The strong stochastic disturbance caused by large-scale distributed energy access to power grids affects the security, stability and economic operations of the power grid. A novel multiple-step greedy policy based on the consensus Q-learning (MSGP-CQ) strategy is proposed in this paper, which is an automatic generation control (AGC) for distributed energy incorporating multiple-step greedy attribute and multiple-level allocation strategy. The convergence speed and learning efficiency in the MSGP algorithm are accelerated through the predictive multiple-step iteration updating in the proposed strategy, and the CQ algorithm is adopted with collaborative consensus and self-learning characteristics to enhance the adaptability of the power allocation strategy under the strong stochastic disturbances and obtain the total power commands in the power grid and the dynamic optimal allocations of the unit power. The simulations of the improved IEEE two-area load-frequency control (LFC) power system and the interconnected system model of intelligent distribution network (IDN) groups incorporating a large amount of distributed energy show that the proposed strategy can achieve the optimal coordinated control and power allocation in the power grid. The algorithm MSGP-CQ has stronger robustness and faster dynamic optimization speed and can reduce generation costs. Meanwhile it can also solve the strong stochastic disturbance caused by large-scale distributed energy access to the grid compared with some existing intelligent algorithms.