As 6G networks expand, they generate large amounts of data and connect various devices, challenging conventional network management techniques. To address these challenges, a Speed-optimized Long Short-Term Memory (SP-LSTM) model and Reinforcement Learning (RL) have been developed to predict network congestion and optimize routing, respectively, by considering link ID, time, throughput metrics, and congestion levels. However, the SP-LSTM may struggle to adapt to sudden changes in network conditions and capture complex spatial dependencies effectively. This limitation could influence its accuracy in predicting congestion in dynamic 6G networks where spatial and temporal interactions play a crucial role. Improving the model's utilization of spatial and temporal data is vital to enhance its predictive capabilities and address network congestion effectively. Hence, this manuscript introduces a novel Speed-optimized Attention-based Hybrid Graph Convolutional Network-LSTM model (SPAH-GCN-LSTM) to predict network congestion in 6G networks. This model combines global and local spatial correlations in traffic data through global and local spatial-temporal modules to enhance prediction accuracy. The global module utilizes a global correlation matrix and SP-LSTM to capture global spatial-temporal relationship. The local module combines a Fully Connected Layer (FCL), GCN, and SP-LSTM to obtain local spatial relationship. Then, the outputs of these modules are fused using a soft attention strategy to focus on important features for accurate prediction. Moreover, the RL approach is used for dynamic routing based on the predicted congestion conditions and real-time feedback. Finally, experimental results show the superior performance of the SPAH-GCN-LSTM model compared to existing models in 6G networks.