Due to the presence of interference between users, power allocation problems in wireless communication networks are often non convex and require a huge amount of computation. The current graph neural network (GNN) has become an effective computational method used to solve this problem. In order to maximize network transmission speed while reducing computational complexity, a flexible duplex network graph representation method that incorporates device and communication connection attributes into GNN is proposed, and a corresponding flexible duplex graph neural network (FD-GNN) model is constructed. For the first time, the distance between node pairs, channel gain, and neighbors are introduced as dynamic thresholds into FD-GNN to adapt to dynamic environments. Excluding channel state information of neighbors in GNN, pruning edges in FD-GNN reduces computation time and network complexity. Simulation results show that the proposed threshold setting method based on channel gain neighbors has the best performance and reaches 97% of the weighted minimum mean square error (WMMSE), reducing the training time required by 24% compared to Full GANN. The proposed threshold based effectively reduces the time complexity of GNN operations and improves the effectiveness of the algorithm.