Modeling multi-regional temporal correlation with gated recurrent unit and multiple linear regression for urban traffic flow prediction

Abstract

Urban traffic flow prediction has received much attention in the past few years, especially after the availability of huge traffic data. In addition, the efficacy of some existing traffic flow techniques heavily depends on some influential external factors (like weather, geographical information, point-of-interests (POIs), or road information). However, exploitable data containing such auxiliary information are extremely limited. One primary reason is that some rising cities lack the resources necessary for collecting such data. Therefore, it is difficult to pick an effective method that can accurately predict the traffic state when the data is limited. Due to that, this paper proposes a framework for achieving accurate predictions based on four data characteristics and the temporal correlation of neighbors without using additional influential factors. Specifically, we first propose a novel Multi-Region Correlation (MRC) method to make regions cooperate and share their traffic flow and patterns. Then, a deep recurrent network is used to capture the neighbors’ temporal correlation by integrating their traffic history that flow among regions and their neighbors. When analyzing the traffic data history, we model four data characteristics (seasonality, trend, residual, and cyclic). And lastly, we employ a Multiple Linear Regression Unit (MLRU) to predict the future traffic directly from the neighbors’ traffic. We conducted extensive experiments on two real-world datasets collected from two major cities in China, Chengdu and Xi’an. The results demonstrate that the proposed model can achieve superior performance compared to many existing models in terms of accuracy and efficiency.

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.