Typhoons are powerful tropical cyclones that often result in devastating secondary effects due to their high destructive potential. Therefore, accurate forecasting of typhoon paths is of great importance for timely and effective disaster warnings. This study proposes a typhoon prediction model based on Seq2Seq attention, which addresses the performance degradation issue in the encoder-decoder structure when predicting long sequences. The model effectively captures long-term dependencies, ensures convergence, mitigates the problem of gradient vanishing or exploding, and improves prediction accuracy. To evaluate the model, typhoon data from the China Typhoon Network spanning from 2000 to 2022 were used. The Seq2Seq attention model is compared with TCN, LSTM and Seq2Seq model to forecast the typhoon trajectory for a 12-hour period. The results indicate that the Seq2Seq attention model outperforms other prediction methods, demonstrating its effectiveness in this study based on superior forecasting accuracy.