• A CC-LFC method is proposed to balance the interests of operators. • A LSCMA-DMPG is proposed to achieve multi-task collaborative learning. • The proposed method can effectively reduce the power fluctuation. Traditional centralized load frequency control (LFC) is vulnerable to power fluctuations in tie-line power due to the conflicting objectives of multiple area controllers and distributors in an isolated multi-area microgrid. To address these problems, a cuttlefish-like cooperative load frequency control (CC-LFC) method is proposed. This AI method imitates the distributed neural network structure of cuttlefish in that it equates the controllers and power distributors of each area as agents in a multi-area microgrid. In online applications, a joint global optimization decision can be obtained from the grid areas without engaging in extensive intercommunication. In addition, this paper proposes a large-scale counterfactual multiagent deep meta-policy gradient (LSCMA-DMPG), which combines centralized training with decentralized execution in a large-scale learning framework. It employs meta-reinforcement learning to realize multitask collaborative learning, which improves the robustness and quality of the obtained CC-LFC policies. The real-time experiments and simulations for a four-area LFC model of Sansha Island in the China Southern Grid (CSG) demonstrate the superior qualities of the proposed method.
Liu et al. (Sat,) studied this question.