Industrial gas turbine emissions prediction using data-driven hybrid stacked ensemble machine under dynamic operating conditions

Industrial gas turbines operate under dynamic conditions in which startup, steady-state, and transient operation exhibit distinct combustion characteristics. Conventional machine learning-based Predictive Emissions Monitoring Systems (PEMS) generally employ a single model trained using aggregated operational data, implicitly assuming a stationary relationship between turbine operating variables and emissions. This assumption often degrades prediction accuracy during operating transitions due to non-stationary combustion behavior. To address this limitation, this study proposes a regime-aware heterogeneous stacked ensemble framework for multi-pollutant emissions prediction. Historical industrial gas turbine data are first segmented into startup, steady-state, and transient operating regimes, after which independent predictive models are developed for each regime. The proposed framework combines XGBoost, Decision Tree (DT), and LightGBM as complementary base learners, while ElasticNet serves as a regularized meta-learner to integrate their predictions. The proposed model predicts CO, CO 2 , SO 2 , and NO x emissions and is evaluated using independent validation datasets. During startup, validation MAPE ranges from 1.38% to 5.88%, with R 2 values between 0.39 and 0.95. Under steady-state operation, validation MAPE ranges from 1.23% to 8.07%, with R 2 values between 0.88 and 0.97, where the largest prediction error is associated with CO while the majority of pollutant–regime models satisfy the preferred industrial target of 5% MAPE. During transient operation, validation MAPE ranges from 1.20% to 5.57%, with R 2 values ranging from 0.73 to 0.98. These results demonstrate that predictive performance depends on both the operating regime and pollutant characteristics, while consistently providing strong explanatory capability across diverse operating conditions. Model interpretability is investigated through complementary sensitivity analyses using Mean Decrease in Impurity (MDI), Permutation Importance, and SHapley Additive exPlanations (SHAP), which consistently identify regime-dependent dominant process variables. Furthermore, measured inference latencies of less than 20 ms per prediction demonstrate the computational efficiency of the proposed framework for near-real-time industrial deployment. Overall, the proposed regime-aware stacked ensemble provides an interpretable, computationally efficient, and practically deployable solution for industrial multi-pollutant Predictive Emissions Monitoring Systems.