After a fire, the exposure temperature and cooling method have lasting effects on the mechanical performance of structural steels, which have an influence on the bearing performance of steel members. To evaluate the residual service capacity of steel structures after a fire, it is meaningful to perform a quantitative analysis of the effects of differences in the stress–strain properties on the bearing performance of steel members. To address this issue, a long short–term memory (LSTM) network was constructed to accurately predict the post–fire load–displacement curves to facilitate the behavioural study of numerous samples of the Q690 high–strength steel (HSS) plater girder under shear. Then, the complex effects of the exposure temperature and cooling method on the bearing capacity of the HSS plater girder were quantified using four performance indexes extracted from the load–displacement curves for three different limit states. The LSTM model was trained and tested on the comprehensively validated finite–element analysis results. For better prediction accuracy of the load–displacement curves, Bayesian optimisation was employed for hyperparameter tuning, followed by ten–fold cross–validation. The developed LSTM model was proven to be accurate and efficient in reproducing the load–displacement curves of plate girders.