Deep LSTM-FCN with Self-Attention for 5G Dual Connectivity Uplink Resource Forecasting

To ensure a smooth transition to 5G, most carriers adopt the Non-Standalone (NSA) mode, integrating it with existing 4G infrastructure via dual connectivity (DC). This approach enhances network performance by leveraging both 5G and 4G resources. However, uplink throughput in DC is more variable than in conventional 4G due to the unstable 5G radio channel conditions, impacting both mobile and stationary scenarios. Traditional network management methods are now inadequate, necessitating autonomous calibration to optimize system parameters and processing time. While many machine learning algorithms focus on downlink (DL) resource prediction, they are less effective for uplink (UL) traffic, particularly in 5G DC where channel asymmetry favors DL throughput. This paper proposes a self-attention mechanism combined with Fully Convolutional Network (FCN) and Long Short-Term Memory (LSTM) to predict uplink resources for real-time scenario. The model captures both global and local features in DC traffic sequences. The evaluation results demonstrate that the model effectively captures the temporal patterns and dynamic behavior of uplink Transport Block Size (TBS) values in both 4G and 5G networks.