Sequence-Based Deep Learning for Handover Optimization in Dense Urban Cellular Network

Abstract: Efficient handover management remains a critical challenge in dense urban cellular networks, where high cell density, user mobility, and diverse service demands increase the likelihood of unnecessary handovers and ping-pong effects. This paper leverages a real-world, multi-operator drive-test dataset of 30,925 labelled records collected within a 2 km area around Sunway City to investigate sequence-based deep learning approaches for handover detection and avoidance. We formulate handover prediction as a sequence problem and evaluate Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Transformer architectures under Reference Signal Received Power (RSRP)-only and all-feature settings. The integration of multi-dimensional features significantly enhanced handover performance in dense urban cellular networks. The proposed GRU-based model achieved a remarkable 98% reduction in ping-pong handovers, alongside a 46.25% decrease in unnecessary handovers, outperforming the baseline RSRP-only approach which yielded a 22.19% reduction. Furthermore, the model demonstrated a 46% improvement in Time of Stay (ToS), indicating more stable user connections. With an inference time of just 0.91 seconds, the solution proves highly efficient and well-suited for real-time edge deployment scenarios. Compared to the conventional 3GPP A3 algorithm, these improvements demonstrate significant gains in mobility robustness and user Quality of Experience (QoE) improvement. The dataset is released to foster reproducibility and further research in intelligent mobility management for 5G and beyond.