The boilers as key equipment in thermal power plants, improving its operating economy directly affects production efficiency of the plant, while the reasonable determination the target value of boiler operation optimization is the key. Firstly, an improved Gaussian mixture model applicable to working condition division was proposed. The evaluation criterion function was constructed by dividing the dispersion of data as basis and based on Mahalanobis distance to determine the number of clusters. Secondly, by building a deep recurrent neural network with long-short term memory function, the economic model under each working condition interval was established. Finally, on the basis of the economic model, an improved particle swarm algorithm was obtained by adjusting the inertia weight and acceleration factor to address the problem that the traditional particle swarm algorithm was prone to fall into local extremes, which could more accurately perform interval optimization under different working conditions and determine the optimized target values of the operating parameters were determined. The experimental results show that the optimization target values determined by this method are better than the historical optimal operating values, indicating that this method has certain advantages in exploiting the potential of boiler optimization, and that the boiler operation can be adjusted according to this scheme to effectively reduce the energy consumption level, so as to achieve the purpose of boiler economic operation.