XGBoost in Public Transportation for Multi-Attribute Data: Delay Prediction in Railway Systems in Real-Time

Predicting delays in metro and tram services is a complex task that requires advanced approaches, such as powerful machine learning tools. This study addresses this topic by applying the XGBoost and Bayesian Optimization (BO) algorithms, which offer a prediction horizon of 15 minutes instead of the next-station prediction, which leaves a very short time for the operator to react if we want to use the prediction for the next station. Our research strongly emphasizes methodological validation, with daily evaluations against real-time data. This process is reinforced by collaboration with the Operational Control Center (OCC) to ensure robustness. The 15-minute delay strikes a balance, giving control center operators sufficient notice to orchestrate traffic management, mitigate disruption, and take timely action. With an exemplary real-world accuracy of 95%, the results of our model have been validated by the OCC. Future efforts will include the seamless integration of predictive capabilities into real-time display systems for the OCC, providing innovative information to optimize traffic flows and ensure punctuality in urban rail systems.