Explainable ensemble learning framework for predicting industrial and energy sector GHG emissions

Accurate prediction of greenhouse gas (GHG) emissions in the energy and industrial sectors are challenging due to nonlinear relationships, data uncertainty, and sectoral variability. Traditional statistical and machine learning models often fail to generalize across diverse emission patterns. To overcome these challenges, this study applies and compares advanced ensemble models Gradient Boosting (GB), Random Forest (RF), XGBoost, CatBoost, and AdaBoost for global multi-gas (CO₂, CH₄, N₂O, and F-gases) prediction over a 172-year period. Performance evaluations of industrial processes and product consumption indicated that the GB and XGBoost models exhibited superiority, with R² values of 0.9997, RMSE ranging from 0.43 to 0.43, and minimal MAE (0.05 to 0.06). In the energy sector, Gradient Boosting attained the maximum accuracy (R² = 0.9997, RMSE = 2.88, MAE = 2.09). This study forecasts future GHG emissions and identifies pertinent variables, in contrast to previous studies. A transparent prediction framework utilizing explainable artificial intelligence (XAI) techniques such as SHAP, LIME, ICEs, and PDPs demonstrated the contribution of each greenhouse gas to emissions. CO₂'s substantial contribution is quantified using XAI, yielding mean values of 0.384 in the energy sector and 0.291 in industrial sectors, which correspond to normalized feature impacts of 68.4% and 64.2%, respectively. The significant influence of CO₂, CH₄, and F-Gases on greenhouse gas emission estimations is evidenced by ICE mean gradients reaching 0.92, LIME local fidelity scores of 0.941 (energy) and 0.924 (industrial), and PDP joint variance increases of 36-39%. The findings of this study can help in better understanding the key drivers of GHG emissions and support data-driven decision-making for emission control, policy development, and sustainable management in industrial and energy sectors.