Analysis,Prediction And Early Warning Of Burst Passenger Flow In Urban Rail Transit System

With the improvement of urbanization and economic development,the transportation demand of people is growing,which has led to more and more transportation problems.These problems not only increase people's travel costs but also make more energy con-sumption and environmental pollution.Some serious security issues are also brought about.To solve these problems,the ITS(Intelligent Transportation System)came into being.The ITS is an intelligent information system that has been developed in recent years to manage and control the transportation system with modern intelligent technology.It plays an important role in alleviating traffic congestion,reducing traffic accidents and decreasing energy consumption.This paper focuses on a key issue in intelligent transportation systems,traffic flow prediction.We focus on sudden passenger flow in urban rail transit system.It makes the analysis,prediction and early warning of sudden passenger flow in the urban rail transit system an important but challenging task that the sudden passenger flow has its sudden and high-risk nature.Traditional traffic flow models usually exploit simulation or predict the future traffic flow with a fitted linear flow model based on historical traffic.However,with the popularity of sensor networks and the development of computer science,the traditional methods have shown many drawbacks:First,the simulation model lacks analysis of real traffic data including passenger behavior;Second,it is difficult for the linear model is to fit the traffic flow with complexity characteristics;Third,the traditional predictive model lacks an early warning mechanism with advance quantity.In view of the above problems,this paper takes the rail transit data collected by the sensor network in the city as the basis and analyzes the flow characteristics and passenger behavior in a data-driven manner with big data technology to provide information for the subsequent modeling tasks.The traditional model is combined with the machine learning model to fit the traffic flow with complex characteristics.Then we design an effective mechanism to early warning of the sudden passenger flow with advance quantity.We solve the sudden passenger flow problem in the subway system through the following three steps:Analyzing flow characteristics and mining passenger behavior;Predicting passenger flow for a period of time in the future;Early warning of sudden passenger flow and send warning signals without delay.Based on the above steps,this paper mainly includes the following four parts:In the first part of this paper,for the challenge of the massive amount and identification lack of hundreds of millions of smart card transactions data,we used distributed big data storage and computing technology to aggregate the traffic of the stations,and complete and extract the passenger trip information.Based on this,we analyzed the characteristics of passenger flow and the spatio-temporal behavior of individuals and groups.Compared with the existing methods,we proposed a novel analysis perspective,the analysis of passengers' round-trip behavior.The advantage of this perspective is that we can understand the distribution of the starting,ending time and the total duration of each passenger at different stations more deeply,so as to understand the passengers'regular behavior and abnormal behavior.The experimental results show that our analysis perspective and method can provide effective insights for forecasting and warning of rail transit traffic.In the second part of this paper,in view of the complex spatial correlation and time dependence of passenger flow,we proposed a deep learning model,Spatio-Temporal Attention Networks(STAN),and used multi-input and multi-output multi-step prediction strategy to solve the error accumulation in long-term passenger flow prediction.Compared with the existing methods,our method took into account the spatial correlation and time-dependent nature of passenger flow,and used the spatial sub-network to calculate the correlation of the traffic of the adjacent stations to the target station.We also used the temporal sub-network to calculate the historical traffic flow to predict the impact of future passenger traffic.The experimental results show that the prediction performance of our method is better than the traditional linear model,machine learning model and deep learning model at each time step,which reduce 7.66%-9.16%error compared to the baseline model.The multiple input and multiple output multi-step prediction strategy effectively reduces the error accumulation of the traditional iterative strategy.In the third part of this paper,we designed a novel responsive prediction model for the challenges of large-scale sudden passenger flow with low frequency,rapid change and mixed components.Based on the responsive prediction model,we designed a hybrid method to predict the peak of the burst passenger flow in rail transit system.Compared with the traditional method,the responsive prediction mode is based on the observed passenger behavior of sudden passenger flow,and a detection-based prediction idea is designed.The time series technology,machine learning algorithm and physical model are merged,and the abnormal outbound passenger flow is detected firstly.The responsive algorithm is then used to predict the peak of the sudden passenger flow triggered by the abnormal outbound passenger flow.The experimental results show that the prediction error of our method for peak passenger flow is 64.2%-70.1%lower than the traditional method,and it has good interpretability.In the fourth part of this paper,aiming at the large-scale sudden passenger flow and the shortage of prediction methods,we proposed early warning technology for sudden passenger flow,including elastic anomaly detection algorithm and online early warning mechanism.The goal is faster and earlier early warning of sudden passenger flow.Compared with the traditional method,the elastic anomaly detection algorithm can detect abnormal outbound passenger flow more efficiently,and the online early warning mechanism can timely send early warning of large-scale sudden passenger flow.The experimental results show that our early warning technology can warn of the sudden passenger flow 2 hours in advance,and the processing efficiency of the elastic anomaly detection algorithm is better than that based on the sliding window.