Enhancing Predictive Emissions Monitoring Performance: Data Preprocessing for XGBoost-Based Model Algorithm

Gas turbines play a crucial role in processing industries and power generation, but their emissions present significant environmental challenges. Accurate emissions prediction is essential, with research focusing on improving statistical models, machine learning algorithms, and hybrid systems. However, varying data quality and dynamic industrial environments often lead to inconsistent model performance. This study examines the impact of data cleaning and splitting on the accuracy of XGBoost-based emissions prediction models, emphasizing the importance of data pre-processing in enhancing performance. The tuning process significantly reduced errors across all metrics, with CO showing the most notable improvement despite high initial errors, while \text{SO}_{2} demonstrated substantial gains, particularly in MAPE. Hyperparameter tuning further enhanced the model's accuracy, reduced error magnitudes, and improved generalization. Regular monitoring and periodic retraining with updated data are recommended to maintain model effectiveness in production environments. If performance improvements plateau, simplifying the model by reducing trees or applying early stopping can optimize efficiency without compromising accuracy.