Abstract: With the deepening of power market reform, short-term generation revenue forecasting in the spot market has become a key link for hydropower plants to formulate generation strategies and improve economic benefits. This paper proposes a short-term revenue forecasting model based on a graph neural network (GNN). The forecasting problem is modeled as a multi-graph spatiotemporal joint optimization problem. A dual-channel GNN framework is constructed to extract grid topology features and historical temporal features. In addition, a quantile loss function and a physical-constraint penalty term are introduced to enhance model robustness, effectively addressing the low prediction accuracy and weak anti-fluctuation ability of traditional models.