Prediction and Comparison of Ash Fusion Temperatures Based on BP Neural Network and Least Squares Support Vector Machine

Abstract

To predict the slagging on heating surface of coal-fired boilers, BP neural network (BPNN) and least squares support vector machine (LSSVM) prediction models were established to predict ash fusion temperature, deformation temperature (DT) and softening temperature (ST). The models take ash metal oxide, SO3 content of ash and slagging evaluation indexes as independent variables, and take DT and ST as dependent variables. Regression analysis and error analysis show that when predicting DT with a large number of samples, the prediction accuracy of the two models is similar, and the confidence of prediction is over 95%. The correlation coefficients are both about 0.92, and the average relative errors are about 3.4%. When predicting ST with less samples, LSSVM model is better with a correlation coefficient of 0.950 52, which is higher than 0.904 26 of BPNN model. The average relative error is 4.98%, and the number of large error points is less than the BPNN model. Therefore, LSSVM model can predict DT and ST of fly ash more accurately.

Keywords

• bp neural network (bpnn)
• least squares support vector machine (lssvm)
• ash fusion point
• ash composition
• slagging evaluation index