Attention-Enhanced LSTM for Short-term Passenger Flow Prediction in Urban Rail Transit

Abstract

Short-term passenger flow forecast is vital for enhancing urban rail transit operations. However, standard LSTM models process all historical time steps equally and fail to reflect the significant impact of peak hours. To remedy this shortcoming, this paper introduces an Attention-Enhanced LSTM model with a streamlined additive attention mechanism that automatically gives higher weights to recent time steps. The model is assessed on the New York MTA hourly ridership data set from 2022 to 2024 at Times Square Forty-Second Street Station. The experimental results show that the Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) of the model are about 1% -2% higher than those of the standard LSTM model. Moreover, the attention weights can be displayed, revealing that the model focuses on the last two hours, which is in line with short-term traffic continuity. Robustness tests over ten independent runs with different random seeds confirm steady performance with small standard deviations. The model is lightweight in structure, consisting of two LSTM layers and one attention layer, making it practical for real-time implementation. This study establishes that simplified attention mechanisms can provide both explainability and stability for short-term passenger flow forecasting. Future work will expand to multi-station prediction using graph neural networks and take into external factors including weather conditions and holidays.