An Improved Generative Adversarial Network for Extreme Scenarios Generation

When facing extreme climate, the high volatility and intermittency of renewable energy output mainly from wind and photovoltaic (PV) bring great challenges to the operation and scheduling of power grids. In order to accurately characterize the uncertainty of the output power of wind and PV in extreme climate, this paper proposed a method for generating extreme scenarios of wind-PV joint output based on multi-channel convolutional generation adversarial network. First, the wind and PV historical output datasets are divided evenly according to the wind and PV output size, and used as the input to the multi-channel convolutional generative adversarial network. Second, the trained generator is utilized to expand the scenarios of wind-PV power output to solve the problem of insufficient samples in historical power output. Then, the extreme degree of the generated scenarios is measured and ranked, and the more extreme scenarios are selected to replace the ordinary scenarios in the original scenarios, thus realizing the transfer of the training dataset. Finally, based on the training dataset after many iterations, a large number of real scenic joint output extreme scenarios are generated. Example analysis shows that the proposed method can not only generate scenarios that are similar to the historical scenarios, but also accurately capture the correlation between wind and PV, and generate a large number of different types of extreme scenarios of wind-PV joint output.