Model-Free Economic Dispatch for Virtual Power Plants: An Adversarial Safe Reinforcement Learning Approach

To address the model inaccuracy and uncertainty of virtual power plants (VPPs), a model-free economic dispatch approach for multiple VPPs is studied in this article, which does not rely on an accurate environmental model. An adversarial safe reinforcement learning approach is proposed, which promotes the safety of the actions and makes the model robust to deviations between the training and testing environments. Moreover, a two-stage reinforcement learning framework is formulated based on the proposed algorithm. The dispatch policy is pretrained in the simulator and then fine-tuned in the real-world environment. The numerical simulations illustrate that the proposed approach is adaptive to the deviation between the training and testing environments, and it provides higher robustness to the noise of the network parameters and uncertainty of the VPPs’ power outputs. The scalability and superiority of the proposed approach are verified by comparing it with existing methods.