Spatio-temporal Machine Learning for Future Mobility: Precision, Resilience, and Adaptability

In a rapidly evolving world with growing population density, urban expansion, and increasing global and local interconnectivity, the necessity for efficient mobility has never been greater. Many regions heavily rely on private cars for transportation, but relying on private cars can not scale with the increase in demand, and it will lead to pollution and inefficient usage of resources. Sustainable and efficient public transport alternatives are needed, especially within and around cities where road networks have limited capacity. One of the main ways to improve public transport is to introduce better Intelligent Transport Systems (ITS). These systems can increase both system efficiency and user experience.
With more and more data sources such as Automatic Vehicle Location (AVL) systems, smartcard data, vast networks of sensors, and even smartphone data becoming increasingly prevalent, the precision and reliability of ITS systems are increasing. Furthermore, with autonomous shuttles entering the final stages of pilot tests, combining these new data sources and new ITS models brings a unprecedented possibility to revolutionize public transport. However, when autonomous public transport becomes the norm, the need for precise and reliable ITS systems also increases. Without human drivers, faulty ITS predictions might lead to misinformed travellers and potentially a complete breakdown of the transport system. Data-driven ITS systems have seen a lot of research in recent years, with machine learning methods outperforming statistical and transport models on benchmarks. Especially the use of spatially aware machine learning models has vastly improved the accuracy of, e.g. traffic flow models - so much that the mean absolute error on test sets is as low as two m/s. However, has the effort been misguided? The value of predicting traffic flow or ETA during free-flow traffic is low - people are just following the traffic laws. The time when ITS models are important is not during ordinary operations but during abnormal scenarios. When traffic is not as expected, travellers need updates to wait and arrival times. Furthermore, what about when we start using autonomous public transport? How much of the recent development in ITS systems also works on autonomous transport? With many of the current deployments of autonomous vehicles being limited to special closed road networks with specific regulations governing their use, it is not unlikely that lessons learned from ordinary buses in mixed traffic will not carry over.

In this thesis, I present work that focuses on leveraging new advances in Spatiotemporal machine learning to extend ITS systems to be more precise, more resilient to unforeseen changes using new data sources, and future-proof by analyzing how current models transfer to autonomous vehicles.

The first chapter presents a new spatiotemporal model for predicting traffic flow and Mobility-as-a-Service demand prediction. Building upon Neural Relational Inference, we create a model that, based on recent history, jointly predicts future traffic and infers the spatial connections that best model these predictions. This allows the model to adapt to new situations and traffic controllers to get a sense of what the model deems important, thus moving towards more interpretable models.

The second chapter presents a new model framework and benchmark datasets to analyze the impact proper usage of incident reports can have on congestion prediction. First, by merging two publicly available datasets together and by extending SUMO, a microscopic simulation framework, we create two new benchmark datasets for congestion prediction. Next, we present a framework for including incident reports in the prediction models for the Spatial Impact Region of the congestion following incidents.

The third chapter presents five new datasets from autonomous shuttle pilot sites under the SHOW project. The sites all have autonomous shuttles serving travellers around local areas but have vastly different fleet types, data amounts, and data quality. We apply different heuristical, statistical, and machine learning models to the sites and gather insights on which kinds of models work for different kinds of sites.

We conclude this thesis with a discussion on future directions ITS and autonomous public transport research should focus on taking this research for the best possible adaptation to autonomous public transport.