Abstract
Passive optical networks (PONs) have matured into a dominant solution for high-speed optical access, leveraging their inherent point-to-multipoint architecture in multi-user scenarios. Recently, emerging applications such as Low-Altitude Economy (LAE) and undersea mining monitoring require technology that supports user mobility and high-speed data links. This paper investigates the feasibility of employing TDM-PON for the above applications. To realize movable Optical Network Units (ONUs), the key challenge is maintaining robust optical link alignment against beam offset induced by moving ONUs. To address this, we propose a Dual-stage Predictive Tracking Network (DPT-Net) that decomposes predictive beam tracking into two sub-neural networks, realizing real-time predictive control. By modifying commercially available 1-Gbit/s Ethernet PON (EPON) transceivers, we experimentally demonstrate a wireless TDM-PON system that supports mobile ONUs for the first time. The neural network-based parallel prediction-and-steering pipeline reduces tracking latency by 38.5% compared to conventional proportional–integral–derivative (PID)-based sequential tracking under the same tracking hardware configuration. The standard deviation of the laser beam position is maintained below 1.64 mm during high-speed motion. Experimental results show that the system achieves average downlink and uplink throughputs of 805 Mbit/s and 798 Mbit/s, respectively, at an ONU velocity of 18 cm/s. In a multi-user scenario, aggregate downlink and uplink rates reach 940 Mbit/s and 868 Mbit/s, respectively, demonstrating seamless integration of optical wireless links into existing PON infrastructures.