Research On Combustion Optimization Of Coal-Fired Boiler Based On Least Squares Support Vector Machine

In recent years, energy problems and environmental protection have gained more and more attention during the development of society. As the main type of installed units in China, clean and efficient electricity generation of coal-fired power plant is crucial to the sustainable development of our country. Combustion optimization techniques based on model prediction and multi-objective optimization algorithms has become a research focus in the study of power plants, which requires little reformation performed to power plants.The development of boiler combustion model entirely based on mechanism analysis is unrealistic due to the complex characteristics of multivariate and strong coupling of boiler combustion process. Artificial intelligence algorithm and multi-objective optimization algorithm is used in this paper to optimize boiler combustion. This study carried out the following areas of research work:Combustion optimization adjustment test was done in a 600MW supercritical once-through boiler through which samples are collected. Analysis of the characteristics of boiler combustion is then performed on the basis of the results of the test.Based on the preprocessed experimental data, least squares support vector machine (LSSVM) is used to build the models of boiler thermal efficiency and NOx emissions of test boiler, which meet the accuracy requirements of practical application. Incremental analysis in which some important input variables of combustion models are changed slightly to verify the validity of boiler combustion model.The multi-objective optimization of combustion models is done through fast non-dominated sorting genetic algorithm with elitist strategy (NSGA-?) based on experimental data from different load range. Recommended value of operating variables is given by choosing the optimal solution from Pareto optima set.Classic pruning algorithm and active learning pruning algorithm are applied to boiler combustion models built earlier to obtain sparse LSSVM models. The rapidity and stability of active learning pruning algorithm is proven through the comparison of the pruning process of the above pruning algorithms. The generated sparse combustion models can facilitate future multi-objective optimization and incremental learning.