Abstract: The real-time prediction of NOx emissions is of great significance for pollutant emission control and unit operation of coal-fired power plants. Aiming at dealing with the large time delay and strong nonlinear characteristics of the combustion process, a dynamic correction prediction model considering the time delay was proposed. First, the maximum information coefficient(MIC) was used to calculate the delay time between related parameters and NO_x emissions, and the modeling data set was reconstructed. Then, an adaptive feature selection algorithm based on Lasso and ReliefF was constructed to filter out the high correlation with NO_x emissions. Finally, an extreme learning machine(ELM) model combined with error correction was established to achieve the purpose of dynamically predicting the concentration of nitrogen oxides. ExpNO_xerimental results based on actual data show that the same variable has different delay time under load conditions such as rising, falling, and steady; and there are differences in model characteristic variables under different load conditions. The dynamic error correction strategies effectively improve modeling accuracy. The prediction error of the algorithm under different working conditions is less than 2%, which can accurately predict the NO_x concentration at the combustion outlet, and provide guidance for NO_x emission monitoring and combustion process optimization.