In the global effort to promote green energy policies, understanding and optimizing boiler combustion processes in coal-fired power plants is crucial. During unit start-ups, shutdowns, and load deep peak regulation, significant energy-saving potential can be harnessed in boilers. This paper focuses on a 600MW supercritical coal-fired power unit and presents an improved Least Squares Support Vector Machine (LS-SVM) model with refined initial parameters. By combining the improved LS-SVM with Affinity Propagation (AP) clustering, a combustion efficiency model for boilers is constructed. The experimental results demonstrate that the AP-based improved LS-SVM model not only reduces computational complexity and training time but also enhances predictive accuracy and generalization performance.