Hierarchical attention network for short-term runoff forecasting

Accurate prediction of runoff is critical concerning reservoir management and disaster preparedness. Data-driven methods are progressively applied to runoff prediction tasks and have led to impressive results. However, existing data-driven methods are hardly considered to the runoff generation process and the spatial characteristics of basins in the models due to the lack of a priori knowledge guidance. Here a structured approach is provided to develop the perceptual model for runoff generation and model the behavior in groups at different locations and scales; considering the hierarchical structure of basin systems, a short-term runoff forecasting model with spatial perception and scale interaction, i.e., the hierarchical attention network, is developed based on the encoder-decoder structure and attention mechanism. Compared to the single- and multi-step prediction performance of the six baseline models, the NSE improved by an average of 2.41, 9.68, and 12.14%, respectively. This implies that incorporating basin-related knowledge in modeling and considering runoff generation processes and spatial connectivity can improve prediction accuracy, and the necessity of considering conceptual mechanisms in data-driven models.