With the rapid development of highspeed railway in China, highspeed railway transport hub (HRTH) has become the highdensity distribution center of passenger flow. In order to accurately detect potential safety hazard hidden in passenger flow, it is necessary to forecast the status of passenger flow. In this paper, we proposed a hybrid temporalspatio forecasting approach to obtain the passenger flow status in HRTH. The approach combined temporal forecasting based on radial basis function neural network (RBF NN) and spatio forecasting based on spatial correlation degree. Computational experiments on actual passenger flow status from a specific bottleneck position and its correlation points in HRTH showed that the proposed approach is effective to forecast the passenger flow status with high precision.
As main influence factors for the safety and sustainability of transportation system, the insecure behaviors and statuses of people are hot issues and difficult problems in traffic safety engineering [
In order to avoid and solve the problems caused by passenger flow abnormal status, many approaches are proposed in literatures which can be mainly classified into two categories. The first category is the studies on passenger flow modeling and simulating in transport hubs. Gipps and Marksjö [
The other category focuses on the modeling and simulating of congestion evacuations which are caused by highdensity passenger flow. Zhong et al. [
According to the literature review above, most studies focus on solving approaches before or after passenger flow abnormal status happened. Specific literature on realtime changing process of passenger flow status is scarce because of limitation from the difficult acquisition of realtime passenger flow status. With the widespread applications of intelligent video surveillance in Chinese HRTH, realtime acquisition of passenger flow status has become feasible. In this paper, passenger flow status is defined as the amount, velocity, and density of passenger flow. We consider the realtime passenger flow status in bottleneck positions of HRTH and propose a hybrid temporalspatio forecasting approach to reflect the change of passenger flow status.
The rest of paper is organized as follows: a hybrid temporalspatio forecasting approach for passenger flow status of bottleneck positions is developed in Section
A hybrid forecasting approach for passenger flow status of bottleneck positions is proposed in this section, which combines temporal and spatio forecasting methods. A temporal forecasting based on RBF NN is proposed to forecast passenger flow status of bottleneck position by using the realtime passenger flow status of the position. The temporal forecasting method can rapidly and precisely reflect the passenger flow status in the bottleneck position but is insensitive for the passenger flow fluctuation from correlation points. So we introduce a spatio forecasting approach based on spatial correlation degree to combine with the temporal forecasting approach for improving the forecasting precision.
The RBF NN is a typical feedforward neural network, which has many merits, such as nonlinear mapping characteristics, selforganized study ability, training fast, and the capability of converging in global optimization and approaching the function in the best way. Simply for its great advantages, RBF NN has been applied in many fields [
The structure of RBF NN is comprised of three different layers: an input layer, a hidden layer and an output layer. The structure of RBF NN for temporal forecasting is shown in Figure
Structure of RBF NN.
The input vector
The output
It is very important to select proper input variables for a neural network. On the one hand, as the number of the input variables increases, the NN architecture will be larger and the computing time will be longer. On the other hand, the irrelevant or mutually correlated input variables are not useful for improving the prediction accuracy. Therefore, how to select a few but sufficient input variables is a key issue.
According to plenty of experiments on the input variables, we choose passenger flow status (amount, velocity, or density) of 15 points in the time series as the variables input our RBF NN and passenger flow status (amount, velocity, or density) of the point after the 15th point as the output variable.
Mapping structure between input variables and output variable.
Input variables  Output variable 









In order to test the accuracy of temporal forecasting, we choose 250 actual passenger flow amounts of one bottleneck position in time series as testing samples. The input and output values after normalization are shown in Table
Input and output values.
No.  Input value  Output value 

