In order to address the issue of reduced prediction accuracy arising from the redundant computation of dynamic parameter associations across multiple spatial dimensions resulting from the inherent token calculation strategy of Transformer based algorithms in short-term flight trajectory prediction, an in-depth analysis was conducted of the spatial-temporal characteristics of trajectory data and the computational logic of Transformer-like algorithms. Building upon a symmetrical multi-scale backbone architecture, a multi-dimensional splitting module, a local feature extraction module, and a global pattern fitting module were innovatively introduced, thereby a short-term flight trajectory prediction algorithm was proposed based on fusing global and local features in multiple dimensions. This algorithm mapped the short-term trajectory prediction task into a pattern regression problem involving an equal-length known trajectory sequence to an unknown sequence, effectively capturing the local spatial-temporal dynamic features and global spatial-temporal dependencies of the data in each dimension, and thereby further enhancing the model’s adaptability and robustness to complex trajectory patterns. Utilizing the massive real-world trajectory data from the OpenSky dataset, comparative experiments and ablation experiments were conducted. The results demonstrate that, compared with the latest Transformer-based baseline algorithm, this method achieves performance improvements of 10.92%, 14.39%, and 14.96% in terms of mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), respectively. Furthermore, through module-by-module ablation verification, the accuracy of the algorithm gradually increases, fully demonstrating the effectiveness, advancement, and contribution of each module to the overall performance of the proposed algorithm.