Handover optimization in O-RAN faces growing challenges due to heterogeneous user mobility patterns and rapidly varying radio conditions. Existing ML-based handover schemes typically operate at the near-RT layer, which lack awareness of the mobility-mode and struggle to incorporate a longer-term predictive context. This paper proposes a multi-modal mobility-aware optimization framework in which all predictive intelligence, including mobility mode classification, short-horizon trajectory and RSRP forecasting, and a PPO Actor--Critic policy, runs entirely inside an rApp in the non-RT RIC. The rApp generates per-UE ranked neighbour-cell recommendations and delivers them to the existing handover xApp through the A1 interface. The xApp combines these rankings with instantaneous E2 measurements and performs the final standards-compliant handover decision. This hierarchical design preserves low-latency execution in the xApp while enabling the rApp to supply richer and mode-specific predictive guidance. Evaluation using mobility traces demonstrates that the proposed approach reduces ping-pong handover events and improves handover reliability compared to conventional 3GPP A3-based and ML-based baselines.

翻译：O-RAN中的切换优化因异构用户移动模式与快速变化的无线条件而面临日益严峻的挑战。现有的基于机器学习的切换方案通常在近实时（near-RT）层运行，缺乏对移动模式的感知能力，且难以融入长期预测上下文。本文提出一种多模态移动感知优化框架，其中所有预测智能（包括移动模式分类、短时域轨迹与RSRP预测，以及PPO Actor-Critic策略）完全运行于非实时RAN智能控制器（non-RT RIC）内的rApp中。该rApp生成面向每个用户的邻小区排序推荐，并通过A1接口将其传递至现有的切换xApp。xApp将这些排序结果与即时E2测量值相结合，执行最终符合标准的切换决策。这种分层设计在保持xApp低延迟执行能力的同时，使rApp能够提供更丰富且针对特定移动模式的预测指导。基于移动轨迹的评估表明，与传统基于3GPP A3的方案及基于机器学习的基线方法相比，所提方法能有效减少乒乓切换事件并提升切换可靠性。