In order to improve the prediction accuracy of nitrogen oxides (NOx) mass concentration at the inlet of denitration system, a combination algorithm prediction model of principal component analysis (PCA) and an extreme learning machine (ELM) with multi-layer self-coding structure based on mutual information is proposed. The selection of input variables is improved, and the network structure of the prediction algorithm is optimized by adding the historical NOx mass concentration. The experimental results show that compared with other prediction algorithm models, the proposed model has higher prediction efficiency and higher prediction accuracy under different working conditions, and shows good anti-noise ability and generalization ability.