Online blend‐type identification during co‐firing coal and biomass using SVM and flame emission spectrum in a pilot‐scale furnace

Abstract

Co-firing coal and biomass has been applied in existing coal-fired power stations recently. Online blend-type identification was investigated by support vector machine (SVM) using flame emission spectrum for combustion optimisation. A spectrometer was used to capture the flame emission spectrum during co-combustion in a 0.3 MW furnace. A total of 22 flame features were defined and extracted from the flame emission spectrum for blend-type identification. ReliefF was applied to calculate the important weights of the extracted fame features. Alkali metals atomic excitation spectral intensities and the means of spectral signals show obviously higher important weights than the other flame features. Ultraviolet signal is more important than visible and infrared signals for blend-type identification. SVM was adopted to identify the blend types. The method of ‘ReliefF + SVM’ was proposed to obtain the optimum feature vector. The number of optimum features can be reduced from 22 to 17 if only the prediction accuracy is considered. The optimum sampling number is 12. At the optimum feature vector (17 features) and the optimum sampling number (12), the average prediction accuracy of the five fuels is 99.67%. The results demonstrate that combining SVM and flame emission spectrum is suitable for online blend-type identification during co-combustion.