Abstract: In order to simplify the complicated process of data mining for power plants by various algorithms, and to solve the problem of incomplete selection of operating conditions in the combustion optimization adjustment test, a combustion optimization method based on KPCA-Kmeans++ data mining was designed. Through the mining of massive historical operating data, the correlation between the main operating parameters and performance indicators of the boiler under steady-state operating conditions in different operating conditions was found, and the operating target values of the parameters corresponding to the boiler’s high-efficiency and low-pollution operation were obtained according to the correlation law. By performing dimensionality reduction clustering on a large amount of data mining results, an intuitive and streamlined combustion control law was obtained, so as to guide combustion optimization. Regarding the grinding method as the working condition division standard, the kernel principal component analysis（KPCA） was used to reduce the dimensionality of the parameters contained in the historical data, thus improving the performance of Kmeans++ algorithm in cluster analysis of massive data under the same working condition. Under each working condition, three combustion optimization modes were determined. Comparing the secondary air optimization effect of Unit 5 in a certain power plant, the NOx concentration at the inlet of the SCR（Selective Catalytic Reduction） device after the optimization decreased by 29.2mg/m~3 on average, and the boiler efficiency increased by 0.11% on average.