To reveal the influence of different physical mechanism fusion schemes on the groundwater level prediction accuracy of graph neural network (GCN), six graph convolutional network (GCN) structures with different physical mechanism fusion schemes were designed based on the daily groundwater level data from 114 monitoring stations in the Poyang Lake Basin by comprehensively considering the physical factors such as spatial correlation of groundwater levels, surface water-groundwater connectivity, and rainfall lag effect. The groundwater level prediction accuracies of the six schemes were compared. The results from the Poyang Lake Basin case validation show that the GCN demonstrated excellent overall simulation capability and accuracy when predicting the dynamic changes of groundwater levels; the R 2 of all schemes exceeds 0.9. The scheme fusing one-month-ahead rainfall and surface water-groundwater connectivity attributes achieves the highest accuracy (with an MAE of 0.060 1 m), improving the accuracy by about 6.8% compared with the single distance matrix scheme. Taking rainfall lag into account characterizes better the delayed response of groundwater levels to rainfall; moreover, the connectivity attribute matrix yields better RMSE improvements for phreatic well water-level predictions than for confined wells.