Adaptive Graph Learning with Transformer for Multi-Reservoir Inflow Prediction

The introduction of AdaTrip marks a significant advancement in reservoir inflow prediction, a critical area for effective water resource management. Traditional approaches have largely focused on single-reservoir models, neglecting the spatial dependencies that exist among interconnected reservoirs. AdaTrip employs an adaptive, time-varying graph learning framework that constructs dynamic graphs, where reservoirs are represented as nodes and directed edges illustrate hydrological connections. This innovative method utilizes attention mechanisms to identify crucial spatial and temporal dependencies automatically. Evaluations conducted on thirty reservoirs in the Upper Colorado River Basin demonstrate AdaTrip's superiority over existing baselines, particularly in scenarios where data is limited, thanks to its parameter-sharing capabilities. Furthermore, AdaTrip offers interpretable attention maps at both edge and time-step levels, providing valuable insights into hydrological controls th…