Passenger Flow Prediction with Transformer : The Shenzhen Metro Case

Traffic flow forecasting has long been a research topic, aiming to predict future flow and provide insights for current traffic control. It is a complex time series problem attempting to capture dynamics and multiple patterns of spatial and temporal dependencies. In this paper, we construct Transformer and other 10 baseline models to make predictions of passenger flow based on the Shenzhen Metro case. More specially, we first perform data pre-processing and feature engineering approach to obtain temporal features and spatial features. Next, we construct 10 baseline models, i.e. four regression models (Lasso, Ridge, ElasticNet and SVR), two Boosting models for bias reduction (XGBoost and LightGBM), two Ensemble methods for variance reduction (Stacking and Blending), and lastly, two Deep -sequence models for generation (RNN and LSTM), for traffic flow prediction. Experimental results show that regression models perform worst and deep sequence models outperform other models dramatically. Most importantly, Transformer achieved a better performance than deep sequence models with a 36.919% reduction in RMSE compared with SVR, the worst performing model.