Short-term traffic flow prediction: From the perspective of traffic flow decomposition

Abstract

Some researchers treat traffic flow as an entirety while predicting short-term traffic flow. Through analyzing real-world traffic flow, we have found that urban traffic shows a stable changing process along with random disturbs. An alternative way is to decompose traffic flow into two components: periodicity and volatility. We propose a hybrid method named Time-Series Analysis and Supervised-Learning (TSA-SL) for short-term traffic flow prediction from the perspective of traffic flow decomposition. In the method, period traffic flow is modeled with a typical TSA method called Fourier Transform (FT), where periodic behaviors are described as the combination of sines and cosines. The volatility of the current location is determined by its surroundings, so spatial–temporal correlations are extracted as input features of SL. Then, three hybrid prediction models, including FT-SVR, FT-GBRT and FT-LSTM, are built with proposed TSA-SL. In the experiment, an Electronic Registration Identification (ERI) dataset including massive real-world individual trajectories is employed. Comparing with classical baseline models, our proposed TSA-SL method has certain superiority. Furthermore, we decompose traffic flow into different components in terms of traveling purposes and vehicle types. The experimental results show that our method performs better in predicting partial traffic flow than predicting all traffic flow.

Introduction

Short-term traffic flow prediction has always been a hot research topic, and is also experiencing a huge breakthrough owing to the advances in sensing technology and artificial intelligence technology. The necessity of predicting urban road short-term traffic flow lies in its benefit to real-time traffic management for government and travel planning for residents. In our opinion, the main challenges of short-term traffic flow prediction can be seen from 2 aspects:

Traffic flow decomposition: On one hand, some researchers treat traffic flow as an entirety. From our observation, urban road short-term traffic can be decomposed by different means. Firstly, urban road short-term traffic flow is a stable general changing process along with some random disturbs [1]. The stability or randomness allows us to model urban road traffic separately. For example, it is believed that time-series can be decomposed into three components: a periodic trend, deterministic part and volatility [2]. Secondly, there are commuting vehicles and non-commuting vehicles from the perspective of traveling purposes and there are private cars, taxis, trucks, buses and other vehicles from the perspective of vehicles’ types. With different traveling purposes and vehicles’ types, vehicles show different traveling behaviors, which cause different traffic flow distribution in different time and areas. It has been proved that decomposing traffic flow data into different components and analyzing them with suitable methods is beneficial [2]. On the other hand, predicting different kinds of traffic flow shows more profound meanings. Traditionally, researchers predict traffic by using loop detector data [3], [4], GPS trajectories [5], and Vehicle License Plate Recognition (VLPR) [6], [7]. However, their work either predicts overall traffic flow or traffic flow of a certain kind of vehicle, such as taxis. Decomposition of traffic flow in case of traveling purposes and vehicles’ types has not been fully discovered yet. Especially in nowadays, the emergence of on-line car-hailing services makes the prediction of private cars’ behaviors more complex.

Spatial–Temporal (ST) correlations: The road network is a complex and connected network where traffic in one area is not only related to historical traffic temporally but also affected by neighbor areas’ traffic spatially. This phenomenon reflects the spatial–temporal correlations of urban road traffic flow. In the very beginning, short-term traffic flow prediction is seen as time-series analysis, and abundant methods have been proposed. Among these methods, Autoregressive Moving Average model (ARMA) and its variants, such as seasonal ARIMA(SARIMA) [8], and Vector ARMA(VARMA) [9] have been proved to be effective against traffic flow prediction problems. Besides, owing to the advanced sensing technologies, such as microwave detecting, Radio Frequency Identification(RFID), GPS and cellular networks, collecting traffic information becomes easier and quicker. With the help of citywide traffic information, such as road networks, traffic flow and individual trajectory, traffic flow prediction from a data-driven view becomes more convenient. Thus, machine learning algorithms, such as regression methods (SVR [10] and GBRT [11]), artificial neural networks(ANNs) have also been used to predict traffic information. In order to well capture spatial and temporal features, some deep neural networks (DNN)-based prediction models have been brought into play, such as DeepSTD [6], MST3D [12] and DMVST-Net [13]. These models generally view the city as a figure, then Convolution Neural Networks (CNN) or Long Short-Term Memory (LSTM) is used to extract spatial and temporal features, thus the citywide traffic flow can be predicted with considering ST features. Recently, attention-based models and graph convolution networks have also drawn much attention and have been widely used in traffic prediction [14], [15], which show some superiority.

To sum up, we draw two gaps from the above analysis:

• Decomposition of traffic flow only considers traffic flow characteristics, such as periodic trend and volatility, and the effectiveness of other kinds of decomposition has not been discovered because of the lack of information.

• Time-series analysis methods have certain advantages in traffic prediction, but they lack ability of dealing with spatial–temporal correlations which has been proved effective in traffic prediction.

In view of above two gaps, we plan to research hybrid prediction methods that can both make use of traffic flow decomposition and deal with spatial–temporal correlations. We believe that the traffic flow consists of a periodicity part and a volatility part. Different from previous work that utilizes time-series or supervised learning, we dedicate to combine them together, aiming at capturing periodicity and volatility at the same time by mining spatial–temporal traffic data. The fundamental assumption of our model is that the periodicity of traffic data is more regular and easy to predict with interpretable methods, such as time-series analysis methods, and the volatility is harder to model with linear equations and more suitable to predict with advanced methods, such as supervised learning. Additionally, we also attempt to decompose traffic flow according to people’s traveling purposes and vehicle’s types to conduct some experiments.

The main contributions of this paper are summarized as follows:

• We perform traffic analysis from 3 aspects: periodicity and volatility, traveling purposes, and vehicles’ types. On one hand, real-world traffic flow shows strong periodicity along with slight volatility which provides us with an opportunity to model them separately. On the other hand, vehicles of different traveling purposes or different types show different traveling behaviors, which cause that different traffic flow components follow different changing process.

• We propose a novel TSA-SL method which combines Time-Series Analysis(TSA) and Supervised Learning(SL). We firstly decompose traffic flow into periodicity part and volatility part. Then, we predict the future periodicity value and volatility value separately. In the periodicity part, we employ a typical time-series analysis method called Fourier Transform(FT) to model periodicity. In the volatility part, supervised learning methods including SVR, GBRT, LSTM are utilized to model volatility for their superiority in dealing with spatial–temporal correlations. Finally, three hybrid prediction models including FT-SVR, FT-GBRT, FT-SVR are built with our proposed hybrid TSA-SL method.

• We conduct abundant experiments on a large Electronic Registration Identification(ERI) dataset including massive real-world individual trajectories. We compare our three hybrid prediction models with some baseline models. The experimental results show the superiority of our models in short-term traffic flow prediction in terms of two metrics: MAE and MAPE. Furthermore, we decompose traffic flow into different components in terms of traveling purposes and vehicle types to conduct additional experiments. The results show that our method has better improvements in predicting partial traffic flow than predicting all traffic flow, which also suggests the positive influence of traffic decomposition.

The remainder of this paper is structured as below. Section 2 reviews related work and states the issue of this paper. Section 3 introduces our data sources and performs traffic flow component analysis. Section 4 presents the methodology, and discusses the modeling of periodicity with time-series analysis and volatility with supervised learning respectively. In Section 5, we detail the baseline models and analyze the experiment results. Section 6 concludes our work and puts forward some future work.