Prediction and Proactive Congestion Mitigation in Software-Defined Networks Using a Hybrid Spatio-Temporal GCN–LSTM and Deep Reinforcement Learning Framework
TL;DR
This paper presents a hybrid Artificial Intelligence (AI)-driven framework that integrates a Spatio-Temporal Graph Convolutional Network with Long Short-Term Memory (GCN–LSTM) for traffic prediction and a Dueling Deep Q-Network (DQN) for proactive routing optimization.
Abstract
Software-Defined Networking (SDN) has emerged as a promising paradigm for centralized network management; however, conventional routing protocols remain predominantly reactive, responding only after congestion has already occurred. This reactive behavior increases end-to-end latency, packet loss, and inefficient bandwidth utilization, particularly under dynamic traffic conditions. This paper presents a hybrid Artificial Intelligence (AI)-driven framework that integrates a Spatio-Temporal Graph Convolutional Network with Long Short-Term Memory (GCN–LSTM) for traffic prediction and a Dueling Deep Q-Network (DQN) for proactive routing optimization. The proposed framework operates within an SDN architecture using the Ryu controller and OpenFlow 1.3 while network scenarios are emulated in Mininet. The GCN component captures spatial dependencies among network nodes, whereas the LSTM models temporal traffic evolution to forecast future congestion. These predictions guide the reinforcement learning agent to proactively update forwarding paths before congestion develops. Experimental evaluation demonstrates significant improvements over conventional routing approaches, including lower prediction error (MAPE 0.94%), reduced average latency (29.1 ms), zero packet loss during burst traffic, and an aggregate throughput of 8.1 Gbps. The proposed framework demonstrates the effectiveness of combining deep learning and reinforcement learning to enable intelligent, predictive, and adaptive traffic management in modern SDN environments. GCN–LSTM MAE of 1.05 Kbps at t+1 and 3.18 Kbps at t+5, RMSE 4.22 Kbps, MAPE 0.94%, average latency 29.1 ms, packet loss 0.00% in the reported burst-traffic experiment, throughput 8.1 Gbps, and Dueling-DQN convergence at approximately episode 350 versus approximately 700 for standard DQN. Calculated comparative improvements were also added.