3D CNN-based Accurate Prediction for Large-scale Traffic Flow

Abstract

Traffic flow prediction is an important part of intelligent transportation system and is of great significance for the dispatching and management of urban traffic. Nowadays convolutional network with excellent spatial feature extraction capability is frequently applied in traffic flow prediction tasks. However, its performance is not as good as expected due to the inability to effectively deal with temporal features. In this study, a three-dimensional convolutional network (3D CNN) called TF-3DNet is proposed to achieve the accurate prediction of large-scale traffic flow. The key idea is to use 3D convolution kernel to simultaneously extract and fuse the spatio-temporal features in the traffic flow data. An effective 3D CNN model architecture is explored to ensure that time information is handled in all network layers as spatial information. Moreover, a missing value completion method is proposed to further improve the prediction performance of TF-3DNet model. The experimental results demonstrate the superior performance of the proposed model in traffic flow prediction task and missing traffic data completion.