ML-based Prediction Framework for varying Traffic Signal Control Strategies and its GLOSA-application

This paper deals with the development and results of a prediction framework for traffic light control systems as well as the usage and benefits of such predictions in green light optimal speed advisory (GLOSA) scenarios. Various machine learning methods like support vector machines, neural networks or reinforcement learning were evaluated for their applicability in the prediction context and compared based on their efficiency and most importantly accuracy. The resulting prediction framework uses decision tree ensemble models combined with certain model knowledge to forecast different control strategies. This method was chosen due to its best performance in various test scenarios. Very high accuracy and fidelity were achieved for standard control methods like fixed-time, time-of-day-based and 'ordinary' traffic-based programs. Only for the more sophisticated model predictive control which was tested lower accuracies were achieved. For the upcoming GLOSA application the penetration of equipped vehicles was varied for different traffic scenarios and control strategies. Results showcase high potentials for enhancing urban mobility and reducing environmental impact by lower emissions and waiting times. However, it is also clear from the studies presented in this contribution that the coordination of the control strategy with the GLOSA vehicles is of enormous importance.