Dynamic Economic Dispatch of Power System Based on Generative Adversarial Imitation Learning

System operation is faced with challenges due to its inherent volatility, intermittence and randomness of new energy. How to develop a dispatch scheme to handle the uncertainties of renewable energy is an urgent problem. This paper proposes a dynamic economic scheduling model based on the generative adversarial imitation learning. A generator network is constructed and the scheduling strategy is given by its observation on the system state. Inspired by the objective function of Proximal Policy Optimization, the loss function of generator network is creatively constructed, and the network parameters are updated by reverse transmission to optimize the scheduling strategy. Then, based on the perfect scheduling idea of American power market, this paper can calculate the ideal scheduling scheme offline and use it as an expert strategy to guide the learning of generator network. Secondly, the discriminant network is constructed to identify the generation strategy and the perfect scheduling strategy and output discrimination results to assist generator network in updating itself. In the game confrontation between the generator and the discriminator, both of them reach the Nash equilibrium state. When put into online application, the output plan of thermal power units considering economy and uncertainty can be arranged according to the new energy and load forecasting data. Finally, an example is given to verify the effectiveness of the proposed model. The model established in this paper does not need to model the random variables of the new energy. Under the guidance of perfect scheduling strategy, the end-to-end strategy learning can be realized the algorithm in this paper converges quickly in off-line training and has high decision efficiency in online application, which provides the objective and effective decision basis for the scheduling department. © 2022 Power System Technology Press. All rights reserved.