Nitrogen oxides (NOx) is one of the important pollutants emitted from cement plants. However, due to the harsh environment and the complex reactions during the cement clinker calcination, the NOx concentration is difficult to be measured accurately. In this paper, we introduce a modular echo state network (MESN) specifically designed for the real-time prediction of NOx concentration. First, the input features for our proposed model are chosen using a mutual information feature selection method. Then, an MESN with small-world features is designed to predict the NOx concentration during the cement clinker calcination. To validate the efficacy of this predictive model, we employ a dataset stemming from a real-world cement clinker calcination system. Experiments show that this method outperforms several comparative models.