Machine Learning-Based NOx Prediction and Combustion Optimization for a Coal-Fired Boiler: The Role of Feature Engineering

Abstract The choice between traditional machine learning and deep learning for industrial time series prediction remains unsettled, particularly when feature engineering is accounted for. Using 10,000 operational samples from a 660 MW coal-fired boiler, we systematically compare support vector machine (SVM), LSTM, XGBoost, Random Forest, and MLP for NOx emission prediction. Using a 139-dimensional representation comprising the current and first-to-fourth-order lagged values of 27 operational variables, together with four autoregressive NOx lag terms, SVM achieves RMSE = 10.124 mg/Nm3 and R2 = 0.934, outperforming LSTM (20.203, 0.738), XGBoost (12.405, 0.901), Random Forest (12.475, 0.900), and MLP (32.236, 0.332). A fair re-evaluation─training LSTM, Transformer, TCN, and BiLSTM on raw 20-step sequences without manual lag engineering─confirms that deep architectures remain less competitive than SVM with explicit feature engineering (best DL model: fair LSTM, RMSE = 18.541, R2 = 0.779). The HAC-adjusted Diebold–Mariano tests (all one-sided p < 8.1 × 10–7), bootstrap confidence intervals, and Friedman–Nemenyi analysis provide convergent statistical evidence of SVM’s superiority. SHAP analysis identifies O2 as the dominant predictor (SHAP = 10.987), and time-series cross-validation confirms SVM’s robustness (mean RMSE = 6.062, std = 1.804). Single-objective differential evolution achieves a 21.16% NOx reduction at 600 MW load. A data-driven efficiency surrogate replaces hand-crafted formulas in multiobjective NSGA-II optimization, producing a scenario-dependent NOx-efficiency Pareto front whose sensitivity to coal calorific value is explicitly evaluated. These findings demonstrate that on industrial data sets of approximately 10,000 samples, domain-informed feature engineering can enable traditional methods to match or exceed deep learning at lower computational cost.