Modeling multi-regional temporal correlation with gated recurrent unit and multiple linear regression for urban traffic flow prediction
Introduction
The rapid development of urbanization has led to the modernization of many metropolitans and deeply affects people’s life. In addition, that growth led to the emergence of huge data, like human mobility, geographic data, and traffic flow. But this development has also caused many issues that need proper management, such as pollution, energy consumption, and traffic congestion. Therefore, urban computing aims to provide a data-driven approach to these problems by leveraging the huge data generated in cities [1]. Urban computing has provided a wide range of applications in different areas such as transportation, environment, social, energy, and urban planning. As one of the major components, urban traffic flow prediction for the Intelligent Transportation Systems (ITS) has received enormous attention in the past decades. Effective traffic flow modeling can reduce traffic congestion or air pollution and helps decision makers achieve good plans for city management. Moreover, it provides early warnings for public safety emergency management. Nevertheless, it is very challenging to perform such a prediction because of the dynamic and complex traffic situations in large cities [1], [2], [3]. Recently, many models were done towards understanding and predicting traffic flow [4], [5], [6]. However, for different reasons such as limited data resources, models and algorithms that depend on external information are not perfect or suitable for some situations. Many previous works on traffic prediction use weather datasets and spatial datasets such as geographical topology, POIs, and road segment information besides historical traffic to predict future traffic flow. However, only a few public datasets publicly contain all of this information, and some newly emerging cities may not have the techniques to collect such data. Due to that, those models will not work for such situations.
In our study, we assume that cities lack the capability or the technology of collecting data from various sources that can be used in the traffic forecasting process. Hence, we aim at developing effective and efficient urban traffic flow forecasting models solely based on traffic data, without using external information. In particular, we focus on modeling the essential forecasting characteristics in time series analysis, i.e. seasonality, trend, residual, and periodicity, as well as modeling the strength of the temporal effect of the traffic flow among neighbors. Such crucial aspects had not been thoroughly examined in most traffic prediction studies. This study will demonstrate that accurate predictions can be achieved by integrating these predictors and neighbors’ temporal correlation.
To introduce our method at a high-level, we analyze the traffic history of neighboring regions and model their four characteristics by using an encoder–decoder model and use a Multiple Linear Regression Unit (MLRU) to aid the decoder in the prediction task. The learning manner behaves as follows: a deep recurrent neural network encodes the temporal correlation between regions and their neighbors; the MLRU is used to predict the future traffic directly from the neighbors’ traffic and feed the output to the decoder that uses attention-based techniques to let regions be aware of their neighbors’ influences. Therefore, our method is based on regions’ correlation and deep sequence learning of the four time series characteristics. This paper’s contributions can be summarized as follows:
- •
We propose a model (MRC-MLRU) to predict urban traffic flow based on the correlation of regions and use a deep recurrent network to capture such correlations. This model uses the attention technique and thus can capture the inherent relationship among regions and their neighbors’ influence.
- •
We employ MLRU to predict the future traffic directly from the neighbors’ traffic by converting the time series prediction problem to a multiple linear regression problem and combining the attention layers and MLRU outputs. As a result, the model can overcome the exposure bias problem and prevent the model’s hidden states from being updated by wrong sequence predictions when the model predictions are very bad at the early stages.
- •
To capture the influence of the temporal correlation, we model four characteristics of traffic data: seasonality, trend, residual, and cyclic, which can model diverse temporal correlations to improve model performance.
- •
We evaluate our model on two real-world datasets and find that the model can greatly enhance urban traffic prediction and gain better traffic-related knowledge from the neighbors’ traffic and the components’ attributes.
- •
We release the datasets for predicting traffic flow derived from the ordered datasets of the ride requests collected by DiDi company. We have processed and reformulated it as a traffic forecasting dataset. The code and datasets have been released.1
this paper is organized as follows. Section 1 contains the introduction,Section 2 presents the related work. Following that, preliminaries and problem definition are given in Section 3. Section 4 details the proposed method. The qualitative and quantitative results of different methods are presented in Section 5. The paper is concluded in Section 6 Finally, In Appendix, we introduce the released datasets and the preprocessing steps.
