Novel Three-Stage Framework for Prioritizing and Selecting Feature Variables for Short-Term Metro Passenger Flow Prediction

Abstract

<jats:p> Short-term metro passenger flow prediction is vital for the operation and management of metro systems. Most studies focus on the higher prediction accuracy with statistical and machine learning methods, but little attention has been paid to the prioritization and selection of feature variables, especially for different metro station types. This study aims to analyze the effect of feature variables on the prediction results, and then select appropriate predictor variables accordingly. A novel three-stage framework is proposed to prioritize feature variables for short-term metro passenger flow prediction, including station clustering, feature extraction, and variable prioritization. A hierarchical clustering algorithm (AHC) is developed for station clustering, the results of which are verified by the K-means and Davies-Bouldin (DB) statistical index. We then extract the temporal, spatial, and external features. Finally, the association between the variables and the prediction results is explored using tree-based models. The proposed framework is demonstrated and validated with data collected from Shanghai Metro Automatic Fare Collection (AFC) system. The results highlight that the importance of feature variables for developing models varies between stations, whereas only a few variables are found to explain most of the variation in the testing dataset; different feature variables lead to distinct differences in prediction accuracy, and simply adding more predictor variables does not necessarily lead to higher prediction accuracy. In addition, the station type and prediction type (i.e., tap-in and tap-out) have little influence on the selection of feature variables. </jats:p>

Passenger Flow Prediction Using Smart Card Data from Connected Bus System Based on Interpretable XGBoost

“…Zhang et al combined with Residual Network (ResNet), Graphic Convolutional Network (GCN), and Long and Short-Term Memory (LSTM) put forward a deep learning architecture [14]. Zhao et al proposed a new three-stage framework based on a hierarchical clustering algorithm (AHC) and tree-based models to select the appropriate feature variables [15]. They proposed a hybrid spatial and temporal deep learning neural network (HSTDL-NET) [16].…”

Naïve Bayes-Based Transition Model for Short-Term Metro Passenger Flow Prediction under Planned Events

“…Table 1 summarizes the relevant literature for passenger flow prediction in metro systems. Most models are proposed under normal scenarios, such as empirical mode decomposition (EMD) with backpropagation neural network (BPNN) ( 13 ), autoregressive integrated moving average model ARIMA ( 14 ), linear regression (LR) ( 15 ), Bayesian model ( 16 ), ARIMA and generalized autoregressive conditional heteroskedasticity ( 9 ), random forest (RF) ( 17 ), gradient boosting decision trees (GBDT) ( 18 ), state-space models ( 19 ), a gravity model with deep learning (DL) ( 20 ), a model based on BiLSTM-CNN (bidirectional long short-term memory neural networks and convolutional neural networks) ( 21 ), a Seq2Seq (sequence-to-sequence) model with attention mechanism ( 22 ), and DL-based models ( 23 – 26 ).…”

“…The naïve Bayes model (i.e., the transition mechanism) is developed to infer the prediction scenario and selects the sub-predictor from GBDT or DL accordingly. The GBDT is used for prediction under normal conditions given its capabilities in handling dense numerical features ( 18 , 36 , 37 ). A DL-based model ( 23 ) is adjusted for prediction under planned events.…”

Short-Term Metro Ridership Prediction During Unplanned Events

“…Many approaches have been proposed for prediction under typical situations, including ARIMA (16,17), fuzzy logic (18), Kalman filtering (19), support vector machines (5), back-propagation neural networks (20), tree-based models-for example, gradient-boosting decision trees (GBDTs) (21)(22)(23) and random forest (RF) (23,24)-and deep learning models (2)(3)(4). They have different capabilities in capturing complex and nonlinear relationships between inputs and outputs.…”