Using Delay Logs and Machine Learning to Support Passenger Railway Operations

Abstract

Passenger railways face reliability challenges resulting from shared track with other trains, limited infrastructure capacity, and rolling stock and trackway that is subject to major failures during service. Dispatchers may have limited contextual information when responding to an emerging delay, and often rely on their own experience to manage an incident. This study leverages various aspects of delay logs—a common set of data collected during railway operations—to arm dispatchers with an understanding of delays, provide contextual information about previous delays that are similar to an emerging event, and make predictions about the size of a delay based on emerging information. Using graph theory, short-text topic modeling, cosine similarity, and machine learning regression models, we demonstrate that agencies can leverage this single data source for insight and operational support. To showcase the potential insights gained by these methods, we apply them to delay log data from the GO Rail network in the Greater Golden Horseshoe area of Ontario, Canada. We find that elastic net and random forest regression models outperform naive models that may be tacitly used in practice today.