Meta-Continual Mobility Forecasting for Proactive Handover Prediction

TL;DR

This paper introduces a lightweight meta-continual mobility forecasting framework that enhances proactive handover prediction in cellular networks by quickly adapting to non-stationary user mobility patterns, reducing missed handovers and ping-pong events.

Contribution

The paper presents a novel meta-continual learning approach combining a GRU predictor, Reptile initialization, and an EWMA residual detector for fast online adaptation to mobility drift.

Findings

Achieves 4.46 m ADE and 7.79 m FDE in zero-shot settings.

Improves few-shot ADE to 3.71 m at 10-shot.

Enables faster recovery from abrupt drift, 2-3 times quicker than offline models.

Abstract

Short-term mobility forecasting is a core requirement for proactive handover (HO) in cellular networks. Real-world mobility is highly non-stationary: abrupt turns, rapid speed changes, and unpredictable user behavior cause conventional predictors to drift, leading to mistimed or failed handovers. We propose a lightweight meta-continual forecasting framework that integrates a GRU-based predictor, Reptile meta-initialization for fast few-shot adaptation, and an EWMA residual detector that triggers compact online updates only when drift occurs. Evaluated on a reproducible GeoLife and DeepMIMO pipeline, our method achieves 4.46 m ADE and 7.79 m FDE in zero-shot settings, improves few-shot ADE to 3.71 m at 10-shot, and enables recovery from abrupt drift about 2 to 3 times faster than an offline GRU. When applied to downstream HO prediction, the approach improves F1 to 0.83 and AUROC to 0.90, with substantial reductions in missed-HO and ping-pong events. The model is lightweight (128k parameters) and suitable for edge deployment in 5G and 6G systems.