Abstract：With the increased penetration rate of new energy year by year，it is difficult to accurately predict

the randomness and fluctuation characteristics of its output，causing a severe challenge to the operation，planning

and scheduling of electrical power system. Therefore，modeling for the uncertainty of new energy has attracted

more and more attention. To obtain the time sequence characteristics of new energy output scenario more

effectively，a new energy scenario generation method was proposed based on data drive，and combined selfattention mechanism with generative adversarial network discriminator with gradient penalty through applying the SA/WGAN model. Through building a deep learning model based on the combination of two models，effectively highlight the timing sequence characteristics of new energy output scenario and enhancing the nonlinear fitting capability in scenario generation. The example results show that，compared with the scenario generation results of original WGAN and WGAN-LSTM，the new energy generation scenario of proposed model can not only

effectively improve the accuracy，but also possess the advantages of stable WGAN-GP training results and quick

SA calculation speed，which can achieve a more efficient generation of scenarios that is close to the distribution of

real new energy scenario.

      Key words：unsupervised learning；self-attention（SA）；generative adversarial networks（GAN）；new energy；scenario generation