This paper introduces an innovative method for predicting wireless network traffic in concise temporal intervals for Open Radio Access Networks (O-RAN) using a transformer architecture, which is the machine learning model behind generative AI tools. Depending on the anticipated traffic, the system either launches a reinforcement learning-based traffic steering xApp or a cell sleeping rApp to enhance performance metrics like throughput or energy efficiency. Our simulation results demonstrate that the proposed traffic prediction-based network optimization mechanism matches the performance of standalone RAN applications (rApps/ xApps) that are always on during the whole simulation time while offering on-demand activation. This feature is particularly advantageous during instances of abrupt fluctuations in traffic volume. Rather than persistently operating specific applications irrespective of the actual incoming traffic conditions, the proposed prediction-based method increases the average energy efficiency by 39.7% compared to the "Always on Traffic Steering xApp" and achieves 10.1% increase in throughput compared to the "Always on Cell Sleeping rApp". The simulation has been conducted over 24 hours, emulating a whole day traffic pattern for a dense urban area.

翻译：本文提出了一种创新方法，利用Transformer架构（生成式AI工具背后的机器学习模型）对开放式无线接入网络（O-RAN）中的短时间间隔无线网络流量进行预测。根据预测流量，系统启动基于强化学习的流量导向xApp或小区休眠rApp，以优化吞吐量或能效等性能指标。仿真结果表明，所提出的基于流量预测的网络优化机制在提供按需激活能力的同时，其性能与持续运行的独立RAN应用（rApps/xApps）相当。这一特性在流量剧烈波动场景中尤为有利。相较于始终开启的最优流量导向xApp，该预测方法无需持续运行特定应用，平均能效提升39.7%；相较于始终开启的小区休眠rApp，吞吐量提升10.1%。仿真在24小时内完成，模拟了密集城区全天流量模式。