Disruptions in public transport have significant consequences for service users and providers. These include longer travel times, increased crowding, and a negative perception of the service among travelers. Therefore, it is essential to maintain the regularity of public transport services to ensure high-quality service. Regularity is measured by how closely the actual service adheres to the planned schedule. However, disruptions and minor disturbances constantly pose a threat to this regularity. To improve the reliability of public transport systems, operators need detailed information about disruptions, including their locations, timing, duration, and their network-wide domino effects. This information helps in developing effective recovery plans and providing real-time updates for passengers. Thus, quantifying and minimizing the impacts of disruptions is crucial for both users and service providers, as it directly affects passenger satisfaction and service provider revenues. To this end, the current research proposes an analytical framework for the automated detection of disruptions in metro services (i.e., primary delays) and prediction of their propagation effects (i.e., the trains and stations that will be affected and the extent of the impact) on the entire metro network (i.e., secondary delays). 

The proposed framework involves overlaying real-time vehicle locations (using AVL data) onto the scheduled train paths (obtained by transforming the static GTFS data into GPS-like data and incorporating day-time-trip-station-dependent dwell times computed using AVL data), thereby identifying deviations from the scheduled train paths. Additionally, a graph modeling framework is proposed to represent the interdependencies between trains and stations in the metro network. In this graph model, running time, dwell time, and headways are depicted as edges, while stations are represented as nodes. A backward pass algorithm is developed to identify the main causes of the observed station delays. After identifying the time-stamped positions of the primary delays, a forward pass algorithm is developed to establish an association between the detected primary delays and their network-wide consequences, resulting in a primary-secondary dataset. The acquired primary-secondary dataset is enriched with additional features (e.g., timetable slacks) and fed into a deep learning architecture to predict network-wide secondary delays following the identification of any service disruptions within the metro system.