Spatio-Temporal Forecasting of Bus Arrival Times Using Context-Aware Deep Learning Models in Urban Transit Systems

Abstract

Accurate forecasting of bus arrival times is critical for enhancing the reliability and efficiency of public transportation systems. However, complex factors such as traffic congestion, weather conditions, and temporal variability make this task challenging. In this study, we propose a context-aware hybrid deep learning framework that integrates Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer architectures to predict stop-level bus travel times. The model is trained on a comprehensive dataset collected from 500 bus routes in Istanbul, spanning six months and incorporating real-time GPS data, GTFS schedules, and hourly weather attributes. A hybrid trend component is selectively introduced for data groups with fewer than 1000 samples to mitigate overfitting under sparse data conditions. Experimental results show that the trend-augmented LSTM model outperforms baseline architectures, achieving up to 28% improvement in MAE. The best-performing model yields an MAE of 2.97 minutes, a MAPE of 14.79%, and an $R >{2}$ value of 0.9272 across all test routes. Furthermore, condition-based evaluations demonstrate that prediction accuracy varies significantly across different time blocks, weather conditions, and day types. The proposed approach is both scalable and adaptable, offering a robust solution for real-time transit forecasting in complex urban environments.