Abstract: In order to realize the distributed active and reactive power coordinated optimal scheduling of networked microgrids, so as to improve the reliability of power supply and reduce the cost, a two-layer multi-agent reinforcement learning method was proposed to train the agents to interact with the environment and learn the optimal scheduling strategy. This method does not rely on an accurate networked microgrids model, and the two-layer multi-agent reinforcement learning algorithm trains continuous and discrete action agents respectively to adapt to the problem that continuous and discrete action devices need to be controlled in each sub-microgrid at the same time. In addition, considering that the topology transformation causes the change of optimization task and the existing agent group is not applicable, the applicable conditions of knowledge transfer were given and the method of knowledge transfer was adopted(the experience of the existing agent was used to train the new agent group), which avoids the ab initio initialization training and reduces the required calculation and time cost. Numerical experimental results show that the proposed method is effective in the distributed active and reactive power coordinated optimal scheduling of networked microgrids.