Train Arrival Delay Prediction by Capturing Spatial and Temporal Features of Train Data Sensor

Predicting train delays accurately is essential for improving operational reliability and passenger satisfaction in urban commuter networks. This study applies spatiotemporal data mining to limited sensor data from Bandung’s commuter railway system to develop a framework capable of reconstructing arrival and departure events despite minimal information. By leveraging the spatial and temporal dimensions of sensor data, this approach enables better delay prediction accuracy comparable to models using more comprehensive datasets. Among the four machine learning models tested, the Long Short-Term Memory (LSTM) model consistently demonstrated the highest accuracy. With an average prediction error of approximately 0.13 minutes, or about 7.8 seconds from the actual delay times, the LSTM model outperformed the Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), and CNN-LSTM models. The study highlights that limited sensor data can be maximized for effective delay prediction, offering a practical approach to enhance scheduling and reliability in data-constrained transit networks.