Real-Time delay prediction for single-track intercity railways

• Thailand’s intercity railway network is dominated by single-track operations, making it highly prone to delay propagation. • A random forest–based machine learning model was developed to predict real-time arrival delays using both scheduled and real-time data, including recent delay states, trip characteristics, and operational conditions. • The model achieves high accuracy (R² = 0.84, MAE = 2.53 minutes) and strong operational reliability (87.5% within 5 minutes and 94.9% within 10 minutes of the actual arrival delay). • SHAP analysis reveals that real-time delay propagation dominates predictions, with service type and infrastructure factors providing secondary contributions. • The framework enhances dispatching decisions, track utilisation, and passenger communication in Thailand’s single-track network. Thailand’s intercity railway system is dominated by single-track operations, where limited passing loops, priority rules and aging infrastructure make services highly vulnerable to delay propagation. Existing analytical and simulation-based methods struggle to capture complex, non-linear interactions between operational factors, infrastructure constraints, and real-time information. This study introduces a real-time delay prediction framework specifically designed for single-track intercity railways, integrating both scheduled features and evolving en-route conditions. A cross-validated random forest model, enhanced with explainable SHAP analysis, is developed to generate real-time predictions of arrival delays at destination stations. The model demonstrates strong predictive accuracy (R² = 0.84; MAE = 2.53 minutes), with 87.5% of forecasts fall within 5 minutes of actual arrival delay and 94.9% fall within 10 minutes, confirming its suitability for operational decision support. SHAP interpretation reveals that real-time delay propagation variables dominate prediction outcomes, while service characteristics and infrastructure factors contribute secondary but meaningful effects. The proposed framework provides practical, real-time value for dispatchers, enabling proactive routing decisions, congestion mitigation, and improved passenger communication. This work offers one of the first explainable machine-learning delay prediction models tailored to single-track railway operations and presents insights applicable to similarly constrained rail systems worldwide.