Active safety correction of a power system is of great importance in ensuring the safe operation of a power grid. The traditional active safety correction method cannot comprehensively consider the system power flow distribution state and the adjustment performance of the units, and has difficulties in practical application because of the low solution efficiency, the adjustment involved in many units, and the need for repeated adjustment. Therefore, an active safety correction strategy based on the deep Q network (DQN) by using a deep reinforcement learning algorithm is proposed. First, a system active safety correction model is established; secondly, convolutional neural networks (CNN) are used to explore the deep features of the grid operation state. The DQN algorithm is used to construct a mapping model of the combination of power grid operation state and optimal adjustment unit through the mechanism of "state-action" and the medium of "reward", and the adjustment unit is determined. Finally, the adjustment quantity is calculated according to the sensitivity of overload line to the adjusting unit. The validation results of the IEEE39-bus system show that the active safety correction strategy proposed can comprehensively consider the overall situation of system power flow distribution and unit regulation performance when dealing with multi-line overload, and effectively eliminate line overload. This work is supported by the Science and Technology Project of State Grid Corporation of China (No. SGTYHT/17-JS-199).