Traditional methods for analyzing network performance have limitations, including high costs and over-simplified assumptions, which are not helpful for network administrators managing increasingly complex networks. Therefore, it is necessary to provide a performance prediction method specifically designed for complex networks. This paper introduces the Attention-based Recurrent Graph Convolutional Network (ARGCN), a tailored performance prediction model for Software-defined Networks (SDNs). SDNs extract network data dynamically, and ARGCN, using a Message Passing Neural Network (MPNN) framework, transmits and aggregates information, incorporating a recurrent neural network with an attention mechanism to handle complex dependencies among link nodes. Experimental validation demonstrates the model’s efficiency in forecasting network metrics with over 95% accuracy, even in worst-case scenarios. ARGCN, integrating MPNN, recurrent neural networks, and attention mechanisms, emerges as a powerful tool for administrators dealing with SDN intricacies.