In cognitive radio network(CRN), power control often faces the complicated iterations and large calculations, resulting in poor real time performance of the system. In this paper, a deep learning-based power control is proposed for CRNs, where the secondary users (SUs) can share the same channel of primary users (PUs) without causing excessive interference to the communication of PU. In the novel scheme, the DNN model is used to treat the input and output of the power control algorithm as unknown non-linear mappings and fit them, which determines the proportion of transmit power allocated to each SU, considering the interference caused to the PU. With this scheme, the maximization of the SUs sum-rate can be achieved. Furthermore, due to some errors in the practical samples of channel information, an auto-encoder is used to compress the channel coefficient through an encoder and reconstruct them through a decoder before DNN training. The simulations results show that the power control method using a combination of auto-encoding and DNN can improve the real-time performance of the system. And the sum-rate of SU is improved while the interference caused to the PU can be regulated even with the inaccurate channel information.