Traffic Flow Model Informed Network for Fine-Grained Speed Estimation with Robustness Enhancement

Traffic state estimation (TSE) is a method for inferring the traffic state on roads by applying partially observed data, forming the basis for traffic management and control. To tackle the problems of observed data sparsity and noise and to finally reach a fine-grained speed estimation, this paper proposes a model-data-fusion driven TSE method, that is, traffic flow model informed deep learning. This method incorporates the macro traffic flow model velocity-LWR as a physical constraint to the deep learning neural network, which optimizes its gradient descent processes and eventually performs a fine-grained speed estimation. Furthermore, in order to deal with the noisy samples, we adopt Huber loss to enhance the robustness of the model. Two case studies are presented to reconstruct the velocity field under the NGSIM data set. The results show that the proposed methods can achieve better traffic state estimation than the baseline.