Abstract
Accurate short-term forecasting of traffic flow on airport access corridors is important for anticipatory traffic management and for preserving passenger and freight reliability. This study evaluates a hybrid modelling framework that integrates Particle Swarm Optimization (PSO) for automated hyperparameter and feature search with Long Short-Term Memory (LSTM) networks, applied to five-minute lane-level sensor data from a monitoring site serving O.R. Tambo International Airport (site 1949) over September 2019. The study formulates short-horizon forecasting as congestion risk prediction derived from traffic flow and speed measurements. A reproducible pipeline was developed that performs deterministic preprocessing, lag and rolling-statistics feature engineering, chronological train–validate–test splitting, and nested time-aware hyperparameter tuning via PSO (20 particles × 30 iterations), with final models assessed on classification, calibration, sequence-shape (DTW), and explainability (SHAP) metrics. The PSO-LSTM hybrid delivered modest but consistent improvements in probabilistic fidelity and produced interpretable, theory-coherent attributions compared with untuned LSTM, stacked GRU, PSO-GRU, and an ARIMA proxy. Threshold analyses showed the model can be calibrated to operating points favoring precision or recall, though trade-offs remain and the stacked GRU was a close competitor on several metrics. Inferential testing indicated mixed statistical significance due to limited temporal scope and class imbalance, and PSO tuning entailed notable computational cost. An ARIMA baseline was included as a classical time-series comparator by modelling the underlying traffic signal and mapping outputs to the same congestion target for like-for-like evaluation. Overall, the results suggest that swarm-guided tuning can yield practically useful calibration and selection benefits in airport access forecasting, while highlighting the need for broader temporal coverage, additional exogenous inputs, and further validation for deployment.