Predicting truck parking occupancy using machine learning

Abstract

The logistics industry faces an increasing shortage of truck parking spots. This results in illegal parking or fatigued driving with hazardous consequences for traffic safety, as truck drivers have no insight into future availability of parking spots. Accurate short-term predictions of parking lot occupation are required to aid drivers in planning their routes and rest stops. To obtain such predictions, this research compares a variety of machine learning algorithms, concluding that decision trees are most suitable for real-time application. The model is trained on real-world data containing 1.5 years of truck parking measurements, obtained from a truck parking in Deventer, the Netherlands. We find that – contrasting to car parking, which is influenced by factors such as the weather – a model using only temporal features and historical occupancy yields the best results. For one-hour ahead predictions, we obtain an RMSE of 0.0029, with a training time of 4 seconds and predictions being sufficiently fast for real-time deployment. The main contributions of this research are (i) a machine learning approach for predicting truck parking occupation, (ii) insights into relevant predictive features, and (iii) a case study. From a practical perspective, we propose an architecture for a dynamic prediction tool, which can be used by truck drivers, parking managers and road authorities to improve truck parking utilization. Future research can build upon the machine learning approach and use the prediction model for other truck parking areas.