Traffic flow prediction over muti-sensor data correlation with graph convolution network
Introduction
Accurate and real-time traffic flow prediction is one of the most critical tasks in intelligent transportation systems (ITS) [1] and of great importance to traffic managers and travelers [2]. Knowing reliable traffic information (such as traffic congestion, traffic volume, etc.) in advance, can help traffic managers better formulate and implement traffic planning strategies, improve the operational efficiency of traffic networks, alleviate traffic congestion, and effectively reduce public safety risks. At the same time, it can help travelers better plan their travel routes, reducing time costs and economic losses. Therefore, traffic flow prediction has become an indispensable part of urban life and has attracted much attention.
Actually, in order to improve the accuracy of traffic prediction, researchers have made a lot of attempts. Traditional statistical methods were first applied to traffic flow prediction problems, such as Autoregressive Integral Moving Average (ARIMA) [3]. This type of model is only suitable for relatively stable and linearly changeable traffic flow prediction, which cannot meet the actual Application requirements [4]. After that, traditional machine learning models, such as Support Vector Machine (SVM) [5], Support Vector Regression Machine (SVR) [6], [7], Bayesian method [8] and K-nearest neighbor [9], were utilized to process highly non-linear data in traffic flow prediction, but their prediction performance relays on careful feature engineering, so these models still cannot be applied to the spatio-temporal correlation analysis of traffic flow data. In recent years, deep learning methods have been applied to traffic prediction problems. For examples, a Graph Convolutional Network (GCN) [10] is used to effectively extract the spatial features of traffic topology network, and a Convolutional Neural Network (CNN) [11] is used to process temporal features with adding the periodicity of the traffic data to improve the traffic flow prediction performance.
Although the above studies have achieved good results in the short term (5–15 min) traffic prediction, most of the existing methods still cannot make satisfactory progress in the medium to long term (15 min–1 h) traffic flow prediction, mainly due to the following challenges. 1) Sensitive periodic data. The periodicity of traffic flow data is greatly affected by people’s daily life, which leads to the difference among data periodic features. If improper data is selected as the input of the model, the prediction accuracy of the model will decrease. Therefore, how to reasonably analyze the data periodic and select the appropriate data makes the prediction of the distant future extremely challenging. (2) Complex traffic pattern changes. Fig. 1(b) demonstrates the changing relationship of traffic patterns among roads. The seven nodes (S1–S7) in the spatial dimension represent the highway network structure composed of sensors, while the three time slices in the temporal dimension represent the current road network structure at each moment. Solid lines between node pairs represent the degree of correlation of traffic patterns, and the darker the color, the higher the correlation. In short-term spatial–temporal dimension, sensor S4 is closely associated with its traffic patterns of the next moment, at the same time the nearby road (S5, S6 and S3) also make influence, however, this influence changes dynamically. For example, when S4 is located in residential areas, S4 is highly correlated with traffic patterns on nearby roads during peak commuting period, while its correlation with nearby roads decreases during non-commuting period. In the long-term spatial–temporal dimension, the traffic flow at the current moment at S4 will not only has a long-term impact on the traffic patterns at the future moment and at the same location, but also has an impact on the future moments of other roads with the traffic flow on the road network (for example, when there is a traffic jam at S4). In Fig. 1(b) when a special event occurs at S1 (such as the big games), many people from S4 (residential area) will gather at S1 and this will continue for a long time. During this period, two road traffic patterns with strong similarity and highly correlated, but this kind of correlation between roads also dynamically changes, so it is often hard to get effective analysis. Therefore, it is a challenging problem to reasonably analyze the changing relationship of traffic patterns among different roads and effectively capture the temporal and spatial correlation among them to predict the long-term traffic conditions.
In response to the above challenges, we propose a traffic flow prediction method based on MDCGCN to predict future traffic data. This model can reasonably analyze the periodicity of traffic data in a data-driven way and select the periodic data highly correlated with the current moment as the input. In addition, it can also analyze the changing relationship of traffic patterns among roads according to the traffic condition of the current period, and effectively capture the dynamic spatial–temporal correlation among them. Our major contributions are summarized as follows:
- •
We propose a benchmark adaptive mechanism to eliminate potential differences among periodic data. Specifically, the benchmark adaptive mechanism automatically selects the data highly related to the current time series from the time series of multiple daily period as the input of the component, improving the overall prediction performance of the proposed model.
- •
We propose a novel multivariate data associative convolution block, which aims to dynamically capture the spatio-temporal correlation caused by traffic pattern changes between roads in a data-driven way, effectively improving the model’s prediction performance in the medium and long term.
- •
We perform extensive experiments on real traffic datasets. These experimental results show that the proposed model is significantly better than the existing time series models and graph convolution models in the medium and long-term prediction effect.
Section snippets
Literature review
To improve the traffic planning capability of ITS, the problem of traffic prediction has attracted widespread attention of researchers, and achieved good results through continuous research and practice [12]. Generally speaking, traffic prediction methods are roughly divided into three categories: traditional statistical methods, traditional machine learning methods, and deep learning methods.
