This study adopted transfer learning for the seismic response prediction of real structures using deep learning. This approach aims to address the challenge posed by the scarcity of observation data. Specifically, the deep learning model Explanianble Seismic Response Networks (ExSRNet) is pre-trained using computationally inexpensive seismic response data from Single-Degree-of-Freedom systems as the source domain. The acquired knowledge is then transferred to the learning process using observation data from an actual structure (a six-story reinforced concrete building) as the target domain. Verification results demonstrate that the pre-trained model exhibits improved training stability compared to the model without pre-training. Furthermore, its prediction accuracy on test data, evaluated by the coefficient of determination (R²) and the maximum response error, is significantly enhanced.