Forecasting boarding passenger volume at PT KAI Daop 5 Purwokerto using SARIMA and LSTM

Abstract

Railway transportation plays a vital role in supporting public mobility, requiring accurate passenger demand forecasting to assist operational planning, particularly during seasonal peak periods. This study aims to apply and compare the Seasonal Autoregressive Integrated Moving Average (SARIMA) and Long Short-Term Memory (LSTM) models for daily passenger boarding volume forecasting at PT Kereta Api Indonesia (KAI) Daop 5 Purwokerto. The dataset consists of daily passenger boarding records from January 1, 2023, to July 15, 2025, treated as a time series. For the SARIMA model, data preprocessing included log transformation, differencing, and seasonal differencing to achieve stationarity. Meanwhile, the LSTM model employed data normalization and sequence construction to capture nonlinear temporal dependencies. The experimental results show that the best SARIMA configuration is SARIMA(3,1,3)(0,1,1)7, while the optimal LSTM architecture uses a window size of 30 with two hidden layers of 64 and 32 units trained for 50 epochs. Performance evaluation on the test data indicates that the LSTM model outperforms SARIMA, achieving lower error values (RMSE = 2632.94; MAPE = 13.2%) compared to SARIMA (RMSE = 3399.48; MAPE = 17.9%). These findings indicate that, in this case study, the LSTM model achieved better forecasting performance than the SARIMA model on the selected hold-out test dataset (882 training observations and 45 testing observations). The obtained MAPE of 13.2% indicates good forecasting accuracy according to the adopted accuracy scale, suggesting that LSTM is a promising approach for supporting passenger demand forecasting at PT KAI Daop 5 Purwokerto.