TransFusion Model Fusion Mechanism Based on Transformer for Traffic Flow Prediction

TransFusion Model Fusion Mechanism Based on Transformer for Traffic Flow Prediction

Xintong Song, Donghua Yang, Yutong Wang, Hongzhi Wang, Jinbao Wang, Bo Zheng
Copyright: © 2023 |Pages: 14
DOI: 10.4018/JDM.325353
Article PDF Download
Open access articles are freely available for download

Abstract

In recent years, the problem of traffic congestion has become a hot topic. Accurate traffic flow prediction methods have received extensive attention from many researchers all over the world. Although many methods proposed at present have achieved good results in the field of traffic flow prediction, most of them only consider the static characteristic of traffic data, but do not consider the dynamic characteristic of traffic data. The factors that affect traffic flow prediction are changeable, and they will change over time. In response to this dynamic characteristic, the authors propose a model fusion mechanism based on transformer (TransFusion). The authors adopt two basic forecasting models (TCN and LSTM) as the underlying architectures. In view of the performance of different models on the traffic data at different times, the authors design a model fusion mechanism to assign dynamic weights to basic models at different times. Experiments on three datasets have proved that TransFusion has a significant improvement compared with basic models.
Article Preview
Top

1. Introduction

The explosive growth of urban population and the increase of vehicles are likely to cause traffic congestion. Traffic accidents have become another turbulence factor in people’s lives. At the same time, traffic congestion has caused a great burden on the environment. The traffic problem has received considerable attention all around the world. There are many factors influencing traffic conditions. Because the traffic state is gathered by human activities, traffic conditions are different at different times and regions (such as the regular congestion the morning and evening traffic peaks, and the more vehicles in the center of the city). In addition to basic factors such as time and region, weather and solar terms are also important factors that affect traffic conditions. Complex factors make it difficult to alleviate traffic congestion. In order to solve the problem of traffic congestion, researchers from various countries have put forward the Intelligent Transportation System (ITS).

Intelligent transportation system (ITS) is a non-linear and time-variant system. Its technical core is traffic prediction. Traffic prediction is to predict the current or next traffic flow through the traffic flow at the previous moments. The status information mainly includes traffic flow, road structure, average vehicle speed, etc. The traffic flow information is particularly critical. Traffic flow refers to the number of vehicles passing through the current section of a highway in a unit period. The traffic flow can clearly see the congestion degree of the current section of a highway. In order to make more accurate prediction of traffic flow in the next time, predicting traffic flow requires not only knowing the traffic flow in the area at historical moments, but also combining local time, region, weather, solar terms and other factors.

The traffic prediction problem is a complex time varying problem. To tackle this problem, an increasing number of traffic conditions forecasting models have been proposed in relevant literature during the past several years, like ARIMA28, SVR9, SAE21, LSTM34, Conv-LSTM20. The spatial-temporal characteristic of traffic data is dynamically related. For example, during the peak period of traffic flow, the central area of the city is inconvenient for vehicles to enter and exit due to traffic congestion. There is little difference in traffic flow between the previous moment and the next moment. The temporal characteristic of traffic data has a greater impact on the predictive effect. At midnight, because the traffic flow of the entire city is small, in a certain period of time, vehicles can reach a farther location from the current location. Therefore, the traffic flow of the area is greatly affected by the traffic flow of the surrounding areas. It is difficult to find a single model with good performance to predict traffic flow at all times, which hampers the improvement of performance. Although some hybrid models simultaneously mine the spatial-temporal relevance of traffic data, the features they get are static, not dynamic. In order to better capture dynamic spatio-temporal characteristic, in this paper we propose a model fusion mechanism based on Transformer, referred to as TransFusion. We select some models that are general in spatio-temporal network data prediction as baseline models. The Transformer layer is used to assign different weights to different models at different times. Dynamic weight distribution can make full use of the advantages of each baseline model at different times to consistently outperforms other baselines. Besides, TransFusion is data-driven and don’t need external information(e.g. location information of sensors, topological map of the road). The contribution of this paper can be summarized as follows:

  • We propose a model fusion mechanism TransFusion to dynamically combine some simple models. TransFusion can extract the dynamic spatio-temporal relevance of traffic data by assigning appropriate weights to different models. TransFusion is data-driven and don’t need external information(e.g. location information of sensors, topological map of the road).

  • Experiments on three datasets demonstrate the superior performance of TransFusion for Traffic Flow Prediction.

Complete Article List

Search this Journal:
Reset
Volume 35: 1 Issue (2024)
Volume 34: 3 Issues (2023)
Volume 33: 5 Issues (2022): 4 Released, 1 Forthcoming
Volume 32: 4 Issues (2021)
Volume 31: 4 Issues (2020)
Volume 30: 4 Issues (2019)
Volume 29: 4 Issues (2018)
Volume 28: 4 Issues (2017)
Volume 27: 4 Issues (2016)
Volume 26: 4 Issues (2015)
Volume 25: 4 Issues (2014)
Volume 24: 4 Issues (2013)
Volume 23: 4 Issues (2012)
Volume 22: 4 Issues (2011)
Volume 21: 4 Issues (2010)
Volume 20: 4 Issues (2009)
Volume 19: 4 Issues (2008)
Volume 18: 4 Issues (2007)
Volume 17: 4 Issues (2006)
Volume 16: 4 Issues (2005)
Volume 15: 4 Issues (2004)
Volume 14: 4 Issues (2003)
Volume 13: 4 Issues (2002)
Volume 12: 4 Issues (2001)
Volume 11: 4 Issues (2000)
Volume 10: 4 Issues (1999)
Volume 9: 4 Issues (1998)
Volume 8: 4 Issues (1997)
Volume 7: 4 Issues (1996)
Volume 6: 4 Issues (1995)
Volume 5: 4 Issues (1994)
Volume 4: 4 Issues (1993)
Volume 3: 4 Issues (1992)
Volume 2: 4 Issues (1991)
Volume 1: 2 Issues (1990)
View Complete Journal Contents Listing