基于SVAE-WGANGP的地点车速数据质量恢复研究

Method for Spot Speed Data Quality Recovery Based on Spatiotemporal Variational Autoencoder with W-generative Adversarial Network-GP

为了解决因设备长期失修造成的数据大量缺失和传统数据修复方法无法表示上下文时空关系以及不规则时序特征的问题,提出一种时空生成对抗变分自编码网络(Spatiotemporal Variational Autoencoder with W-Generative Adversarial Network-GP, SVAE-WGANGP),用以恢复地点车速数据质量。该方法以生成对抗变分自编码网络为模型基本框架,直接学习自然缺失数据集的概率分布;基于改进时空信息单元的变分自编码生成网络提取数据在缺失模式下的隐式不规则时序特征与显式上下文时空相互依赖信息;利用对抗训练策略(Wasserstein GAN with Gradient Penalty, WGAN-GP)优化深度全连接判别网络,以获得最优重构数据。借助乌鲁木齐市某路网46天实际卡口地点车速实例验证模型合理性,结果表明:与其他6个基准模型的评估指标均值相比,PMCR机制下,所提方法的均方根误差(RMSE)和平均绝对误差(MAE)降低幅度分别在0.794~0.332和0.899~0.321,决定系数R^(2)升高幅度在3.175%~60.918%;LMR机制下,所提方法的RMSE和MAE平均降低幅度分别在0.600~0.222和0.773~0.208,R^(2)平均升高幅度在4.681%~91.518%;BMR机制下,所提方法的RMSE和MAE平均降低幅度分别在0.212~0.625和0.269~0.715,R^(2)平均升高幅度在5.309%~49.671%。SVAE-WGANGP在恢复不同缺失机制下的路网地点车速数据质量时具备较优精确性和良好普适性,交通时空信息和不规则时序特征对该模型的数据质量恢复性能具有一定贡献性。此外,在BMR机制下,SVAE-WGANGP的运算耗时均值较VAE-GAN的均值降低0.421 s,与其他5个基准模型相比,增长幅度在0.155~12.518 s。从整体来看,该方法在恢复数据时具有较高的时效性。

A novel generative adversarial variational autoencoder model called spatiotemporal variational autoencoder with W-generative adversarial network-GP(SVAE-WGANGP)for spot speed data quality recovery was developed.The aim was to address the problems of various data missing due to equipment disrepair and the inability of traditional data repair methods to represent the context relationship in spatiotemporal domain and irregular temporal features.A variational autoencoder with a generative adversarial network was used as the model framework to directly learn the probability distribution of natural missing data.The variational autoencoder generator,which is based on an improved spatiotemporal information unit,captured implicit irregular temporal features and explicit contextual spatiotemporal interdependence.An adversarial training strategy based on the Wasserstein generative adversarial network with gradient penalty(WGAN-GP)optimized a deep fully connected discriminator network to obtain optimal reconstruction data.The rationality of the model was verified using an example of the bayonet spot speed in an Urumqi road network for 46 days.The experimental results indicate that under point missing completely at random(PMCR)mechanism,compared with the average values of the six benchmark models,the decrease in the root-mean-square error(RMSE)and mean absolute error(MAE)of the proposed method ranges from 0.794 to 0.332 and from 0.899 to 0.321,respectively,and the increase in the coefficient(R^(2))ranges from 3.175%to 60.918%.In the line missing at random(LMR)mechanism,the decrease in the RMSE and MAE of the proposed method ranges from 0.600 to 0.222 and from 0.773 to 0.208,respectively,and the increase in R^(2) ranges from 4.681%to 91.518%.In the block missing at random(BMR)mechanism,the decrease in the RMSE and MAE of the proposed method ranges from 0.212 to 0.625 and from 0.269 to 0.715,respectively,and the increase in R^(2) ranges from 5.309%to 49.671%.The SVAE-WGANGP model exhibits improved accuracy and universality for recovering spot speed data quality in road network in different missing mechanisms,whereas the spatiotemporal traffic information and missing characteristics significantly contribute to the data quality recovery performance of the proposed model.Furthermore,in the BMR mechanism,the average running time of SVAE-WGANGP was shortened by 0.421 s compared with that of variational autoencoder with generative adversarial network,and the growth rate ranges between 0.155 and 12.518 s compared with those of the other five benchmark models.Overall,the proposed method has superior efficiency in recovering missing data.