Section snippets
Related work
Machine learning theories have been widely used for time series forecasting in many areas like weather forecasting [7], financial market prediction [8], and the ITS [9], [10]. As traffic prediction is such an essential part of ITS, thus many works on prediction have been published. The following is an overview of literature related to traffic prediction. Ma et al. [11] use the Long–Short-Term Memory (LSTM) neural network [12] to predict traffic speed by capturing nonlinear and dynamic traffic
Overview
Here, we will introduce some preliminaries, notations, and problem formulation.
Methodology
In order to solve the traffic flow forecasting problem in the urban regions, we propose a model that captures the neighbors’ temporal correlation. Our model uses a seq2seq structure that includes an encoder for historical learning and decoder for making the future predict [14], and a multiple linear regression unit (MLRU). However, unlike the existing works, which only use the region’s traffic history to predict future traffic, we combine the region’s and its neighbors’ traffic history. This
Experiments
This section describes the datasets, the baseline methods, the parameter settings, and the implementation details; finally, we discuss the results of our model and compare them with the baseline methods.
We use two real-world datasets from two different cities in China: Chengdu and Xi’an cities to test the performance of the proposed model and the baseline models, as shown in Table 2. Both of the datasets are from the trajectory data of the DiDi drivers in Chengdu and Xi’an cities at the second
Conclusion
In this paper, we focused on a practical challenge in a traffic flow prediction scenario, which includes traffic prediction without external data. In order to achieve effective traffic flow prediction when the data is limited in terms of the availability of auxiliary information. We proposed a deep-learning-based approach for the traffic flow prediction problem, which incorporates the advantage of the region’s traffic history and the neighbors’ traffic. In addition, we integrate the traffic
CRediT authorship contribution statement
Taha M. Rajeh: Conceptualization, Methodology, Software, Investigation, Writing – original draft. Tianrui Li: Writing – review & editing, Funding acquisition, Supervision. Chongshou Li: Writing – review & editing. Muhammad Hafeez Javed: Software. Zhpeng Luo: Writing – review & editing. Fares Alhaek: Software.
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.
Acknowledgments
This research was supported by the National Key R&D Program of China (2019YFB2101802) and the National Natural Science Foundation of China (Nos. 62176221, 62276215). Data source: DiDi Chuxing GAIA Open Datasets.
References (48)
- et al.
Prediction of taxi destinations using a novel data embedding method and ensemble learning
IEEE Trans. Intell. Transp. Syst.
(2019) - et al.
Long short-term memory neural network for traffic speed prediction using remote microwave sensor data
Transp. Res. C
(2015) - et al.
Short-term traffic flow prediction based on a hybrid optimization algorithm
Appl. Math. Model.
(2022) - et al.
MapReduce accelerated attribute reduction based on neighborhood entropy with Apache Spark
Expert Syst. Appl.
(2023) - et al.
Urban computing: concepts, methodologies, and applications
ACM Trans. Intell. Syst. Technol.
(2014) - et al.
Data-driven intelligent transportation systems: A survey
IEEE Trans. Intell. Transp. Syst.
(2011) - et al.
A short-term traffic flow prediction model based on an improved gate recurrent unit neural network
IEEE Trans. Intell. Transp. Syst.
(2021) - et al.
When will you arrive? estimating travel time based on deep neural networks
- et al.
Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms
- et al.
Geoman: Multi-level attention networks for geo-sensory time series prediction
Fine-grained photovoltaic output prediction using a bayesian ensemble
A dual-stage attention-based recurrent neural network for time series prediction
Cooperative shock waves mitigation in mixed traffic flow environment
IEEE Trans. Intell. Transp. Syst.
Long short-term memory
Neural Comput.
Learning phrase representations using RNN encoder-decoder for statistical machine translation
Sequence to sequence learning with neural networks
Neural machine translation by jointly learning to align and translate
Attention is all you need
Graph attention networks
Semi-supervised classification with graph convolutional networks
Improving traffic prediction with tweet semantics
Collective traffic prediction with partially observed traffic history using location-based social media
Deep sequence learning with auxiliary information for traffic prediction
Deep spatio-temporal residual networks for citywide crowd flows prediction
Cited by (17)
A novel partial grey prediction model based on traffic flow wave equation and its application
2024, Engineering Applications of Artificial IntelligenceData-unbalanced traffic accident prediction via adaptive graph and self-supervised learning
2024, Applied Soft ComputingA generative adversarial network-based framework for network-wide travel time reliability prediction
2024, Knowledge-Based Systems