Abstract: Dynamic modeling of selective catalytic reduction（SCR） denitrification system is of great significance for optimization of ammonia injection, reducing of the NOx emission and ammonia slip. The modeling methods based on machine learning have advantages in accuracy, but most methods lack self-learning or adaptive strategy, and are difficult to keep good performance in long-term operation. In order to adapt to the dynamic characteristics of SCR denitrification system caused by load and coal quality change, selective ensemble model library was proposed including a construction method of library based on time-phased data, a combined strategy based on selective real-time error weighting method and an updating strategy based on model evaluation method. The model was trained, tested and verified by a 50-day operation data of SCR system in a 660 MW coal-fired boiler, and compared with the traditional adaptive model. When the traditional model failed, the proposed model can still keep high prediction accuracy. The results show that selective ensemble model library performs better in prediction accuracy, robustness and stability.