The wake effect is the major obstacle to reaching the maximum power generation for wind farms, since choosing the suitable wake model that satisfies both computational cost and accuracy is a difficult task. Deep Reinforcement Learning (DRL) is a powerful data-driven method that can learn the optimal control policy without modeling the environment. However, the “trial and error” mechanism of DRL may cause high costs during the learning process. To address this issue, we propose an ensemble-based DRL wind farm control framework. Under this framework, a new algorithm called Actor Bagging Deep Deterministic Policy Gradient (AB-DDPG) is proposed, which combines the actor-network bagging method with the Deep Deterministic Policy Gradient. The gradient of the proposed method is proved to be consistent with the DDPG method. The experiment results in WFSim show that AB-DDPG can learn the optimal control policy with lower learning cost and a more robust learning process.