Research On Prediction Of Oxygen Content In Fiue Gas Based On Partial Least Squares

The oxygen content in flue gas is one of the most important parameters for both security and economy of the boiler units, which reflects operating conditions. Due to the influence of various factors in field, it is difficult to ensure the measuring of oxygen online achieve a certain precision. Soft-sensing, a new type of technique in measuring field, provides a new effective way, which makes it possible to get the uneasy measured parameters in an indirectly way, applied to the complex thermal process of the power plant. There are many ways to set up a soft-sensing model. In order to reduce error and measure oxygen in an effective way, this thesis chooses the partial least squares (PLS) to establish both linear and nonlinear models, then compares the results after simulation. The main contributions of this dissertation can be summarized as followed:(1) To enhance the accuracy of the forecast, the thesis proposes the application of partial least squares linear regression and nonlinear regression based on spline transform to the modeling and prediction of flue gas oxygen. Set up the model by selecting12auxiliary variables that can affect the flue gas oxygen content obviously. The establishment of the two models can overcome the high degree of correlation between multiple independent variables, identify outliers in sample library effectively, reduce the error caused by abnormal data, simplify the model and improve the efficiency of computing.(2) After a review of the history and current situation about PLS method, the simply description of the principles in how to set up the model, choose auxiliary variables, determine the number of ingredients, analyze the precision and specific outliers using PLS method are given. Considering there may exist nonlinear factors, propose to build nonlinear oxygen model. Original nonlinear components turned into linear ones through spline transform, simplifying the model as well.(3) The simulation results show that, in either sufficient sample numbers and lack of cases, the accuracy of oxygen content prediction model based on partial least squares can meet the requirements. It proves that this method is reasonable and worth to be put into reality, which can be more effective for small sample with multivariable data.