Abstract: Near space refers to the complex transition region between the Earth's atmosphere and space, with its height between 20-100 km, including the stratosphere, the mesosphere and the lower thermosphere which are important components of the Earth's atmospheric system. The near space is affected not only by small-scale disturbances in the troposphere but also by planetary waves in the middle and upper atmosphere. There are also atmospheric dynamic processes in the region and physical and chemical interactions between layers of the atmosphere, which make the study of near space difficult. Scholars at home and abroad have done a lot of research on the application of machine learning and deep learning algorithms in the field of space physics, and achieved good results. However, due to the complexity and variability of atmospheric environment, there is little research on using deep learning algorithm to forecast environmental temperature in near space. The study of temperature forecasting in near space is of great scientific significance for understanding atmospheric dynamics, cross-layer coupling and analyzing atmospheric fluctuations. In this paper, CNN-LSTM algorithm is used to forecast temperature in near space based on MERRA-2 datasets. The CNN part can effectively extract spatial features and the LSTM part can capture temporal dependencies. Its output window is 7 days, and the relative error is controlled within 2% compared with the original data. On this basis, we do some research on the sensitivity of the algorithm in season, altitude and latitude. The results show that in summer and autumn of the Northern Hemisphere, the forecast results are better, and the accuracy of the forecast model is more reliable at low latitudes.