Predicting runoff with precision holds immense importance for flood control, water resource management, and basin ecological dispatch. Deep learning, especially long short-term memory (LSTM) neural networks, has excelled in runoff prediction, often outperforming traditional hydrological models. Recent studies suggest that deep learning models employing the self-attention mechanism, such as Transformer and Informer, can achieve even better results than LSTM. However, research exploring the multi-step runoff prediction capabilities of these novel models across diverse scenarios remains scarce. In this investigation, we introduce a relative location coding-enhanced Informer model, termed Rel-Informer, and compare its performance in rainfall-runoff prediction against the standard Informer, Transformer, and LSTM models. The publicly available CAMELS dataset is utilized for training and validating the models, and four experiments are designed: (1) Individual rainfall-runoff modeling (one model per catchment)
(2) Regional rainfall-runoff modeling (one model per region)
(3) Fine-tuned regional rainfall-runoff modeling (fine-tuned from Experiment 2)
(4) Large-scale rainfall-runoff modeling for ungauged catchments (one model for all catchments). The findings reveal that Rel-Informer consistently performs better than the other models, particularly in short-term runoff predictions (1-3 days ahead). Although regional modeling is less precise than individual modeling, it significantly benefits from fine-tuning. The large-scale regional Rel-Informer model effectively predicts runoff in ungauged catchments, showcasing its potential for widespread runoff prediction. This study underscores the influence of hydrological characteristics, such as snowmelt and baseflow indices, on prediction accuracy. In conclusion, the Rel-Informer model, enhanced with improved relative position encoding, emerges as a promising tool for runoff forecasting, especially in data-rich catchments.