Outliers Treatment for Improved Prediction of CO and NOx Emissions from Gas Turbines Using Ensemble Regressor Approaches

Gas turbines are widely used in power generation plants due to their high efficiency, but they also emit pollutants such as CO and NOx. This study focuses on developing predictive models for predicting CO and NOx emissions from gas turbines using machine learning algorithms. The dataset used includes pollutant emission data from a combined cycle gas turbine (CCGT) in Türkiye, collected hourly between 2011 and 2015. Various outlier treatment methods such as Z-Score, Interquartile Range (IQR), and Mahalanobis Distance (MD) are applied to the dataset. Machine learning algorithms including Random Forest, Extra Trees, Linear Regression, Support Vector Regression, Decision Tree, and K-Nearest Neighbors are used to build the predictive models, and their performances are compared. Additionally, Voting Ensemble Regressor (VR) and Stacking Ensemble Regressor (SR) methods are employed, using Gradient Boosting, LightGBM, and CatBoost as base learners and XGBoost as a meta-learner. The results demonstrate that the SR model, when applied to the dataset processed using the IQR method, achieves the highest prediction accuracy for both NOx and CO emissions, with R² values of 0.9194 and 0.8556, and RMSE values of 2.7669 and 0.4619, respectively. These findings highlight the significant role of the IQR method in enhancing model accuracy by effectively handling outliers and reducing data noise. The improved data quality achieved through this method contributes to the superior performance of the SR model, making it a reliable approach for predicting NOx and CO emissions with high precision.