Non-sampling estimation of waste composition based on incineration flue gas pollutant fingerprints and machine...

Understanding mixed waste composition is crucial for controlling pollutants in incineration flue gas. Real-time detection of pollutants is challenging owing to the complexity and heterogeneity of mixed waste, complicating incineration optimization and pollutant control. This study introduces a novel method to rapidly predict mixed waste composition using waste incineration flue gas "fingerprint" and machine learning regression models, bypassing traditional sampling processes. A comprehensive "fingerprint" dataset was established via incineration experiments with various mixed waste compositions, featuring multiple waste components and their associated flue gas pollutant concentrations. Predictive performance was compared for five machine learning models, including extreme gradient boosting tree (XGBOOST), K-nearest neighbor (KNN), random forest (RF), light gradient boosting machine (LGBM), and support vector regression (SVR). After feature importance analysis optimization, RF and XGBOOST models achieved the best performance, with R² values exceeding 0.92 for key waste types. The accuracy of the models in predicting waste composition was significantly improved compared to that without optimization. Beyond predictive accuracy, the proposed method enables near real-time waste composition estimation, offering significant advantages for dynamic adjustment of feedstock and operating parameters. This facilitates intelligent incineration control, enhances energy efficiency, and supports proactive pollution management at the emission source.