Elsevier

Neurocomputing

Volume 398, 20 July 2020, Pages 477-484
Neurocomputing

A novel learning method for multi-intersections aware traffic flow forecasting

https://doi.org/10.1016/j.neucom.2019.04.094Get rights and content

Abstract

Recent advances in machine learning have helped solve many challenges in artificial intelligence applications, such as traffic flow forecasting. Traffic flow forecasting models based on machine learning have recently been widely applied because of their great generalisation capability. This study aims to construct a multi-intersection-aware traffic flow prognostication architecture considering recent information of a nearby road, which is a significant indicator of the near-future traffic flow, and considering the selection of appropriate and essential sensors significantly correlated to the future traffic flow. To capture the inner correlation between sequential traffic flow data, a novel learning method involving the relevance vector machine is employed for the traffic flow forecasting. To optimise the kernel parameters of the relevance vector machine, a combination of the chaos theory and a simulated annealing algorithm is adopted. The proposed model is verified with the real-world data of six roads in a Minnesotan city. Then, the forecasting results of the new model are compared with those of some state-of-the-art models. These results indicate that the application of relevance vector regression to short-term traffic flow forecasting combined with a chaos-simulated annealing algorithm to optimise the corresponding parameters is a high-precision and -scalability short-term traffic flow forecasting method. The multi-intersection-aware mechanism helps improve the forecasting accuracy.

Introduction

Intelligent transportation is one of the important current research topics in the fields of intelligent transportation and urban intelligent big data processing [1], [2], [3]. A prerequisite of real-time traffic signal control, traffic assignment, route guidance, automatic navigation and accident detection in intelligent transportation systems is real-time and accurate traffic flow forecasting. Sensor systems installed on various road types provide support to traffic flow data to realise short-term traffic flow forecasting. These sensor systems typically provide information about the flow, speed and lane occupancy rate of a transportation network and use this information to forecast possible traffic flow conditions in 5 to 10 min [4], [5], [6], [7]. Two of the most prominent research areas in recent years are traffic flow forecasting and estimation of travel time delays, which can be used to assist forecasting by traffic control centres to improve traffic mobility. Traffic flow forecasting systems must be able to forecast the traffic flow in the next decision time and even several later moments at the time of making a decision about the control variables. Generally, a forecasting time span not greater than 15 min(or even less than 5 min) constitutes a short-term traffic forecast [8], [9], [10]. The accurate description and prediction of a rapidly changing urban road traffic flow are very complex; therefore, it is difficult to describe this with a single linear statistical algorithm. Existing research on traffic flow forecasting has concentrated primarily on fast urban roads. Accordingly, different forecasting methods, such as time series models [11], [12], artificial neural networks (ANNs) [13], [14], support vector machines (SVMs) [15], [16], [17] and deep learning-based methods [18], [19], have been used extensively in short-term traffic flow forecasting. These approaches, particularly SVMs and ANNs, are useful in discovering the inner statistical rules hidden in the sequential traffic flow data. The wide application of ANNs and SVMs offers significant opportunities in terms of the forecasting performance improvement of complicated nonlinear regression problems [7], [15], [16], [17], [18], [19]. Tipping [20] proposed a sparse Bayesian learning model relevance vector machine (RVM). Compared to SVMs, the relevance vector regression (RVR) demonstrates an outstanding performance in the regression forecasting accuracy rate. Regression forecasting models based on the RVR have already been applied in various aspects and, consequently have achieved excellent results, such as tracking of objects, software reliability prediction, fault diagnosis, pose estimation and channel equalisation forecasting [21], [22], [23], [24], [25], [26]. Motivated by the above mentioned considerations, this study establishes a RVR model to forecast short-term traffic flow based on chaotic simulated annealing optimisation(RVMCSA). The main contributions of this paper are summarised as follows:

  • First, for RVR, we incorporate a fantastic searching algorithm, called chaotic simulated annealing algorithm, to adapt its parameter. The RVRCSA model is adopted to forecast the traffic flow for the first time to describe the randomness in the sequential traffic flow data.

  • Second, a multi-intersection-aware traffic flow prognostication architecture, which can exploit the recent traffic information on a nearby road for the near-future traffic flow, is constructed.

  • Third, the experimental results demonstrate that the new method offers the advantages of higher forecasting accuracy compared with well-known techniques.

The remainder of this paper is organised as follows: after the related research works are discussed in Section 2, Section 3 outlines the principle of the proposed RVM used for regression estimation. Section 4 introduces the framework of multi-intersection-aware traffic flow prognostication based on the RVMCSA. Section 5 analyses the performance of the proposed novel model and, finally, Section 6 concludes the paper.

Section snippets

Related works

The most commonly used time series model is the autoregressive integrated moving average model (ARIMA). Williams et al. [27] proposed a multivariate ARIMA model comprising upstream traffic flow data and considering seasonal variation, which made a seasonal distinction between the peak and off-peak flows to improve the models forecasting ability. Thomas et al. [12] combined the traffic flow data collected by an upstream detector with ARIMA to build a real-time traffic flow forecasting model, and

Relevance vector machine used to regression estimation

The RVM can be employed to solve regression problems. Given a set of vectors {xi}i=1N and the corresponding target values as input, the correspondence relation between x and t conforms to the following function:t=y(x;ω)=i=1Mωik(x,xi)+ω0where k is a nonlinear kernel function; and ωi is the weight of the model that is non-zero only when xi belongs to the relevance vector. Use the probability as follows to describe the influence of ti in addition to the error:p(ti)=N(ti|y(xi;ω),σ2)A reasonable

The architecture of multi-intersections aware traffic flow prognostication

The primary factors associated with the traffic flow are the traffic flow of adjacent road sections, the traffic flow of the previous time period in a road section, average speed and lane occupancy. The architecture of the multi-intersection-aware traffic flow prognostication model is constructed, as shown in Fig. 1. If a detected sub-section set of the current road section is {Di,i=1,2,3,,L}, the forecast the flow data qnd(t+1) at time t+1 in nth detected sub-section Dn must be calculated.

