Abstract: In order to enhance the accuracy of the combustion model in reversing the burnout of pulverized coal in the boiler, the author constructs a boiler combustion optimization model based on self - learning optimization. The model is achieved by combining the improved neural network model of genetic algorithm with the CO online monitoring system. Besides, a relationship between CO volume fraction and boiler thermal efficiency is established. The outlet oxygen, air distribution methods, and burnout air(SOFA air) are adjusted based on self - learning optimization results. It is found that adjusting the export oxygen content from 3.0% to 2.5% and 3.5% increases the boiler thermal efficiency by 0.53% and 0.49%, respectively. Adjusting the air distribution method of the boiler to waist reduction and positive tower air distribution resulted in an increase in the boiler′s thermal efficiency by 0.57% and 0.73%, respectively. The opening of the SOFA air distribution doors on both of A an B sides is adjusted from 87.4% to 86.7%, resulting in a 0.71% increase in boiler thermal efficiency, which reduces heat loss.