A combined application of advanced statistical, soft computing and least square methods for the prediction of higher heating value of coal

초록

The reliability of some empirical models for the prediction of higher heating value (HHV) based on the proximate analysis by using a total of 147 coal samples collected from Witbank coalfields in South Africa is evaluated. The most statistically reliable model is selected among the empirical models using the nonparametric Mann - Whitney's p-value after subjecting the models to normality tests using Shapiro-Wilk, Lillefors, and Anderson-Darling methods. This selected model is then enhanced using an artificial neural network (ANN) to improve its predictive performance and extrapolation capability. The most statistically reliable model among the models subjected to the rigorous statistical tests is compared with the measured HHV and the ANN predicted HHV. The coefficient of determination (R2) between the selected model and the measured HHV is 0.9482, while that between the selected model and the ANN output is 0.9625. A direct comparison between the ANN predictions and the measured HHV yields a higher R2 value of 0.9868. A new hybrid equation is developed by integrating the selected empirical model with the ANN and further optimized using the least squares method via Newton's iterative approach. To validate the models, an additional 50 datasets from existing literature are used. The ANN model gives very low R2 value of 0.2921 for the existing literature data indicating its low extrapolation potential while the new hybrid model gives R2 value of 0.9372.

키워드