Abstract: Aiming at the problem of improving the short-term load forecasting accuracy and forecasting stability of the power system, a new type of cycle-consistent generative adversarial network(CycleGAN) is proposed. The generator and discriminator are gated recurrent unit(GRU) and temporal convolutional network(TCN), respectively. The generator uses a gated recurrent unit neural network, which is better adapted to the timing prediction task and solves the model gradient problem. The discriminator model uses a temporal convolutional neural network, which has better results in capturing long-term dependencies in the data of the temporal task and is more effective in identifying the differences between the fake samples generated by the generator and the real samples. Meanwhile, a cycle consistency loss function is introduced in CycleGAN that allows the model to learn the prediction laws more fully during the training process. Through the example experiments, it is proved that the proposed new model has better prediction accuracy and stability.