Conventional communication systems are generally based on modular design, since the modules are optimized separately, the system can not achieve the optimal performance. An end-to-end communication system model can be implemented by deep learning, which can improve the transmission performance. However, the channel environment is changeable and unknown, which make the optimization of the end-to-end communication system impossible. Recently, the birth of the deep convolutional generative adversarial networks (DCGAN) can simulate unknown channels and solve the optimization problem of end-to-end systems. Then, the DCGAN has poor training stability, and the problems of over-fitting and gradient disappearance caused by it will lead to performance degradation. In this paper, we propose a residual-based DCGAN model to alleviate these problems. Specifically, we introduce a residual block structure, which effectively alleviates the over-fitting problem of the gradient. In addition, we introduce the Wasserstein distance to measure the difference between the generated data and the real data distribution, and further solve the problem of model training instability. Simulation results show that our proposed Residual DCGAN-based model effectively improves the block error rate (BLER) performance compared with traditional methods.