Modeling traffic in the signal controlled road network is critical to urban traffic simulation and management. However, Existing traffic simulation methods often have difficulty in balancing computational efficiency and simulation accuracy. In this paper, We propose a macroscopic traffic simulation framework for machine learning methods used to simulate traffic flow changes in road networks under signal control. The computational efficiency is ensured by the macroscopic simulation granularity, while the accuracy of the simulation is maintained by extracting the traffic variation patterns through machine learning methods. We design a deep learning simulation model integrated with graph neural networks called SimNET, which accurately mines the macroscopic characteristic changes of traffic flow from data. We compare SimNET, other state of art deep learning methods for Spatio-temporal problems and a traditional traffic simulation model on real world and generated datasets. Our model demonstrates better accuracy while being more convenient in simulation setup compared to traditional macro models. Furthermore, our model can realistically reproduce some typical traffic phenomena under a signalized road network.