Aiming at the problem of how to maintain stable and secure operation of the smart grid after a false injection attack, the tolerant intrusion control of the system after a false data injection attack is investigated. First, since there is no publicly available FDIA dataset, two datasets are generated: one contains random single-node attacks and the other contains random multi-node attacks. Second, combining AE and Generative Adversarial Network (GAN), and introducing Wasserstein distance and gradient penalty term to avoid the problems of gradient vanishing and pattern collapse that GAN is prone to in adversarial training; at the same time, adversarial training is used to achieve Nash equilibrium, learn spatial relationship between data, and complete the data reconstruction; the method does not need labeled data from measurement instruments, and only requires attack-free historical state data to complete the training, thus it is unsupervised learning. Finally, the simulation results show that the adopted recovery scheme can accurately restore the contaminated state quantities to their pre-attack values, and has high recovery accuracy for FDIA contamination.