Aiming at the insufficient new energy consumption capacity in actual power grid operation, an intelligent generation method is proposed for operating mode samples of optimal new energy consumption based on generative adversarial networks (GAN) and reinforcement learning (RL). The method is used to establish generator and discriminator models through deep neural networks with a small sample set of actual power grid operation modes. And based on the objective function of GAN, the generator and discriminator can be trained for maximum and minimum adversarial training to generate false operating mode samples, distribution characteristics of which approach the real samples. During the training process of the generator, a reward and punishment feedback function is defined for new energy consumption to provide a quantitative good or bad response to the “generation action”. On this basis, the reinforcement learning method is introduced. Following the optimization principle of accumulating the most rewards, the generator parameters are continuously optimized in real-time by the reinforcement learning method with policy gradients. Operating mode samples are intelligently generated with optimal new energy consumption characteristics. The proposed method is tested in the actual power grid of a certain region. The results show that it can effectively generate operation mode samples that meet the optimal new energy consumption, and provide data support for improving the current operation scheduling decision of insufficient new energy consumption.