In the traditional statistical methods, the Historical Average method (HA) and Autoregressive
Preliminary
In this section, we give definitions of related concepts in the paper. Definition 3.1 (Traffic topology network): The topological structure of the traffic topology network is defined as a weighted undirected graph . In this paper, the road section covered by the road sensor is used as a node, where V is the node set, , and N is the number of node. E is the set of connected edges of the traffic network, indicating whether there is a real connection between the nodes. The adjacency matrix of
Multisensor data correlation graph convolution network model
Fig. 2 shows the multisensor data correlation graph convolutional network (MDCGCN) model proposed in this paper. The model is mainly composed of three parts, which make use of three attributes of traffic data: recent, daily period and weekly period components. In MDCGCN model, the daily period component includes a layer of benchmark adaptive mechanism (BA-Block). In addition, the three components also include two layer multisensor data correlation convolution modules, 2D standard convolution
Experiment
To evaluate the performance of the model, we performed comparison experiments on two real highway traffic datasets.
Conclusion
In this paper, we propose a MDCGCN model for traffic flow prediction. The model combines a benchmark adaptive mechanism and a multi-sensor data connection convolution block. The former considers the difference between periodic data to effectively improve the input quality of the data, and the latter considers the changing relationship between traffic patterns between roads, combining the traffic topology network and traffic pattern relationship graph to capture long-term spatiotemporal
CRediT authorship contribution statement
Wei Li: Resources, Supervision. Xin Wang: Conceptualization, Methodology, Investigation, Writing - original draft. Yiwen Zhang: Resources, Writing - review & editing, Supervision. Qilin Wu: Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
This work was supported by the National Science Foundation of China (No. 61872002) and the Natural Science Foundation of Anhui Province of China (No. 1808085MF197). Yiwen Zhang is the corresponding author of this paper.
Wei Li received her PhD degree in computer science in 2006 from Anhui University. She is a professor in the School of Computer Science and Technology at Anhui University. Her research interests include software engineering, deep learning, computer recognition technology.
References (51)
- et al.
Statistical methods versus neural networks in transportation research: Differences, similarities and some insights
Transportation Research Part C: Emerging Technologies
(2011) Intersection traffic flow forecasting based on ν)gsvr with a new hybrid evolutionary algorithm
Neurocomputing
(2015)- et al.
Combining kohonen maps with arima time series models to forecast traffic flow
Transportation Research Part C: Emerging Technologies
(1996) - et al.
A novel wavelet-svm short-time passenger flow prediction in beijing subway system
Neurocomputing
(2015) Traffic flow forecasting by seasonal svr with chaotic simulated annealing algorithm
Neurocomputing
(2011)- et al.
A distributed spatial–temporal weighted model on mapreduce for short-term traffic flow forecasting
Neurocomputing
(2016) - et al.
Urban traffic flow forecasting through statistical and neural network bagging ensemble hybrid modeling
Neurocomputing
(2015) - et al.
Long short-term memory neural network for traffic speed prediction using remote microwave sensor data
Transportation Research Part C: Emerging Technologies
(2015) - et al.
Long-term time series prediction with the narx network: An empirical evaluation
Neurocomputing
(2008) - et al.
Traffic flow prediction using lstm with feature enhancement
Neurocomputing
(2019)
Lstm-based traffic flow prediction with missing data
Neurocomputing
Lstm network: a deep learning approach for short-term traffic forecast
IET Intelligent Transport Systems
Data-driven intelligent transportation systems: A survey
IEEE Transactions on Intelligent Transportation Systems
Applications of computational intelligence in vehicle traffic congestion problem: a survey
Soft Computing
Short-term prediction of traffic volume in urban arterials
Journal of Transportation Engineering
A tutorial on support vector regression
Statistics and Computing
Travel time estimation for ambulances using bayesian data augmentation
The Annals of Applied Statistics
Nonparametric regression and short-term freeway traffic forecasting
Journal of Transportation Engineering
Deep learning
Nature
Joint optimization of offloading utility and privacy for edge computing enabled iot
IEEE Internet of Things Journal
Modeling and forecasting vehicular traffic flow as a seasonal arima process: Theoretical basis and empirical results
Journal of Transportation Engineering
T-gcn: A temporal graph convolutional network for traffic prediction
IEEE Transactions on Intelligent Transportation Systems
Pa-crt: Chinese remainder theorem based conditional privacy-preserving authentication scheme in vehicular ad-hoc networks
IEEE Transactions on Dependable and Secure Computing
Cited by (60)
FD-TGCN: Fast and dynamic temporal graph convolution network for traffic flow prediction
2024, Information FusionBayesian graph convolutional network for traffic prediction
2024, NeurocomputingSpatiotemporal hierarchical transmit neural network for regional-level air-quality prediction
2024, Knowledge-Based SystemsRegional traffic flow combination prediction model considering virtual space of the road network
2024, Physica A: Statistical Mechanics and its ApplicationsResearch on intelligent vehicle Traffic Flow control algorithm based on data mining
2024, International Journal of Intelligent Networks
Wei Li received her PhD degree in computer science in 2006 from Anhui University. She is a professor in the School of Computer Science and Technology at Anhui University. Her research interests include software engineering, deep learning, computer recognition technology.
Xin Wang received his bachelor degree in computer science and technology in 2016 and now is a master student in the School of Computer Science and Technology at Anhui University. His research interests include machine learning and deep learning.
Yiwen Zhang received his PhD degree in management science and engineering in 2013 from Hefei University of Technology. He is a professor in the School of Computer Science and Technology at Anhui University. His research interests include neural computing, service computing and machine learning. More details about his research can be found athttps://bigdata.ahu.edu.cn.
Qilin Wu received his PhD degree in computer application technology in 2011 from Hefei University of Technology. He is a professor in the School of Information Engineering at Chaohu University. His research interests include resource allocation and optimization for wireless networks, edge computing and service computing.