Machine Learning Prediction of CO and NO_x Concentrations from Tea Waste Pellet Combustion

Tea residues can be converted into pellet fuel, but interactions between fuel properties and combustion conditions complicate emission prediction. This study evaluated 60 independent combustion tests, comprising two moisture classes (8–9% and 11–12%), three pellet diameters (6, 8, and 10 mm), and ten tests per combination. Ridge regression, radial-basis support vector regression (SVR), and random forest regression were compared with global-mean and treatment-mean benchmarks. Two predictor sets were assessed: moisture class and diameter alone, and an expanded set that also included O₂, CO₂, lambda, and combustion-air temperature. Hyperparameters were selected using nested cross-validation. Repeated treatment-stratified validation assessed new tests under represented conditions, whereas leave-one-combination-out validation examined transfer to unobserved treatment combinations. Mean reported CO concentrations ranged from 1476.9 to 8786.4 mg m⁻³, and NOx from 87.3 to 607.5 mg m⁻³. With fuel properties alone, SVR achieved mean cross-validated R² values of 0.869 for CO and 0.903 for NOx, with RMSE values of 946.1 and 51.1 mg m⁻³, respectively. However, RMSE improvements over the treatment-mean benchmark were only 3.2% and 6.8%. Adding gas-state variables did not improve the best internal prediction performance. For unobserved treatment combinations, fuel-only SVR yielded negative R² for both targets, demonstrating limited transferability. The results support restricted prediction within the tested design rather than general fuel optimization. Confirmation of the concentration-reporting basis and prospective validation across additional operating conditions are required before operational deployment.