Network performance prediction is the key to achieving efficient network management of software defined optical networks(SDON), but there is an urgent need for a network performance prediction model that can accurately predict key indicators at limited cost. A graph neural network-based SDON performance prediction model is proposed, which combines BiGRU and Self-Attention mechanisms to learn the complex relationships between network topology, routing, and traffic matrices, accurately estimating the packet delay, jitter, and packet loss rate of the source/destination in the network. This model can be applied to networks that have not been encountered during training. The experimental results show that in different traffic model tests, the proposed model has a significant improvement in average absolute percentage error (MAPE) performance compared to the baseline model.