Study On Modeling Method Based On Noisy Data Driven In Power Plant Boiler

Thermal power units account for about75%of the power generation capacity inour country. With the power demand growth and the increasing concern onenvironmental quality, improving the combustion efficiency of boilers in power plantand reducing pollution emissions are more and more important. The power plant boileris a complex system, which includes physical and chemical processes, such as fuelcombustion, the working material flow and heat transfer, et al. It is difficult to describethe process with precise mathematical model, and causes difficulties on optimizing thesystem operation. Some important parameters in the boiler can reflect the boilercombustion efficiency, such as carbon content in fly ash, flue gas oxygen content,wind powder concentration and nitrogen oxide content. But it is difficult to obtainthese parameter values by direct measurement because of the complex combustionprocess, strong correlation and the technical or economic reasons. It is an effectivemethod to establish the soft sensor model of these key parameters through the dataaccumulated in the unit during the operation of the unit.This paper focuses on the flue gas oxygen content modeling. A data-driven softsensor modeling method is used to solve the problem of measurement. The data isvulnerable to noise pollution under outside interference or internal equipment vibrationand so on. In the paper, the noisy data is processed by wavelet denoising; the softsensor model based on least squares support vector machine (LS-SVM) is established.To improve the measurement accuracy and generalization ability, an improved particleswarm optimization (PSO) is applied to optimize the parameters of LS-SVM. Themain work of this paper is as follows:The wavelet shrinkage threshold de-noising method is proposed by analyzingseveral commonly used wavelet transform method, the effects of threshold function,decomposition level, wavelet and threshold selection method on denoising areanalyzed. The noisy data collected from boiler operation is processed by wavelet shrinkage threshold denoising method. Simulation results show that the waveletshrinkage threshold method can remove the noise effectively.Considering the work process of power station boiler, and taking flue gas oxygencontent as an example, the flue gas oxygen content soft sensor model based on leastsquares support vector machine is established. Taking into account the impact of datanoise on the accuracy of LS-SVM model, wavelet shrinkage threshold denoisingmethod is used to process the collected sample data, and establish the LS-SVM model;simulation shows that combining wavelet threshold denoising and LS-SVM modelingcan improve the accuracy of the model compared with that without denoising.The parameters of the model are important factors to the performance of themodel. A parameter selection method based on improved particle swarm optimization(IPSO) is presented to optimize the LS-SVM parameters. The simulation results showthat IPSO can improve the efficiency of parameter optimization. Then the soft sensormodel based on IPSO-LSSVM is proposed to model the flue gas oxygen content. Thesimulation results show that the model accuracy can be improved by IPSO-LSSVMcompared with LS-SVM model.