Use of traffic flow data, parameter instructions and performance comparison

The traffic flow data employed in the experiment were collected by the Transportation Data Research Laboratory (TDRL) at the University of Minnesota Duluth, which used a remote transportation microwave detector to collect data from urban roads located in Minnesota Twin Cities and the City of Rochester [36]. The information for the six selected traffic sections in the experiments is as follows:

  • Traffic section 1: TH-252SB, station ID N89132

  • Traffic section 2: TH-77CDSB, station ID N88125

  • Traffic

Conclusions and further research

To develop a prediction model for the traffic flow data, the basic principles of RVMs and forecasting of short-term traffic states were employed. The application of RVMs combined with chaos genetic optimisation was discussed in the context of short-term traffic forecasting. The real-time data of roads in Minnesota were utilised to verify the model. The results indicated that the method proposed herein achieved some improvements in the prediction ability compared to similar methods. Moreover,

Declaration of Competing Interest

None.

Acknowledgment

This work was supported by the Natural Science Foundation of China [61772199, 61802123].

We would thank the Transportation Data Research Laboratory (TDRL) at the University of Minnesota Duluth for their released traffic flow data.

Zhangguo Shen, is a Ph.D. candidate in the College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou, China. He received the B.E. degree in Computer Science from Zhejiang University City College, Hangzhou, China, in 2004 and the M.S. degree in Computer Science from Hangzhou Dianzi University, Hangzhou, China, in 2007. His research interests include Traffic Flow Forecasting, Machine Learning and Cloud Computing.

References (36)

  • LingS.H. et al.

    A novel genetic-algorithm-based neural network for short-term load forecasting

    IEEE Trans. Ind. Electron.

    (2003)
  • ChenD.

    Research on traffic flow prediction in the big data environment based on the improved RBF neural network

    IEEE Trans. Ind. Inform.

    (2017)
  • P. Dell’Acqua et al.

    Time-aware multivariate nearest neighbor regression methods for traffic flow prediction

    IEEE Trans. Intell. Transp. Syst.

    (2015)
  • YuJ. et al.

    Leveraging content sensitiveness and user trustworthiness to recommend fine-grained privacy settings for social image sharing

    IEEE Trans. Inf. Forensics Secur.

    (2018)
  • ZhangJ. et al.

    Local deep-feature alignment for unsupervised dimension reduction

    IEEE Trans. Image Process.

    (2018)
  • YuJ. et al.

    Learning to rank using user clicks and visual features for image retrieval

    IEEE Trans. Cybern.

    (2015)
  • LouJ. et al.

    Improving traffic flow forecasting with relevance vector machine and a randomized controlled statistical testing

    Soft Comput.

    (2019)
  • HeP. et al.

    Model approach to grammatical evolution: deep-structured analyzing of model and representation

    Soft Comput.

    (2017)
  • Cited by (0)

    Zhangguo Shen, is a Ph.D. candidate in the College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou, China. He received the B.E. degree in Computer Science from Zhejiang University City College, Hangzhou, China, in 2004 and the M.S. degree in Computer Science from Hangzhou Dianzi University, Hangzhou, China, in 2007. His research interests include Traffic Flow Forecasting, Machine Learning and Cloud Computing.

    Wanliang Wang, received the B.S. degree in industrial automation from Jiangsu University, Jiangsu, China, in 1982, and the Ph.D. degree in control theory and engineering from Tongji University, Shanghai, China, in 2001. He is currently a Professor with the School of Computer Science and Technology, Zhejiang University of Technology, Hangzhou, China. His current research interests include Wireless Network, Artificial Intelligence, and Intelligent Control.

    Qing Shen, received her Master’s degree in Computer Application Technology in July 2007 from North University of China. She is an associate professor in the Computer Science Department of Huzhou Teachers College, Huzhou, Zhejiang. Her current research interests include Artificial Intelligence, Machine Learning and Software Reliability Evaluation, etc.

    Shaojun Zhu, received his Ph.D. degree from Ningbo University, China, in 2014. Is an associate professor in the Computer Science Department of Huzhou Teachers College, Huzhou, Zhejiang. His research interests include Machine Learning and Image Processing.

    Jungang Lou, received his M.Sc. in Computational Mathematics (2006) from Tongji University and Ph.D. in Computer Software and Theory (2010) from Tongji University. He is now a professor in the School of Information Engineering at Huzhou University, and he is also a Postdoctoral in the institute of Cyber-Systems and Control, Department of Control Science and Engineering, Zhejiang University. He was a Visiting Scholar with the department of Computer Science at The University of Texas at San Antonio between Nov. 2017 and May 2018 (advisor Professor Qi Tian, IEEE Fellow). His research interests include dependable computing, software reliability evaluation, computer system performance evaluation, neural network optimization and time series prediction. He has published over 60 papers in refereed international journals including IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on Cybnetics, Neurocomputing and so on.

    View full text