Abstract: As the ultra-low emission requirement for thermal power units in China is that the concentration of sulfur dioxide emissions should be less than 35 mg/m~3,accurate prediction and control of the SO₂ emission concentration is of great significance for the environmental protection operation of thermal power units.In order to deal with the prediction of SO₂ emission concentration in circulating fluidized beds,a prediction model of SO₂ emission concentration based on deep belief network（DBN）is established by introducing deep machine learning method.Firstly,the operational variables affecting the concentration of SO₂ emissions are determined as model inputs through mechanism analysis;secondly,the DBN network is used to extract the deep features of the model inputs,and ELM is used as the regressor to establish the prediction model;finally,the DBN-ELM model is compared with three prediction models of SO₂ emission concentration.The results show that the root-mean-square deviation and average absolute error of the model are 175.3mg/m~3 and 117.6 mg/m~3 respectively.This prediction accuracy is much higher than that of the other three comparison models.And it has more application value in practical engineering.