Abstract: In order to reduce the pollutant emissions of a circulating fluidized bed boiler in a certain power plant and improve the economy of the boiler combustion operation, this article adopts the data-driven technology to achieve the multi-target combustion optimization for circulating fluidized bed boilers. Improved particle swarm optimization-based long short-term memory neural networks is used to establish the boiler’s mathematic model with NO_x emission, SO₂ emission and exhaust gas temperature as outputs, respectively. The relative error is regarded as a predictive evaluation index to determine the optimal network parameters. Secondly, the NO_x emission prediction model, the SO₂ emission prediction model and exhaust gas temperature prediction model are constructed based on improved particle swarm optimization-based long short-term memory neural network, long short-term memory neural network（LSTM）,generalized regression neural network（GRNN）, and a backpropagation neural network（BPNN）. By comparing the evaluation indicators, the effectiveness of the predictive models constructed was testified in this paper; Finally, based on the non-dominated sorting genetic algorithm（NSGA-II）, the combustion optimization adjustment schemes for CFBB under different operating conditions are obtained so as to reduce NO_x/SO₂ emission and maintain the stability of exhaust gas temperature at the same time. The results showed that compared with before optimization, the average NO_x emission was decreased by 10.583%, the average SO₂ emission was reduced by 25.812%, and the maximum reduction of SO₂ emission was 650 mg/m~3. In addition, the average exhaust gas temperature was decreased by 0.143%.