Exploiting Stage Information for Prediction of Switching Times of Traffic Actuated Signals Using Machine Learning

Traffic jams and stops in front of (signalized) intersections are causing a significant amount of road traffic emissions. Reliable signal timing estimation integrated into Green Light Optimal Speed Advisory systems is one of the keys to minimizing unnecessary braking and accelerating in front of traffic lights. In this paper, we present experiments based on recorded real-world process data of traffic signals to predict the switching times of traffic actuated signals using machine learning. Traffic lights sign a limited number of signal combinations. We summarize these combinations into sets to obtain one feature depicting the signal state. This paper focuses on whether the inclusion of the (estimated) state as an input feature can achieve higher accuracy for the estimation of the switching times than those based only on signal information. Implementing XGBoost and Bayesian Networks, we found that a higher accuracy and a lower root mean square error are achieved by using states. Whether the actual or estimated next state improves the quality of the prediction, the most depends on the degree of traffic dependency and traffic load of the traffic light.