Boiler Combustion Modeling And Optimization Based On Improved LSSVR Algorithm

The comprehensive benefits of coal-fired generating units need to be improved urgently under the background of increasing demand for electricity and increasingly strict requirements for energy conservation and environmental protection.So boiler combustion optimization has become a research focus.The optimization of boiler combustion can be divided into two parts:the establishment of accurate prediction model of boiler,and the efficient multi-objective optimization method of boiler.The boiler combustion process is complex,the mechanism model cannot accurately reflect the influence of operating parameters on boiler efficiency and NO_xemission.And mechanism model is inconvenient for follow-up optimization.Data modeling and optimization has been widely used because of its fast and accurate advantages.The least squares support vector regression(LSSVR)algorithm has the advantages of fast training and small sample modeling.However,the algorithm selects all the samples as support vectors,resulting in the algorithm’s shortcomings such as poor generalization and sensitivity to error samples.Therefore,an improved algorithm,Constrained Support Vector Regression(CSVR)algorithm,is proposed.Based on the LSSVR algorithm,the new algorithm analysis and simplifies the parameter,and the support vector constraint is proposed to optimize the support vector.Two test functions are used to test the new algorithm,and the results show that the new algorithm has high fitting accuracy for multi-modal functions while choosing fewer support vectors.In addition,when radial basis function is used in the algorithm,the influence of kernel parameter and normalized parameter on the algorithm is also analyzed.The analysis shows that selecting the appropriate parameter combination can improve the algorithm accuracy,so it is necessary to optimize the parameters in the subsequent modeling.1000 sets of operation data of a 600MW unit were collected and treated by normalization and principal component analysis as samples.The CSVR algorithm was used to establish the prediction model of boiler efficiency and NO_xemission concentration,and particle swarm optimization(PSO)algorithm was used to optimize the modeling parameters.The traditional PSO algorithm has the problems of slow convergence and prematurity,so the adaptive inertia coefficient and fixed interval particle perturbation strategy are proposed.The test functions were used to test the improved PSO algorithm.The results show that the improved algorithm converges faster and can jump out of local optimum when dealing with complex multimodal functions.The modeling results of boiler efficiency and NO_xemission concentration show that compared with BP neural network and LSSVR algorithm,the CSVR algorithm has the highest prediction accuracy after parameter optimization,and the CSVR algorithm is not sensitive to individual error samples,and has better generalization performance.Based on the boiler combustion prediction model,linear fitting method and fast non-inferior sorting genetic algorithm(NSGA-II)were used respectively to optimize the boiler efficiency and NO_xemission concentration of two high NO_xemission samples.The optimization results show that both methods can get the optimal solution,but NSGA-II can get more feasible solutions in one run,which can reflect the different boiler efficiency corresponding to different NO_xemission concentration after optimization.NSGA-II can provide different optimization reference for operators after accumulating optimization data.