Explainable and Physics-Aligned Handover Prediction for Cellular-Connected UAV Networks
Year: 2026
Type: article
Abstract: Cellular-connected unmanned aerial vehicles (UAVs) experience frequent handovers at altitude because line-of-sight propagation exposes each device to many cells with comparable signal strength, causing failures that threaten safety-critical missions. Machine learning can anticipate handovers proactively, but most predictors are opaque, and existing interpretable alternatives ignore domain knowledge such as the monotonic relationships defined by 3GPP A3 events. This paper introduces the Hierarchical Attribution Neural Additive Model (HA-NAM). HA-NAM decomposes each prediction into per-feature shape functions in three groups (spatial radio quality, UAV mobility, and handover history), enforces hard monotonicity on 12 of 15 features, and adds lightweight cross-group interaction modules. We evaluate HA-NAM on a 500 ms look-ahead A3 prediction task in a 57-cell urban macro deployment with L3-filtered RSRP, spatially consistent LoS/NLoS states, and trajectory-level data splitting. HA-NAM achieves AUC$0.952 \pm 0.003$and the lowest missed-handover rate among all intrinsically interpretable models. LSTM and GRU baselines yield only a$+0.8$pp AUC gain at the cost of per-feature transparency. On real 5G UAV measurements from the AERPAW testbed, HA-NAM achieves AUC 0.934 to 0.961 across flight-level splits, and simulation-trained shape functions show statistically significant correlation with real handover labels. Edge benchmarks confirm sub-10 ms inference on a Raspberry Pi 4 and sub-6 ms on an NVIDIA Jetson AGX Orin. Physics-constrained additive models can therefore match black-box accuracy while providing auditable, per-feature explanations that transfer from simulation to real aerial networks.
Institutions: Ninevah University, Taibah University
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7164136306