基于生成对抗网络的大规模路网交通流预测算法

对大规模路网交通流进行准确预测,能够应用于区域交通协同控制与管理,提高路网运行效率.针对如何高精度地拟合大规模路网交通流时空分布并对其进行准确预测,提出基于梯度惩罚的Wasserstein生成对抗网络(Wasserstein generative adversarial network with gradient penalty,WGAN-GP)的大规模路网交通流预测算法.根据大规模路网交通流数据特点,为了增加模型对时间相关性和远距离空间相关性特征的抽象能力,采用残差U型网络作为生成器来增加网络深度;采用多重判别器分别从时间和空间特征来对生成数据进行判别,从而提高判别器的判别能力.所提算法能够解决判别型深度学习模型仅能针对路网整体误差最小化,而忽略各交通流观测点预测误差最小化原则的问题,能够更好地满足现实交通场景需求.实验结果表明,所提算法能够有效地学习路网交通流数据内部多因素耦合特性,具有更高的预测精度.

Traffic flow forecasting algorithm for large-scale road network based on GAN

Accurate traffic flow forecasting can be applied to traffic control and management to improve operating efficiency of the lagre-scale road network. Aiming at how to fit the spatio-temporal distribution of traffic flow with high precision and accurate forecasting, an algorithm of lagre-scale road network traffic flow forecasting based on the Wasserstein generative adversarial network with gradient penalty is proposed. According to the characteristics of traffic flow data for large-scale road networks, the proposed algorithm uses the residual U-Net as a generator to increase the network depth for improving the ability of model to abstract the characteristics of temporal correlation and long-distance spatial correlation. The proposed algorithm can solve the problem that the discriminant deep learning models can only minimize the whole error of the road network while ignoring the error minimization of each observation point, then meet the demand of real traffic scenes better. Experimental results show that the proposed algorithm can learn the coupling characteristics of multi-factor inside the traffic flow data in lagre-scale road networks effectively and improve the prediction accuracy.