Abstract: Seismic disasters can severely impact the aseismic performance of building structures; assessing the aseismic performance of structures post-earthquake is a crucial step in mitigating losses caused by seismic events. This study proposes a composite feature intelligent prediction method based on a deep learning model to accurately and efficiently predict the secondary aseismic performance of post-earthquake damaged structures. A large-scale two-stage time-history analysis was conducted on a ten-storey steel-concrete composite frame structure based on a three-dimensional refined fiber beam model, constructing a composite feature dataset that includes the state of seismic damage and secondary seismic input. A composite feature deep learning surrogate model was established based on multi-layer perceptrons and long short-term memory neural networks, capable of accurately predicting structural secondary post-earthquake response indicators according to the data of structural monitorable indices and seismic wave sequences during the first earthquake. Comparative experiments indicate that the method proposed improves the average accuracy by over 24.8%, compared to mainstream machine learning algorithms such as Support Vector Machines and XGBoost (eXtreme Gradient Boosting), with an R² value reaching above 0.91. The excellent performance of this method makes it a core intelligent model for the evaluation of the aseismic performance of multiple structures in subsequent engineering applications.