Multi-component hybrid deep learning model for railway passenger demand forecasting
TL;DR
This study presents a systematic evaluation of decomposition-based forecasting frameworks for railway passenger demand prediction by integrating Seasonal-Trend Decomposition using Loess, EMD applied to residual components, and Fuzzy C-Means clustering to demonstrate the framework's robustness and generalization capability.
Abstract
Railway transportation plays a critical role in supporting sustainable mobility in Indonesia, yet significant fluctuations in passenger demand often lead to congestion and operational challenges. This study presents a systematic evaluation of decomposition-based forecasting frameworks for railway passenger demand prediction by integrating Seasonal-Trend Decomposition using Loess, EMD applied to residual components, and Fuzzy C-Means clustering. Using the Argo Muria train service as a case study, multiple deep learning models, including LSTM, GRU, RNN, CNN, and BiLSTM, are trained on decomposed components, and their forecasts are combined linearly. Model performance is evaluated using a rolling-origin strategy across multiple stations. At the primary destination station, Semarang-Gambir, the best configuration achieves an MAE of 19.88, RMSE of 26.79, sMAPE of 8.97, and R2 of 0.84. Consistent results across stations demonstrate the framework's robustness and generalization capability.