Abstract:

The highly dynamic nature of subway airport line passenger flows and their susceptibility to the influence of airport flight schedules present challenges for accurate short-term forecasting of passenger flow. This study integrates airport flight information and historical passenger flow data from airport lines to construct a short-term passenger flow forecasting model based on a stacking ensemble model. The model incorporates random forest (RF), LightGBM (light gradient boosting machine), gradient boosting decision tree (GBDT), and logistic regression algorithms to act as ensemble learners. The proposed model is validated using data from the Beijing Subway Daxing Airport Line and is compared against two baseline models, namely informer and long short-term memory (LSTM) networks. The results indicate that the dual-channel prediction, which considers flight information and historical passenger flows, outperforms the single-channel prediction solely based on historical passenger flows. The results also indicate that the stacking model demonstrates superior performance across all metrics. Particularly, the best prediction performance is achieved at a 96 step (24 h) forecast horizon, with mean absolute error of 7.66 and 4.67 for inbound and outbound passenger flow predictions, respectively. Analysis of the impact of flight information characteristics on the prediction model reveals that departure flight information is of relatively lower importance than that of arrival flights, which is attributed to large differences in advance arrival times for departing passengers.

Key words: airport line, short-term passenger flow forecasting, Stacking model, flight information