基于时间卷积神经网络的短时交通流预测算法

Short-term Traffic Flow Prediction Based on Temporal Convolutional Networks

短时交通流预测是智能交通系统实现交通控制与交通诱导的关键所在。传统一维卷积神经网络(CNN)在短时交通流预测上难以获取长时记忆,同时存在信息泄露的问题。文中提出扩张-因果卷积神经网络(DCFCN),引入扩张卷积来增加感受野大小,获取序列的长时记忆;同时,引入因果卷积来解决信息泄露问题。DCFCN由6层卷积层堆叠而成,每层通过Padding的方式实现因果卷积,扩张系数逐层呈指数增长。实验结果表明,文中提出的DCFCN在短时交通流预测上优于其他对比模型,且在GPU上计算效率明显提升。

Short-term traffic flow prediction is the key for Intelligent Transport Systems(ITS)to achieve traffic control and traffic guidance.Traditional convolutional neural networks are difficult to obtain long-term memory in short-term traffic flow prediction,and there is information“leakage”from future to past.Dilated-Causal Fully Convolutional Networks(DCFCN),which introduces dilated convolutions to increase the reception field and obtain long term memory of sequences,was proposed in this study.At the same time,the causal convolutional was introduced into DCFCN to solve the problem of information“leakage”.DCFCN is composed of 6 convolutional la-yers,each layer achieves causal convolution by padding,and the dilated increases exponentially with the depth of the network.Experiments show that the proposed DCFCN is superior to other comparison models in short-term tra-ffic flow prediction,and the computational efficiency on the GPU is significantly improved.