基于动态生成对抗网络的路网缺失交通数据修复

Missing Traffic Data Recovery for Road Network Based on Dynamic Generative Adversarial Network

针对智能交通系统数据采集过程中发生的数据缺失问题,本文提出一种基于动态生成对抗网络(dynamic generative adversarial network,D-GAN)的路网交通数据修复方法。该方法首先依据交通数据的时空特性与设定的缺失类型和缺失比例来构造各种缺失交通数据矩阵,然后基于博弈思想迭代训练由2个全连接神经网络构成的生成对抗网络。引入一种新颖的动态自适应机制,研究能在模型计算过程中自动识别生成器与判别器的最佳迭代次数,最终生成完整的交通数据矩阵并修复缺失值。采用加州PeMS和广州交通速度数据集来完成D-GAN模型的构建,并使用多种评价指标评估D-GAN的修复性能。实验结果表明:相对于非随机缺失模式,D-GAN对随机缺失模式的修复精度更高;随着缺失率增加,D-GAN的修复精度加速下降。但在各种缺失条件下,D-GAN模型的修复性能要优于现有模型(例如BGCP、prophet-RF和GAIN)。

For the issue of missing data occurring in the data collection process of intelligent transportation systems,a road network traffic data recovery model is proposed in this paper based on the dynamic generative adversarial network.Firstly,the approach in this paper are constructed various missing traffic data matrices by considering the spatial-temporal properties of traffic data and the set missing patterns and missing rates.Then iteratively trains a GAN was composed of two fully connected neural networks based on a game idea.By introducing a novel dynamic adaptive mechanism,this study can automatically identify the optimal number of iterations of the generator and discriminator during the model computation,and finally generates the complete traffic data matrix and repair the missing values.California PeMS and Guangzhou traffic speed datasets are used to complete the D-GAN model construction,and multiple evaluation metrics are employed to assess the repair performance of D-GAN.Experimental results show that the repair accuracy of D-GAN is higher for random missing patterns compared with non-random missing patterns;and the repair accuracy of D-GAN degrades accelerated with increasing missing rates.However,the repair performance of D-GAN outperforms the baseline models(e.g.,BGCP,prophet-RF,and GAIN)under various missing conditions.