higher accuracy (average NSE = 0.655 vs. 0.58 for NWMv3.0), with over 91% of observed values falling within its 95% prediction intervals. The model identifies dominant hydrologic drivers, predicts short‐term floods with high accuracy across lead times, and removes the need for catchment‐specific model training and tuning, thereby simplifying model deployment and improving scalability for operational hydrology.
Heidari et al. (Wed,) studied this question.