Integrated Spatio-Temporal Data Mining for Quantitative Assessment of Road Network Performance

Recent traffic surveys and analysis of road network performance in London show a decline in traffic flows and perversely a decline in speeds and increase in congestion. It is believed that the increases in congestion reflect travellers' responses, both temporary and longer-term, to competition for road network capacity. Continuing adjustments to network capacity in pursuit of mayoral transport priorities, for example, improved safety and amenity, and increased priority for buses, taxis, pedestrians and cyclists, has led to increasing delays for private vehicular traffic. The current annual cost of congestion on London's main roads is estimated to be in the range of 1.8 to 3 billion.Analysis of road network performance is intricate. This is because the road network is essentially an open system with many factors and in which travellers can respond by modifying their choices in many different ways that will affect monitored performance outcomes. The form of these factors, their direction of causality, the fact that some of them interact strongly, and their sheer numbers all contribute to the complexity. These factors have different patterns of influence in both time and space, and analysis of the distinct cause-effect patterns is complicated by the non-linearity of the effects, including the possibility of abrupt growth in congestion once it sets in. Modelling spatial-temporal dependency of the factors is the bottleneck in analysis of the network performance. The challenge is to model dependency in both space and time seamlessly and simultaneously so that the accuracy of analysis can be improved. Another challenge is to fully consider the topology (links and hierarchies) and geometry (distances and directions) of real road networks in the analysis. These are also fundamental challenges in modelling complexity of other types of networks.This research will tackle these challenges. It will be achieved by innovative combination of two chosen novel machine learning methods (Dynamic Recurrent Neural Networks - DRNN and Support Vector Machines - SVM) with the most advanced statistical space-time series analysis (Spatio-Temporal Auto-Regressive Integrated Moving Average - STARIMA) and Geographically Weighted Regression - GWR. These methods are selected because their applications in transport studies are relatively new compared with conventional statistical methods, and, more importantly, they have the potential to improve the representation of the network complexity. The DRNN and SVM can model the non-linearity and non-stationarity existing in most spatio-temporal data which may not be fully accommodated by STARIMA. The STARIMA has the explanatory capability which is missing in DRNN and SVM. The GWR can model the heterogeneity of the networks and improve the understanding of the scales of the networks. Their use in combination will improve the sensitivity and explanatory power of the analysis, to enable the effects of the factors to be assessed separately (isolatable). These methods will also be explored, refined and further developed in the light of experience in this study.The outcome of this research will advance the new and emerging fundamental researches in agent simulations, dynamic network analysis, and computational models and architectures of artificial neural networks, which are widely involved in space-time analysis of social-economic phenomena. It will offer TfL better tools and techniques to manage the road space and mitigate congestion more effectively thereby improving person journey times and overall journey reliability, and in doing so also deliver large economic benefits to London. The benefits of the research will accrue widely to both public and private transport users. The methodology developed here will be transferable to understand the congestion in other big cities around the world with economic, monetary, social and environmental benefits.