LSTM-based Packet Loss Prediction for Adaptive Polar Coding in QUIC enabled Wireless Networks

This paper proposes a long short-term memory (LSTM)-based packet loss prediction enhancing polar code adaptation scheme in QUIC (LLPA-QUIC), addressing the challenge of degraded QUIC transmission efficiency caused by spurious loss detection and inefficient forward error correction (FEC) in dynamic wireless networks. Unlike prior works focusing solely on FEC redundancy optimisation, LLPA-QUIC pioneers a comprehensive FEC framework that integrates both loss detection threshold adaptation and packet loss rate (PLR) prediction. Firstly, an adaptive loss detection based on LSTM is designed to adjust detection thresholds by learning from historical reordering patterns, mitigating false packet loss judgments. Secondly, we utilise the packet loss statistics to develop a Kalman filter-LSTM model for real-time noise suppression and PLR prediction. Thirdly, based on the predicted PLR, we devise an adaptive packet-level polar coding scheme to balance recovery efficiency and bandwidth overhead. Experimental results indicate that our proposal improves the efficiency of QUIC transmission. Specifically, compared to rQUIC, our proposed scheme reduces file transfer completion time by up to 47.3% in emulated satellite environments and by 33.8% and 29.1% in the controlled 5G and Wi-Fi networks, respectively. These gains highlight LLPA-QUIC’s potential to support reliable and efficient QUIC transmission in next-generation wireless systems.