Accurate flight trajectory prediction is essential for identifying route deviations and easing air traffic management pressures. However, existing methods often have limitations, such as short prediction horizons, inconsistent performance, and low accuracy. To address these challenges, we propose a flight trajectory prediction method based on a TCN-SAM-Transformer model to enhance the accuracy of long-term predictions. The Temporal Convolutional Networks captures multi-scale time patterns of long series by expanding causal convolution, alleviating the gradient attenuation problem of traditional cyclic networks. The SAM module explicitly models the spatial topological relationship between the way points, and enhances the perception of spatial structure. To address the limitations of the small dataset and validate the proposed model, we built the dataset using 12-month historical flight trajectory data from real flights. Our approach demonstrates a better performance over existing methods, particularly in accuracy and consistency, highlighting its potential to enhance flight trajectory prediction.