Traffic congestion undermines mobility, economic productivity, and safety in rapidly urbanising regions, yet evidence-based hourly traffic-flow forecasting studies using real-world South African freeway data remain limited. This study addresses this gap by developing and rigorously benchmarking machine-learning models for hourly traffic-volume prediction on the N3 Durban – Pietermaritzburg corridor national motorway using continuous count-station observations. A unified modelling pipeline was implemented comprising preprocessing, standardisation, an 80:20 holdout split, and 10-fold cross-validated hyperparameter optimisation using GridSearchCV. Four algorithms namely k-Nearest Neighbours, Decision Tree, Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP) were trained and evaluated using R 2 , RMSE, and MAE. SVR achieved the best overall performance, with an R 2 = 0.9952, RMSE = 36.25 and MAE = 36.11, closely followed by MLP (R 2 = 0.9952). The key advance is a reproducible, like for like comparative evaluation of multiple machine learning models under identical validation and tuning conditions for an under-studied African freeway context, providing an empirical performance ranking and practical guidance on model selection for hourly forecasting. These findings support data-driven operational planning and intelligent transport interventions on strategic corridors and offer a transferable benchmarking framework for similar emerging economy highway networks.
Emenike et al. (Wed,) studied this question.