Combustion optimization is one of the effective techniques to enhance boiler efficiency and reduce nitrogen oxide (NOx) emissions from coal-fired boilers. A precise NOx emission model and a boiler efficiency model are the basis of implementing real-time combustion optimization and are required. In this study, to obtain very precise models and make full use of abundant real-time operational data easily collected from supervisory information systems (SIS), a novel deep learning algorithm called a deep bidirectional learning machine (DBLM) is proposed to set up the correlation between NOx, emissions, boiler efficiency, and operational parameters from a 300 MW circulating fluidized bed boiler (CFBB). Experimental results indicate that, in comparison to other recently published state-of-the-art modeling methods, the models built by DBLM could own much better generalization performance and high repeatability, which may be a better choice for modeling NOx emissions and efficiency in achieving boiler combustion optimization and improving power plant performance.