Using a Separable Convolutional Neural Network for Large-Scale Transportation Network Speed Prediction

Abstract

This paper proposes the reduction of the convergence time on a Convolutional Neural Network (CNN) method for traffic speed prediction, without reducing the performance of speed prediction method. The proposed method contains two procedures: The first one is to convert the traffic network data to images; in this case the speed variable will be transformed. The second step of the procedure presents a modification of the CNN method for speed prediction in which a separable convolution is used to reduce the number of parameters. This separable convolution helps to reducing the convergence time of speed predictions for large-scale transportation network. The proposal is evaluated with real data from the Caltrans Performance Measurement System (PeMS), obtained through sensors. The results show that Separable Convolutional Neural Network (SCNN) reduces convergence time of CNN method without losing the performance of the predictions of traffic speed in a large-scale transportation network.