This paper employs a Transformer model to address the challenge of real-time temperature prediction for optical elements under multi-physical field coupling. Experimental results demonstrate that, compared to empirical model methods, the Transformer model achieves reductions of 70% and 32% in root mean square error (RMSE) and mean absolute error (MAE), respectively. When compared to LSTM-based methods, the Transformer model reduces RMSE and MAE by 66% and 23%, respectively. Additionally, the coefficient of determination (R²) of the Transformer model approaches 1 more closely, indicating higher consistency between predicted and actual values.