Deep learning-based traffic flow prediction and traffic management system for urban transportation networks

Due to the growth of the population and improved living standards, a lot of urbanization policies are implemented by the governments in many countries.As a result, the total amount of vehicles on the streets is growing rapidly, and traffic congestion occurs frequently.The recurring and non-recurring congestion has resulted in high economic costs, environmental pollution and severe noise pollution.Hence, how to alleviate traffic congestion and delays have become an essential problem in this urbanized world.In this thesis, we proposed a deep learning-based traffic prediction and management system to tackle the traffic congestion problem.First, we developed a domain-adversarial-based graph network for short-term traffic flow prediction.Contrast experiments have been conducted on several opensource traffic datasets, which proved that our proposed method had made some improvements compared to other baselines.Secondly, we proposed an optimized vehicle route guidance and navigation system.The employed deep reinforcement learning method and graph convolutional network can extract several features from traffic data and topological data.The proposed route guidance system can assign the best route and re-route the vehicles in a complex urban transportation network with dynamic traffic conditions.Last but not least, we proposed a deep reinforcement learning-based traffic signal control strategy to reduce the travel delay time and queue length by adjusting phase splits.TD3 algorithm is employed to generate optimal phase splits for the next cycle of traffic signals.The performance of traffic signal control by this strategy is evaluated through several urban road networks with true traffic demand or manually generated demand via VISSIM.