Abstract: To address the difficulty in dynamic modeling of power-plant condensers under unsteady conditions such as load variations and circulating-water pump switching—caused by thermal inertia and fluid transport delay—a transfer-learning-based Transformer dynamic modeling method (TL-Transformer) is proposed. The method extracts multivariable long-term temporal features with the self-attention mechanism of Transformer, pretrains on pre-maintenance historical operating data, and fine-tunes the model with a small number of post-maintenance samples, thereby alleviating the problem of insufficient data in dynamic processes. Validation on a dual-pressure condenser of a 1000 MW coal-fired unit shows that the R² values of temperature and pressure output parameters on the target-domain test set are no lower than 0.98, andthe root mean square errors (RMSEs)of the dual-pressure condenser pressures are 0.03 and 0.04 kPa. In tests under actual circulating-water-flow step-change conditions and load-variation conditions, the R² values of all output parameters are no lower than 0.97. It can be seen that this method can still maintain high prediction accuracy under limited samples or shifted operating conditions, and has good model adaptability.