Abstract: In order to solve the problem of manpower and time cost consumption caused by strict non-convergence of power flow in large-scale power grid calculation, an automatic adjustment method of power flow convergence based on knowledge experience and deep reinforcement learning was proposed. Firstly, the knowledge experience of power flow convergence adjustment and the basic concepts and principles of deep reinforcement learning were introduced. Then we designed state space, action space and multiple rewards of reinforcement learning, as well as the framework of deep neural network. Next, by adding knowledge and experience into reinforcement learning, the search space was narrowed. And the process of human adjustment was simulated, which is balancing active power and then balancing reactive power. So that the search has directivity, and the power flow adjustment strategy was constructed. When balancing reactive power, Dijkstra algorithm was used to locate capacitors and reactors near the nodes with potential reactive power deficiency by means of optimal path. Finally, the CEPRI36-node system and the northeast power grid were used to verify the effectiveness of the method.