Ground motion prediction equations (GMPEs) are crucial for the seismic hazard analysis of infrastructures. Currently, nearly all GMPEs are designed to predict only mainshock; in addition, they are generally based on a pre-assumed function form for data fitting. Historical earthquake records show that the mainshock (MS) is always followed by several aftershocks (ASs), aggravating the structural damage caused by the MS. The direct application of the traditional earthquake-oriented GMPE to predict ASs may not properly reflect their spectral characteristics and relationship with the MS. All pre-defined functions are generally low-order functions, i.e., they include limited number of variables. The newly emerged deep learning method is a powerful tool for revealing and presenting correlations among high-dimensional variables. Thus, the deep learning method was used as a GMPE to predict the spectral accelerations (Sa) of aftershocks in this study. Two popular networks (the deep neural network (DNN) and conditional generative adversarial network (CGAN)) were adopted to build the prediction model. Eight seismic variables and Sa of the mainshock at 21 periods were used as inputs for the deep learning model, and the Sa of the aftershocks at 21 periods were the outputs. A total of 503 sets of MS–AS records were used to develop the model, and the prediction results were compared with real data and that obtained by traditional GMPEs. The comparisons indicated that the deep learning model is a promising tool for predicting the Sa of aftershocks, and the CGAN model is slightly better than the DNN model because of the former's random nature in the generation of new data.