Improvement Of Extreme Learning Machine And Teaching Learning Based Optimization Algorithms And Its Application In Boiler Combustion Optimization

Extreme Learning Machine?ELM?is a kind of single hidden layer feed-forward neural network with high computation speed and good generalization ability.Teaching-Learning-based Optimization algorithm?TLBO?is a heuristic algorithm,which is inspired by the behavior of class teaching.ELM and TLBO can solve the modeling and optimization problems of complex systems,which have been applied in many fields.In view of the complex combustion process of circulating fluidized bed boiler?CFBB?,it is difficult to establish the combustion characteristic model by traditional method and to realize the optimization of boiler combustion process.Therefore,the ELM and TLBO algorithms are studied intensively in our dissertation,and they are applied to model and optimize the CFBB combustion process,which can realize the boiler combustion with high efficiency and low pollution.The main research content has theoretical significance and practical application value,which is summarized as follows:In the dissertation,three kinds of improved extreme learning machines are proposed to solve the problems of setting the parameters and model structure of the original ELM.In order to improve the self-adaptability and model stability of ELM to training sample data,a sample adaptive extreme learning machine is proposed,whose input weight threshold can be determined according to training sample data.And the proposed method has a new hidden layer activation function.In addition,an optimized extreme learning machine is proposed,which uses artificial intelligence optimization algorithm to find the optimal input weight threshold of the extreme learning machine,so that ELM has good model generalization ability.Finally,based on the ELM,a sample incremental quantum double parallel feed-forward neural network with online learning capability is raised,which initializes the input weight threshold according to the quantum calculation rules.Moreover,the hidden layer neuron is a quantum neuron and the model parameters are dynamically adjusted on line with the increment of data samples.In view of all the features of the sample incremental quantum double parallel feed-forward neural network mentioned above,the dynamic modeling of complex systems can be achieved.Compared with many famous neural networks,the improved ELM method has the better stability and generalization ability through the test of UCI data set.For the deficiencies of the original TLBO algorithm,three improved teaching learning optimization algorithms are put forward.Firstly,for the purpose of improving the convergence accuracy of TLBO algorithm and reducing the number of convergence iterations,a chaotic grouping teaching and learning optimization algorithm?CLF-TLBO?is proposed,which initialize individual populations of populations in means of chaotic cubic sequences to increase population diversity.Furthermore,the introduction of inertial weights to the teaching phase can improve the search speed of the algorithm and the combination of grouping idea of the random frog leaping algorithm and learning stage can enhance the local search ability of the algorithm.Secondly,on the basis of CLF-TLBO,a new type of individual location updating mechanism was introduced to improve the speed of CLF-TLBO via replacing the original frog leaping algorithm grouping idea.The second proposed algorithm is called CTLBO.Finally,a kind of teaching learning optimization algorithm in the light of the real teaching phenomenon is presented,which is called MTLBO.It uses chaotic cubic sequences to initialize the population individuals and updates the position of individual individuals by improving grouping mechanism in teaching and learning stages.It can be seen that the algorithm can improve the convergence speed and solution quality of the algorithm and balances the global exploration capabilities and local development capabilities.The initial TLBO population initialization,the individual updating mechanism in the teaching phase and the individual updating mechanism in the learning phase are improved by the three improved TLBO algorithms mentioned above.The simulation results of several classical test set functions show that the improved mechanism has improved the convergence accuracy and reduced the number of convergence iterations of original TLBO algorithm.In this dissertation,the 300MW circulating fluidized bed boiler of a thermal power plant is taken as the research object.Aiming at the combustion process modeling and optimization problem,three improved ELM algorithms are used to establish off-line and on-line models of the boiler thermal efficiency,the concentrations of nitrogen oxide?NOx?and sulfur dioxide?SO₂?,respectively.It not only verifies the performance of the improved ELM,but also solves the problem that the combustion process of CFBB is difficult to model based on mechanism.The simulation results show that the improved method has good stability and generalization ability,and the established model has high accuracy.Based on the established model,the MTLBO algorithm is used to optimize the boiler thermal efficiency and NOx and SO₂ emission concentration,respectively.The simulation results show that,without considering the NOx and SO₂ emission concentrations,the thermal efficiency of CFBB can be improved according to the optimized operating parameters.Common knowledge,the NOx and SO₂ emission concentrations of the boiler are decreased to some extent respectively.In addition,a multi-objective MTLBO algorithm is proposed to achieve simultaneous optimization of the three objectives of CFBB.Based on the comprehensive model of three objectives,the multi-objective MTLBO algorithm is used to optimize the operation parameters of the boiler.Finally,an optimal solution set is obtained.According to the demand of the power plant,one set of optimized operation parameters can be selected to realize the boiler combustion with high efficiency and low pollution emission.