Case-based traffic analysis and prediction model for transportation incident management

Abstract

The intricate nature of modern urban traffic systems necessitates innovative methodologies for efficient transportation incident management. Addressing these disturbances is vital for both maintaining traffic flow and ensuring road safety, especially in burgeoning urban areas with increasingly complex transportation networks. This research presents a methodological approach that utilizes a data-driven paradigm to enhance our understanding and prediction capabilities in this domain.

Central to this study is the development of the Transport Incident Information Database. By adopting ontology-based modeling in tandem with advanced Natural Language Processing (NLP) techniques, this research efficiently processes a vast array of data from sources like traffic news texts. By integrating rule-based text mining with the BERT-based model, the research achieves a precise and comprehensive data synthesis, resulting in a continuously updated and relevant repository.

The introduced traffic performance evaluation model is another focal point of this research. It leverages datasets from the Hong Kong Transport Department, applying time series similarity techniques to identify and comprehend traffic patterns in non-incident scenarios. Such patterns offer a foundational baseline, upon which the LSTM time series model generates predictions related to traffic performance during incidents.

A significant innovation presented is the AISTGCN model, which represents a notable evolution in the realm of traffic prediction. This model incorporates the foundational principles of the ASTGCN framework but is enhanced with incident-based attention mechanisms. It adeptly captures the relationships between time, space, and incident attributes, delivering more precise traffic speed predictions and allowing for an in-depth analysis of different road network segments' resilience.

In terms of contributions, the research is manifold. On a theoretical level, it underscores the essential role of ontologies in traffic data analytics, with NLP serving as a crucial instrument for data interpretation. From a methodological perspective, the research introduces a unique convergence of rule-based techniques, the BERT model, LSTM, and the augmented ASTGCN with incident attention features. These methodological strides aim to push the boundaries of current transport analytics methodologies. On the practical side, the presented tools and models cater to stakeholders, ensuring a data-informed decision-making process that can effectively respond to real-world transportation challenges.

In conclusion, this study offers a comprehensive examination of traffic analysis and prediction, presenting strategies and models to elevate transportation incident management. The research emphasizes the integration of advanced analytical techniques with pragmatic solutions, aiming for a transport system that is resilient, adaptive, and informed by a wealth of data insights.