摘要
城市轨道交通作为一种安全、运量大、环保节能的交通工具,能有效缓解城市交通压力,逐渐成为大中城市居民最重要的出行方式之一。准确可靠的城市轨道交通短期客流预测对旅客出行与客流管控有重要意义。有鉴于此,提出一种新型的生成对抗网络(GAN)模型,即CWGAN-div模型,以实现对地铁OD需求的短期预测。CWGAN-div模型融合条件生成对抗网络(CGAN)模型以及基于Wasserstein散度的生成对抗网络(WGAN-div)模型,结合2种模型的特点,来提高原始生成对抗网络模型的稳定性和生成精度。考虑到地铁客流量变化的时间周期性,使用一种融合2类周期信息的时间标签作为条件信息,并与历史OD数据一起作为模型的输入。为了更充分、更稳定地挖掘地铁客流需求的时空相关性,采用一种改进的卷积神经网络,即残差神经网络构建CWGAN-div的内部结构。以深圳地铁1号线和4号线的44个站点为例,数值实验表明,CWGAN-div模型具有较好的稳定性和预测效果,相比传统预测方法和普通深度学习方法在预测精度上分别提高了27.97%和6.59%,相比其他组合模型提高了3.26%,相比基础CGAN模型和WGAN-div模型预测精度分别提高了3.83%和9.51%,且残差神经网络结构能够提升模型的稳定性,加快模型收敛。由此可见,CWGAN-div模型在预测短期地铁OD需求方面具有研究意义与现实意义。
As a safe,large volumed,and environmentally friendly public transport,urban rail transit can relieve the pressure of the urban traffic system and gradually become an important transport in cities.Accurate and reliable prediction of short-term metro demand is essential to passengers and traffic managers.Thus,this paper proposed a novel Generative Adversarial Network(GAN) model that combined Conditional GAN(CGAN) and GAN with Wasserstein divergence(WGAN-div) named CWGAN-div to achieve short-term prediction of metro origin-destination(OD) demand.The CWGAN-div combined the advantages of CGAN and WGAN-div to improve the stability and accuracy of the original GAN.The input of the model consisted of historical data and conditional information to consider temporal dependences.To fully utilize the temporal-spatial correlations of the data,the residual neural networks(ResNets) were taken as the internal structure of the model.The 44 stations of Shenzhen Metro Line 1 and Line 4 were taken as examples.The results of numerical experiments show that the proposed model not only has good robustness due to the use of ResNets,but also outperforms the classic model,traditional deep learning model,other hybrid models and basic GAN models.The proposed CWGAN-div has practical significance in predicting short-term metro OD demand.
作者
申慧涛
郑亮
李树凯
王璞
SHEN Huitao;ZHENG Liang;LI Shukai;WANG Pu(School of Traffic and Transportation Engineering,Central South University,Changsha 410075,China;State Key Lab of Rail Traffic Control&Safety,Beijing Jiaotong University,Beijing 100044,China)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2022年第6期1530-1539,共10页
Journal of Railway Science and Engineering
基金
轨道交通控制与安全国家重点实验室(北京交通大学)开放课题基金资助项目(RCS2022K004)
国家自然科学基金资助项目(71871227)
中南大学创新驱动计划(2019CX018)
湖南省自然科学基金资助项目(2021JJ30888)。
关键词
城市轨道交通
地铁OD需求
短时交通预测
生成对抗网络
残差神经网络
urban rail transit
metro OD demand
short-term traffic prediction
Generative Adversarial Network
residual neural network