The Prediction and Mitigation of Road Traffic Congestion Based on Machine Learning

Bartlett, Zoe Elaine (2023) The Prediction and Mitigation of Road Traffic Congestion Based on Machine Learning. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (7MB) | Preview

Abstract

Traffic congestion is a major issue for all developed countries. In most urbanised areas space is a scarce commodity. Therefore, better management of the existing roads to increase or maintain their capacity level is the only viable solution. Research in the last two decades has focused on Intelligent Transport Systems (ITS) development. Predicting traffic flow in real time can be used to prevent or alleviate future congestion. The key to an effective proactive method is a model that produces timely and accurate predictions. However, despite extensive research in this area, a reliable method is still not available. Therefore, in this thesis, we developed an accurate online road traffic flow prediction model, with a particular focus on heterogeneous traffic flow, for urbanised road networks. The contributions of this work include: Firstly, we conducted a comprehensive literature review and benchmark evaluation of existing machine learning models using a real dataset obtained from Transport for Greater Manchester. We investigated their prediction accuracy, time horizon sensitivity, and input feature settings (different classes of vehicles), to understand how they can affect their prediction accuracy. The experimental results show that the artificial neural network was the most successful at predicting short-term road traffic flow. Additionally, it was found that different classes of vehicles can improve prediction accuracy. Secondly, we examined three recurrent neural networks (a standard recurrent, a long short-term memory, and a gated recurrent unit). We compared their accuracy, training time, and sensitivity to architectural change using a new performance metric we developed to standardise the accuracy and training time into a comparable score (STATS). The experimental results show that the gated recurrent unit performed the best and was most stable against architectural changes. Conversely, the long short-term memory was the least stable model. Thirdly, we investigated different magnitudes of temporal patterns in the dataset, both short and long-term, to understand how contextual temporal data can improve prediction accuracy. We also developed a novel online dynamic temporal context neural network framework. The framework dynamically determines how useful a temporal data segment is for prediction, and weights it accordingly for use in the regression model. The experimental results show that short and long-term temporal patterns improved prediction accuracy. In addition, the proposed online dynamical framework improved prediction results by 10.8% when compared with a deep gated recurrent model. Finally, we investigated the dynamic nature of road traffic flow’s input features by examining their spatial and temporal relationships. We also developed a novel dynamic exogenous feature filter mechanism. The feature filter mechanism uses ’local windows’ to filter input features in real-time to improve prediction accuracy. The results show that a global correlation was insufficient to describe the complex and dynamic relationships between the input features. The local correlations (local windows) were able to identify additional geospatial and temporal relationships. Furthermore, the proposed feature filter mechanism was compared to a state-of-the-art method, a dynamic rolling window feature filter model. The experimental results showed that the proposed model was the most accurate, with an RMSE of 10.06%, closely followed by the dynamic rolling window feature filter model, with an RMSE of 10.98%. However, the proposed model was computationally much lighter than the rolling windows model.