Deep Learning-Based Approach for Public Transportation Estimated Time Arrival Prediction

Efficient public transportation is essential for sustainable urban mobility and achieving the 20-minute city vision. This paper presents a deep learning-based framework for accurately predicting the Estimated Time of Arrival (ETA) of buses using real-world GPS data from Ulaanbaatar’s metropolitan transit system. The study evaluates multiple models—including Fully Connected Multi-Layer Perceptron (FCMLP), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Random Forest, and ensemble methods—on both single-route and city-wide (139-route) datasets. Results show that deep learning models trained on individual routes achieve high accuracy (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ≈ 0.93, MAE ≈ 18.5 seconds), effectively capturing localized spatiotemporal patterns. On the city-wide dataset, the FCMLP model achieves the best performance (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.98, MAE = 20.54 seconds). In contrast, the GPS-sequence-based LSTM model yields lower performance (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.66, MAE = 100.43 minutes). These findings highlight the effectiveness of route-specific modeling and structured feature engineering. The proposed approach offers practical insights for improving transit reliability, adaptive scheduling, and real-time passenger information, supporting data-driven planning for resilient and accessible urban transportation systems.