Abstract: In view of the problem that it is difficult to accurately predict the SO₂ emission mass concentration of thermal power units due to numerous influencing factors, a combined model named as improved INFO-Bi-LSTM model was proposed with the combination of improved weighted mean of vectors (INFO) algorithm and bi-directional long short term memory (Bi-LSTM) neural network. The high quality initial population was generated by adopting Circle chaotic mapping and reverse learning, while the ability of jumping from local optimal solution and global searching of INFO algorithm was improved with the application of adaptive t-distribution. Improved INFO-Bi-LSTM model and several other prediction models for a combined desulfurization process inside and outside the furnace were selected to predict the SO₂ emission concentrations under four typical conditions, after which, verifications and comparisons were conducted on the prediction results. Results show that, the optimization ability of INFO algorithm is improved, while improved INFO-Bi-LSTM model has a higher accuracy, and which is more suitable for the application of SO₂ mass concentration prediction. This can provide a reference for control theory in flue gas desulfurization process under variable conditions.