For the reduction of pollutant emissions from power plants, the ways to accurately estimate the emission value of nitrogen oxide (NOx) from gas turbines and identify its key influencing factors are crucial for effective optimization design. The accuracy and generalization performance of a single model can hardly meet the requirements of industrial applications because of variable operating conditions in the gas turbine operation process. The combined modeling of the partial least square (PLS) method and mutual information (MI) ensures the effectiveness and accuracy of the feature variable selection of NOx. Specifically, PLS is employed to determine the number of feature variables affecting the NOx of gas turbines, which can avoid subjective factors of variable selection and reduce dimensions. Then, mutual information (MI) is applied to select optimal feature variables. Different prediction models are used for simulation analysis, and single and combined feature selection is compared. The results show that in this study, the combined model PLS-MI can select more representative feature variables and can ensure the generalization accuracy of the prediction model, reducing the model complexity and providing a theoretical basis for optimal control of power plants with certain application prospects.