Application of lstm and ann models for traffic time headway prediction in expressway tollgates

Abstract

Traffic time headway is essential to support decision-making in safety management, capacity analysis, and service provision. Many studies on the time headway distribution on highways and urban roads serve two primary purposes. The studies that serve the latter purpose, service level, have not been given adequate attention. In fact, at manual toll stations, traffic congestion is still a severe problem. Predicting the time headway at toll stations becomes extremely meaningful when the service providers can allocate resources reasonably, minimizing waiting time in off-peak periods and utilizing resources during high-demand periods. This study applies two modern machine learning methods to predict the time headway at Niigata toll stations, Japan, namely Long Short-term Memory (LSTM), which only requires simple input of time series, and Artificial Neural Network (ANN), which requires some additional external features. The data set is the time headway of vehicles on expressways, along with the weather information and the vehicle’s average speed for five working days. There needs to be a trade-off between computation time, input data complexity, and model accuracy. Thus, tollgate operators could choose a suitable model based on their actual situation.