Abstract: Aiming at the problem that passive tracking of servo antennas cannot meet the performance requirements of spatial terahertz communication links caused by narrow terahertz beams, a pre-judgment method of terahertz beams based on optimized deep networks is proposed. Firstly, by analyzing the variation of the azimuth and pitch angle errors with the operation period and the initial time of extrapolation, the characteristic of the periodic divergence of the pointing errors of high-orbit satellites to low-orbit satellites is obtained. Then, the long-short-term memory network parameters are optimized by particle swarm optimization, the pointing error of the future time is predicted and corrected. In view of the poor global search ability of the particle swarm optimization and easy to fall into the local optimum, the inertia weight of the particle swarm optimization is dynamically adjusted to achieve the purpose of optimizing the long-short-term memory network. The simulation results show that the long-short-term memory network optimized by the improved particle swarm optimization can effectively predict the pointing error in the future, and the average absolute percentage error is reduced by 13.08% compared with the unimproved network in the same scenario.