Abstract: In response to the problem that the strong randomness and intermittency of the new energy power output affects the power balance of distribution network, we propose a two-layer power allocation strategy combining modified DDQN strategy incorporating multi-step greedy policy with consensus protocol. The proposed method can adaptively optimize the output of each unit in the distribution network under the fluctuation of source and load, and ensure the rapidity and economy of power regulation. Firstly, a hierarchical distributed power allocation framework is proposed based on the classification of "resource clusters", and the power allocation of smart distribution network is decomposed into optimization allocation models of the coordinated scheduling layer and the autonomous layer. Secondly, for the coordinated scheduling layer, we apply the modified DDQN strategy with a multi-step greed policy to achieve power allocation among "resource cluster"; meanwhile, for the autonomous layer, we use a dynamic iterative power allocation method with cost micro-increment as the consistent state variable. Finally, the simulations on a typical smart distribution network example show that the proposed two-layer power allocation strategy can be adopted to solve the optimal power allocation problem under the new energy fluctuation. Compared with various methods, the proposed method can achieve a faster convergence rate and lower regulation cost.