Improving syngas yield and quality from biomass/coal co-gasification using cooperative game theory and local interpretable model-agnostic explanations

Abstract

The co-gasification of waste biomass and low-quality coal to produce syngas as fuel is an effective and sustainable approach in the waste-to-energy paradigm. The modeling of this process is however complex and time-consuming. The data-driven machine learning (ML) approaches enhanced with explainable artificial intelligence (XAI) are capable of solving this issue. Hence, in this study, five different ML techniques including Linear Regression (LR), Support Vector Regression (SVR), Gaussian Process Regression (GPR), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) were employed for the model-prediction. The ultimate analysis, proximate analysis, and operation setting data were employed for the control factors syngas yield and lower heating value (LHV) prediction. The prediction results showed that XGBoost was superior to other ML approaches with an R² value of 0.9786, mean squared error (MSE) of 10.82, and mean absolute percentage error (MAPE) of 9.8% during model testing of the syngas yield model. In the case of the syngas LHV model an R² value of 0.9992, MSE of 0.03, and MAPE of 0.83% was observed. XGBoost was superior for both syngas yield and LHV models. The analysis of feature importance and its quantification was conducted by Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). SHAP and LIME approaches revealed that reaction temperature and biomass mixing ratio were the most important control factors for the syngas yield model.