Explainable AI-based optimization of facility operations for excessive NOx emission control: A case study of a solid refuse fuel combustion facility

Thermal facilities have complex operational processes, which hinders optimizing facility operation for excessive greenhouse gas emissions control. This study proposes an XAI framework to identify key operational variables and optimize the facility operation for excessive NO x emissions control (< 40 ppm) at a Solid Refuse Fuel (SRF) combustion facility in South Korea. In this study, a harmonized Multi-Ensemble Machine Learning (hMEML) model is proposed to predict excessive NO x emissions using the data of 30 input operational variables at the SRF combustion facility over 2019–2023. Then, the trained hMEML are utilized to develop the optimized facility operation for excessive NO x emissions control via the permutation importance score analysis, sensitivity test, and SHAP analysis. For predictive performance, the hMEML model is superior to individual AI models, with an R 2 of 0.978 (0.895) for training (testing) samples. The permutation importance scores determine the two key operational variables as the combustion furnace temperature and secondary forced draft fan electric current. Results show that excessive NO x emissions can be controlled efficiently by decreasing both key variables by −3.6%. This study also finds that abrupt changes in facility operation decrease the predictability of the hMEML model, indicating the potential of the hMEML model to detect and monitor changes in facility operation. This study highlights the trustworthiness of XAI in smart management for NO x emissions control at thermal facilities. • A harmonized multiensemble machine learning (hMEML) model was proposed. • hMEML outperformed eight different AI models in the prediction of NOx emissions. • The XAI framework identified key operational variables of NOx emission. • −3.6% of two key operation variables was efficient excessive NOx emission control. • This study proved the potential of hMEML in detecting facility operation changes.