In the communication systems of intelligent micro-grids, the intrinsic security of the physical layer is insufficient, with high computational complexity and poor robustness. To address these issues, a beamforming algorithm based on deep learning and successive convex approximation (DL-SCA) is proposed. Initially, under the premise that the norm of channel estimation error is bounded, a secrecy rate maximization problem is formulated. A deep neural network is employed to learn the intricate mapping between channel state information (CSI) and the optimal beamforming matrix, facilitating the extraction of profound channel features. Subsequently, building upon the initial solution obtained, the beamforming precoding optimization problem is further solved using the successive convex approximation (SCA) algorithm. Ultimately, through the Monte Carlo experiment based on imperfect CSI, the security performance and convergence speed of different algorithms are compared. In comparison with the zero forcing (ZF) beamforming algorithm, there is an approximate 20% increase in the system secrecy rate; and when compared with the traditional SCA algorithm, there is a reduction of about 10% in the execution time. This substantiates that the DL-SCA algorithm can effectively enhance the intrinsic security performance of the micro-grid, with high robustness and rapid convergence.