Transition-Aware Conformal Prediction for GPS-Aided UAV-to-Ground mmWave Beam Tracking

Abstract

Reliable receive-beam alignment at a ground base station (BS) is essential for uplink millimeter-wave (mmWave) communication from a moving uncrewed aerial vehicle (UAV), but existing conformal beam-selection methods largely treat successive decisions independently. We propose Sequential Conformal Prediction with Empirical Transition-Distance Penalties (SCP-ETP), a transition-aware conformal beam-tracking method that augments the Adaptive Prediction Sets (APS) nonconformity score with a negative-log empirical transition-distance penalty conditioned on the previously observed beam. The transition-distance statistics are estimated once from training trajectories before conformal calibration, so the resulting score function is fixed before calibration and no retraining of the underlying beam predictor is required. Under exchangeability of the augmented calibration and test tuples, the standard split-conformal marginal coverage result applies; for the trajectory-level evaluation considered in this paper, coverage is reported empirically. On DeepSense 6G Scenario 23, SCP-ETP achieves empirical coverage of 0.945 with an average prediction-set size of 3.61 under realistic closed-loop operation, reducing the probing budget by approximately 45–50% relative to Global APS. Additional transition-only and subsampling analyses show that temporal persistence is the dominant source of compact prediction sets at the native sampling rate, while SCP-ETP provides additional efficiency gains over transition-only conformal prediction within the empirically supported persistence regime.