1 

−0.7561 
 
2 

−0.8049 
 
3 

−0.8537 
 



 
234 

0.6585 
 
235 

0.5610 
A comparison between actual value and forecasting value based on RBF NN is shown in Figure
Comparison between actual value and forecasting value based on RBF NN.
According to the spatial correlation degree and passenger flow status of bottleneck position
According to the passenger flow status of
In order to reduce the errors which are caused by uncertainties passenger flow status, we use the change of initial forecasting value
The velocity in correlated points prominently affects the density of congested point; the faster the velocity of correlated points, the weaker affection on the density of congested point. So we adopt velocity to improve the density forecasting. The spatio forecasting model of passenger flow density is shown in (
Similarly, the density in correlated points prominently affects the velocity of congested point; the higher the density of correlated points, the faster the velocity of congested point. So, we adopt density to improve the velocity forecasting. The spatio forecasting model of passenger flow velocity is shown in (
A hybrid forecasting approach to combine the spatio and temporal forecasting is proposed in this section. The forecasting value
The parameters of the model are determined by numerical fitting of actual passenger flow status values and forecasting values generated by spatio and temporal forecasting methods.
According to the spatio and temporal forecasting methods and the combination model, the temporalspatio forecasting model of passenger flow density is described by
According to the spatio and temporal forecasting methods and the combination model, the temporalspatio forecasting model of passenger flow velocity is described by
To illustrate the proposed forecasting approach, computational experiments are performed by using the actual passenger flow status from a specific bottleneck position in the Chinese HRTH. In order to assess the improvement of our approach, the forecasting of passenger flow density and velocity among temporal forecasting based on RBF NN, spatio forecasting based on spatial correlation degree, and hybrid temporalspatio forecasting approach are compared. These numerical experiments are performed based on a personal computer with Intel Core(TM) i52450M @ 2.50 GHz processors and 4 GB RAM.
To implement the proposed forecasting approach, the parameters related to specific bottleneck position of HRTH are needed. The specific bottleneck position is a ticket entrance
According to RBF NN designed in Section
Comparison between actual value and forecasting value of passenger flow density based on RBF NN.
Comparion between actual value and forecasting value of passenger flow velocity based on RBF NN.
According to spatio forecasting model proposed in Section
The forecasting value of passenger flow density based on spatio forecasting model is calculated by (
The comparison between actual value and forecasting value of passenger flow density based on spatio forecasting model is shown in Figure
Comparion between actual value and forecasting value of passenger flow density based on spatio forecasting model.
Comparion between actual value and forecasting value of passenger flow velocity based on spatio forecasting model.
Based on the numerical fitting of actual and forecasting passenger flow status value in
The comparison between actual value and forecasting value of passenger flow density based on hybrid temporalspatio forecasting approach is shown in Figure
Comparion between actual value and forecasting value of passenger flow density based on hybrid temporalspatio forecasting approach.
Comparion between actual value and forecasting value of passenger flow velocity based on hybrid temporalspatio forecasting approach.
The average forecasting precision comparison of three approaches mentioned above is shown in Table
Average forecasting precision comparison of three approaches.
Forecasting approach  Average forecasting precision of density  Average CPU time of density forecasting (s)  Average forecasting precision of velocity  Average CPU time of velocity forecasting (s) 

Temporal forecasting approach based on RBF NN  95.21%  9.8  94.73%  9.6 
Spatio forecasting approach based on spatial correlation degree  91.23%  11.2  88.21%  12.1 
Hybrid temporalspatio forecasting approach  96.83%  12.9  96.10%  13.7 
As observed in Table
In this paper, we considered the forecasting approach for passenger flow status in the Chinese HRTH. A hybrid temporalspatio forecasting approach was proposed, which combined temporal forecasting and spatio forecasting. The temporal forecasting based on RBF NN could fast and accurately forecast the status change of passenger flow but was insensitive for the influences from correlation points. A spatio forecasting approach based on spatial correlation degree was introduced to combine with the temporal forecasting approach to avoid the influences and improve the forecasting precision. Computational experiments on the actual passenger flow density and velocity from a specific bottleneck position and its correlation points in Chinese HRTH showed that the approach proposed in this paper is effective to forecast the passenger flow status of bottleneck position in HRTH with high forecasting precision for different types of passenger flow status. In the future, considering the passenger flow abnormal status forewarning of bottleneck position based on the passenger flow status forecasting is a possibility for further research.
The authors declare that there is no conflict of interests regarding the publication of this paper.
This research was supported by a Grant (no. I11A300010) from the National Natural Science Foundation of China.