The Coupling Cross Section(CCS) of aperture is an important parameter to evaluate the effect of aperture penetration. Using BP neural network to predict CCS has a much higher prediction speed than full-wave analysis and better accuracy than traditional formula methods. This paper focuses on the prediction model which can be applied to multi-shape aperture array. Three neural network models are proposed to predict the CCS of aperture array, including one traditional single-stage model and two two-stage models. Taking the regular hexagonal aperture array as an example, the performance of the three models is compared. These results show that the double-level model with the most prior information performs the best. The Root Mean Square Error(RMSE) of the CCS prediction for the regular hexagonal aperture array by this model is 0.017 2, and the coefficient of determination(R) is 0.999 1. When this model is transferred, it can predict the CCS of circular and square aperture arrays, with an average relative error of 1.94% for the samples. The prediction results confirm the precision, efficiency, and universality of the model.