In order to reveal the traffic decay patterns in terminal areas under convective weather conditions, and further enhance existing models’ ability to capture complex dependencies in long sequences, a model integrating a multi-head self-attention mechanism was developed for predicting the degree of terminal area traffic impact. This approach first established an airspace block partitioning framework, incorporating both weather hazard index and airspace blockage rate as dual-dimensional features. Subsequently, a long short-term memory (LSTM) network improved with a multi-head self-attention mechanism (MHSA) was constructed. The model was validated using actual operational data from the Guangzhou terminal area. The results show that the model predicts best when the input sequence length is 12. Compared with classical models such as support vector regression (SVR), random forest (RF), recurrent neural network (RNN), long short-term memory network (LSTM), and gated recurrent unit (GRU), the mean absolute error (MAE) is reduced by 61.97%, 53.45%, 43.75%, 10.00%, and 20.59%, respectively. The root mean square error (RMSE) is reduced by 58.89%, 53.16%, 41.27%, 13.95%, and 15.91%, respectively. Additionally, the coefficient of determination (R²) is improved by 48.10%, 30.18%, 15.13%, 2.30%, and 2.97%, respectively. It is concluded that the proposed model can effectively capture the complex relationship between meteorological factors and traffic flow decay. The traffic decay effect is most significant when convective weather is located in the airspace adjacent to the runways and in the key arrival/departure areas in the northern part of the terminal area. Therefore, it is suggested that air traffic controllers enhance real-time monitoring and early warnings for these critical airspace sectors.