Applying a multistage of input feature combination to random forest for improving MRT passenger flow prediction

As one of the main public transport systems all over the world, mass rapid transit (MRT) is widely served in the metropolitan areas. To meet the increasing travel demands in the future, accurately predicting MRT passenger flow is becoming more and more urgent and crucial. This paper aims to use an experimental way to objectively quantify and analyze the impacts of various combinations of traditional input features to improve the accuracy of MRT passenger flow prediction. We have built a series of passenger flow prediction models with different input features using a random forest approach. The features of passenger flow direction, temporal date, national holiday, lunar calendar date, previous average hourly passenger flow, and previous k-step hourly passenger flow and their trends are selected and applied in a multi-stage of the input feature combination. The typical encoding strategies of the input features have been further discussed and implemented. Finally, the optimal combination of the input features has been proposed with a case study at Taipei Main Station. The experimental results show that the proposed optimal combination of the input features and their appropriate codes can be helpful to improve the accuracy of passenger flow prediction, not only for the prediction results on weekdays and weekends, but also for them on national holidays.