With the increasing complexity of wireless environment, scene, frequency band, antenna scale and other factors, it brings new challenges to wireless channel modeling. On the other hand, wireless communication solutions based on artificial intelligence have been constantly proposed in recent years, which are highly dependent on the quality and quantity of channel data. However, in the actual communication system, the acquisition of real channel data is restricted by the high cost of channel data acquisition. In this paper, an improved Wasserstein generative adversal network with gradient penalty (WGAN-GP) is proposed to solve the channel modeling and generation problem in MIMO communication system. The experimental results show that the improved WGAN-GP can generate fake channel data with the same distribution as the real channel data. In addition, the validity of the fake channel data is cross-verified by a channel feedback scheme based on deep learning.