Deep Learning for Metro Delay Propagation Prediction : Combining Neural Network Architectures into Hybrid Models

Sammanfattning

Public transportation, especially metro systems, is essential for urban mobility, with millions worldwide depending on timely services for their daily commutes. However, various factors can cause delays that disrupt the entire network. If not managed effectively, these primary delays can propagate through the system, leading to secondary delays and a cascading effect that impacts city operations and economic productivity. This thesis investigates the use of neural network models to predict delays propagation (both single-step prediction and multi-step prediction) in metro systems, focusing on the Stockholm metro. By utilizing General Transit Feed Specification (GTFS) static data and real-time vehicle position data accessed through Trafiklab’s API, the research explores how these models can forecast arrival delays of different metro trains at different stations. The study combines several neural network architectures— CNN, LSTM, and FCNN—into hybrid models, such as CNN+LSTM, CNN+FCNN, 2dLSTM+FCNN, 2dLSTM+FCNN and 2dLSTMt+LSTM, to identify the most effective approach for delay prediction. The CNN, 2dLSTM, and 2dLSTMt components identify patterns in operational data such as planned and actual running, interval (headway) and dwell times but also departure and arrival delays. At the same time the FCNN and LSTM components focus on non-operational features such as track length and number of shared tracks. The outputs are then merged, with an additional FCNN component analyzing relationships between the two types of data. Further, the research details the data processing, formatting, and model construction methods. Since planned arrival times were unavailable, they were estimated from vehicle position data, crucial for calculating key features in the study. The findings reveal that while no single model outperformed across all tasks, the iv CNN+LSTM model generally excelled, especially in complex scenarios involving multiple stations and trains. The best-performing model achieved an MSE of 7.065, an RMSE of 22.66, and an R2 of 0.8817. The results also showed that model performance is more dependent on the number of trains and stations considered rather than the prediction horizon. Additionally, the project analyzed feature contributions using SHapley Additive exPlanations (SHAP) plots, finding that the actual departure delay at the current station was the most significant predictor of delays. These results contribute to developing automated metro dispatching systems that reduce human error and enhance operational efficiency, leading to a more reliable and satisfying service for passengers.