Predictive Real-Time Traffic Management in Large-Scale Transportation Network

Operating transportation highway networks in real-time presents a formidable challenge. Planned and unplanned incidents (e.g. hazardous weather conditions, accidents, local events, etc.) on the highway networks can catastrophically impact mobility and safety. It remains unknown to traffic operators which time and which strategy to engage for mitigating non-recurrent impacts, and additionally, how to incorporate overwhelmingly increasing traffic data. 

Mitigating non-recurrent impacts requires: accurate and ahead-of-curve real-time prediction, and proactive operational management. Unfortunately, both are not fundamentally addressed despite decades of research. Transportation Systems Management and Operations (TSMO) refers to a set of strategies that could be utilized to mitigate nonrecurrent impacts, such as pricing, signalization, and access control. Although TSMO are technically available to practitioners, but what time and what strategy to engage remain unknown. Being proactive is the key to effective management of non-recurrent impacts – vehicles need to be nudged before the substantial impacts reveal. For this purpose, accurate and prompt real-time prediction is paramount. 

In this dissertation, we focus on the prediction and management of traffic in large-scale networks in a predictive manner. We develop models and algorithms that leverage traffic sensing and machine learning to achieve two main goals: predicting non-recurrent traffic conditions in large-scale networks at least 30 minutes in advance and providing real-time recommendations for operational management. Ultimately, our goal is to devise an automated and holistic framework for predictive real-time traffic management, consisting of two tightly integrated modules: a prediction module that can swiftly adapt to real-time incidents, and an operational strategy module that offers traffic operators a significant window of time to assess conditions and respond appropriately through recommended strategies. 

Specifically, our proposed framework interconnects various sub-task models that (1) integrate data from multiple sources to capture mobility patterns, (2) learn the underlying processes from diverse mobility data in an offline setting, (3) adapt the model online to address new and previously unseen non-recurrent events, and (4) recommend optimal proactive traffic management strategies to minimize total congestion. Firstly, we develop a traffic sensing data pipeline for automating the transformation of inexpensive, replicable, and openly-accessible traffic data related to highways into coherent and synchronized traffic features of significant predictive power. We then propose an principled learning framework for accurately estimating both the recurrent and non-recurrent traffic components in an offline setting, as well as adapting the non-recurrent model online to traffic incident influences. The framework enables fast online inference and low-cost adaptation to non-recurrent traffic incidents in an explainable, modular way. Finally, we develop a predictive recommendation model, which associates the traffic predictions with incident signal timing plans, for automating operational management under realtime traffic incidents.