Passenger Train Delay Prediction using Machine Learning

Passenger train delay (PTD) is a challenging problem in rail transport. PTD means more uncertainty and unreliability for rail operators and higher dissatisfaction for passengers. Accurately predicting PTD can help rail operators adjust their operation and reduce passenger dissatisfaction. Leveraging advances in Machine Learning, this dissertation focuses on modelling Passenger Train Delay Prediction (PTDP) for Amtrak, the primary rail operator in the United States. Specifically, this dissertation offers three main contributions: (1) a Python-based Amtrak and Weather Underground (PAWU) data retrieving tool, 2) a station-to-station PTDP model, and 3) a universal PTDP model. In the first contribution, the PAWU tool is developed to retrieve historical train data from Amtrak and weather information from Weather Underground (used later to model PTDP). In the second contribution, Random Forest (RF), Gradient Boosting Machine (GBM), and Multi-Layer Perceptron (MLP) are utilized to build station-to-station PTDP models for individual station-to-station trip combinations. Two different data-frame structures are proposed, one using only real-time data and another using both real-time and historical data. The results show that using MLP with real-time and historical data works best. In the third contribution, a universal PTDP model is developed that can be used to predict PTD between any pair of stations by using RF and MLP. Unlike station-to-station model, here, the results show that RF performs best. In addition, the influence on PTD of several exogenous variables such as ridership, population, infrastructure and geography, historical delay profile at destination (HDPD), calendar, and weather information are also examined and discussed.