In this paper

an optimal secondary control strategy is proposed for islanded AC microgrids considering communication time-delays. The proposed method is designed based on the data-driven principle

which consists of an offline training phase and online application phase. For offline training

each control agent is formulated by a deep neural network (DNN) and trained based on a multi-agent deep reinforcement learning (MA-DRL) framework. A deep deterministic policy gradient (DDPG) algorithm is improved and applied to search for an optimal policy of the secondary control

where a global cost function is developed to evaluate the overall control performance. In addition

the communication time-delay is introduced in the system to enrich training scenarios

which aims to solve the time-delay problem in the secondary control. For the online stage

each controller is deployed in a distributed way which only requires local and neighboring information for each DG. Based on this

the well-trained controllers can provide optimal solutions under load variations

and communication time-delays for online applications. Several case studies are conducted to validate the feasibility and stability of the proposed secondary control.