In order to simulate driver's memory effects and fuzzy perception characteristics， a deep learning car-following model was designed based on fuzzy perception time window. Taking 3 s continuous speed， leading-following car speed difference and headway distance as model inputs with a minimum time interval of 0.2 s，the driving memory was simulated. A gated recurrent unit （GRU） network with a single layer of 32 output dimensions could fit the actual data well by training multiple groups of deep learning car-following models. In each input time series data of the model， part of the real car-following state value was replaced by the predicted value of the model as the driver’s estimation of the scenario， that is， fuzzy perception. The experiment results show that different fuzzy perceptions to the same scenario can produce different car following behaviors， and the heterogeneity of driving behaviors can be simulated， which provide a method for heterogeneous traffic behavior simulation.