Evaluation of Temporal Fusion Transformers for Streamflow Forecasting: Effects of Input Configuration and Cross-Basin Performance

Accurate short-term streamflow forecasting is essential for water resources management, motivating the evaluation of deep learning (DL) frameworks using hydrometeorological inputs. This study assesses the Temporal Fusion Transformer (TFT) across Peninsular India, focusing on lookback length, predictor selection, and transferability. Results show that predictability is dominated by short-term hydrological memory, with lagged streamflow as the primary driver. Linear autoregressive models perform best with short lookbacks but degrade with longer inputs, whereas TFT captures nonlinear interactions and can extract long-term information from meteorological forcing. Optimal model memory is basin-dependent based on their underlying hydrological processes. Static catchment attributes provide modest basin-specific improvements, emphasizing parsimonious feature selection. Multi-basin training enhances robustness, particularly in heterogeneous settings. Transfer experiments show that prediction in pseudo-ungauged basins is feasible, with intra-cluster transfer preserving performance and cross-cluster transfer revealing regional asymmetries. Overall, TFT provides a scalable framework for forecasting in hydroclimatically diverse, data-scarce regions.