Abstract: As a core power generation equipment in modern power systems, ultra-supercritical boilers play a significant role in national energy strategies due to their high efficiency, environmental benefits, and stable operation. However, traditional control methods struggle to achieve global optimization of their performance. This paper proposes a multi-objective adaptive control strategy for boiler combustion based on deep learning. The strategy designs a deep prediction model incorporating the Transformer architecture and integrates a soft actor-critic (SAC) deep reinforcement learning decision model. By constructing a dynamically weighted reward function, it achieves collaborative optimization of multiple objectives, including improved combustion efficiency, reduced emissions of pollutants (NO_x, CO), and enhanced operational stability. Furthermore, an adaptive mechanism combining online learning and offline pre-training ensures the control system's robustness and continuous optimization capability in the face of equipment aging, fuel variations, and external disturbances. Experimental results demonstrate that the proposed adaptive control strategy significantly improves performance across all aspects, enhancing the system's response speed and robustness to load disturbances.