Near Real-Time Train Location Prediction Using LSTM Neural Network With Attention Mechanism

Accurate and continuous train location tracking is essential for urban rail systems to ensure safety and efficient scheduling. However, traditional sensors often produce irregular updates due to delays and signal interference, creating data gaps that disrupt operations. This study presents a near real-time prediction model. First, data interpolation is used to create consistent 60-second interval inputs, filling sensor gaps. Second, stacked Long Short-Term Memory (LSTM) layers capture long-term dependencies in the data. Third, an attention mechanism enables the model to focus on relevant time steps, enhancing prediction accuracy. Using more than six months of data from the Commuter Line Jawa Barat dataset, our model achieved an average Mean Absolute Error (MAE) of 0.0013 and Root Mean Square Error (RMSE) of 0.0016, outperforming other state-of-art deep learning models, including GRU, RNN, and CNN-LSTM, and showing a 38% improvement in MAE and 41% improvement in RMSE compared to the traditional LSTM model.