To address the issues of low computational efficiency and insufficient modeling of multi-altitude physical characteristics in traditional numerical prediction methods for the aerospace transition zone, a HAFNO (height-aware Fourier neural operator) model was proposed. A height-aware weighting mechanism was introduced to adaptively distinguish physical features across different altitude layers while maintaining O(n log n) computational complexity. Additionally, a multi-altitude coupled preprocessing framework and an adaptive loss function incorporating spatial gradient constraints were constructed. Experimental results based on MERRA-2 data demonstrate that HAFNO outperforms benchmark models such as ConvLSTM and DeepONet in the 50~70 km altitude range. The average RMSE (root mean square error) is reduced by 12.8% compared to the standard FNO, with a maximum correlation coefficient of 0.994, providing an efficient deep learning technical approach for environmental forecasting in the aerospace transition zone.