Abstract
This paper considers the problem of short-term estimation of the link flow on an isolated intersection by means of artificial intelligence (AI). The problem is formulated as a supervised learning regression task. Several architectures of neural networks are presented to predict the link flow: recurrent neural network (RNN), convolutional neural network (CNN), and a combination of the two, with RNN treating the time dependencies and CNN extracting the spatial properties from traffic. The models are trained on synthetic data that are generated by a simulator. These models are compared with a fully-connected network. Numerical experiments are carried out on CPU and GPU. They show that deep learning models with special architecture can estimate the link flow much better than a traditional neural network on a single intersection. The obtained link flow estimates can be used in different traffic control algorithms in order to minimize the transport network congestion.