Uncertainty and sensitivity analyses of co-combustion/pyrolysis of textile dyeing sludge and incense sticks: Regression and machine-learning models

Abstract

Bioenergy generation from biomass waste through co-combustion/pyrolysis fulfills simultaneously multiple objectives of reductions in fossil fuel use, greenhouse gas emission, and solid waste stream. This experimental study aimed to quantify the multiple co-combustion/pyrolysis responses of textile dyeing sludge (TDS) and incense sticks (IS) as a function of blend ratio (BR), heating rate (HR), atmosphere type (Atm), and temperature (Temp). Joint optimizations, and predictor importance, sensitivity, uncertainty and interaction analyses were conducted using data-driven models for the responses of remaining mass (RM), derivative thermogravimetry (DTG), and differential scanning calorimetry (DSC). The data-driven models compared in this study were Box Behnken design (BBD)-based regression models, general linear models (GLM), and the six full models with all the predictors included of multivariate adaptive regression splines, multiple linear regressions, random forests (RF), regression decision tree (RDT), RDT with ensemble and bagger, and gradient boosting machine. BBD, GLM, and Sobol’s total and first-order indices indicated HR as the most important and sensitive predictor in the joint optimizations. GLM pointed to a three-way interaction among HR, BR, and Atm, while BBD, and Sobol’s second-order index showed a two-way interaction between HR and BR as the most important ones. RF outperformed the other full models for all the responses in terms of validation metrics. RF showed the two most important predictors as Temp and BR for RM; HR and Temp for DSC; and Temp and HR for DTG, respectively, which also constituted the most important two-way interactions.