Equivalent modeling method for real time active power dispatch of virtual power plant based on deep reinforcement learning under privacy protection

With the rapid development of distributed generation, virtual power plant (VPP) technology has been proposed and become an important technical means to solve a large number of grid connected scheduling problems of distributed generation. The VPP is usually composed of distributed photovoltaic, micro gas turbine, energy storage and other different types of distributed generation. Its output characteristics and cost characteristics are jointly determined by multiple distributed generation, and are time-varying. The equivalent modelling method for VPP to take part in the real-time active power dispatch of power grid under privacy protection is studied in this paper. Firstly, the internal real-time optimal dispatching model of VPP in response to the real-time dispatching instructions of grid active power is established; Then, an equivalent modeling method of VPP participating in real-time active power dispatching based on reinforcement learning under privacy protection is proposed to obtain the mapping model between the real-time active power dispatching instructions and the cost of VPP dispatching without displaying physical model information of VPP, and then a A3C deep reinforcement learning algorithm based solution method is proposed for VPP ; Finally, the feasibility and effectiveness of the proposed VPP real-time active power scheduling equivalent model based on deep reinforcement learning are verified by comparing with the VPP real-time optimal scheduling model based on physical model information, which provides a scientific foundation for the global dispatching optimization of power grid under privacy protection.