A Lightweight Transformer Framework for Multi-Scale, Spatiotemporal Air Traffic Density Forecasting in Regional Airport Safety Research

Abstract

Large-scale global air traffic models impose computational demands that can limit their practical deployment in resource-constrained, region-specific aviation safety research. This paper presents a lightweight, reproducible Transformer pipeline for localized, multi-scale air traffic density forecasting. Using five years of publicly available weekly (Monday) ADS-B snapshot data for the San Francisco Bay Area (254 observed Mondays, 2017-2022; 3,051,163 position records), we use K-Means to group hourly air traffic conditions into eight states. We then train three compact encoder-only Transformers (0.8M parameters each), one per operational scale: same-day tactical (12-hour), weekly (7-day), and strategic (two-month) forecasting. The learned advantage is concentrated at the tactical scale, where same-day context is available; at the two-month horizon, the Transformer does not outperform hourly climatology, the strongest non-learned baseline. The compact model achieves 60.4% tactical accuracy, compared with 57.8% for climatology. This advantage is centered on atypical hours, when the observed traffic state differs from the state normally seen at that hour. Such deviations occur in 42% of test hours. The model correctly predicts the observed state in 23% of the atypical hours, while climatology, which only predicts the typical state, cannot predict any of them. The entire pipeline runs in about an hour on a 2019 laptop CPU, and all code and derived data are released to support reproduction. We argue that such compact, transparent, locality-first models offer a practical and underexplored approach to applied aviation safety research, particularly in regions where large infrastructure investments are infeasible.