Congestion estimation and turning ratio prediction based on machine learning with applications in urban traffic light control

Abstract

Increasing transportation efficiency is an interesting and important problem. In the world with convenient means of ICTs, the concept of “smart city” emerged. In the meantime, a lot of data-driven traffic network optimization algorithms have also been developed and applied widely. However, the performance of some optimization algorithms can be improved with some pre-works added. This thesis discusses two such pre-works. The first pre-work is urban traffic network congestion region identification and prediction with two case studies at NTU campus and Jurong area, which utilizes the vehicle data (average speed, GPS-based location, heading direction) via V2X to analyse the traffic condition of each link. Links with similar congestion levels will be clustered together into a region. Our simulation-based case studies show that about 75% of the total queue delay could be reduced with good knowledge of congestion regions in the network. The second pre-work is about traffic network turning ratio prediction, which may be useful in developing more accurate network dynamic models. By constructing a recurrent neural network to predict the vehicle turning ratios at the next time step with prior or online-learned knowledge of network supply functions, traffic light schedules and historical vehicle turning ratios as inputs. This prediction model can be integrated with a real-time traffic signal control algorithm to form an adaptive closed-loop traffic signal control strategy, which in our simulated case studies decreases 24% of the delay time compared to the case without turning ratio prediction.