A hybrid machine learning model for NOx emission concentration prediction from sludge incineration

Abstract

Accurate prediction of NO_x emission concentration is crucial for optimizing combustion processes and enhancing flue gas treatment in incineration systems. However, the traditional prediction models that employ data-driven methods face significant challenges due to insufficient input feature information, as well as low computational efficiency and robustness. These limitations hinder the accurate real-time prediction of NO_x emission concentration. To address this issue, this paper proposes an advanced hybrid model for predicting NO_x emission concentration. Initially, static and dynamic flame features are extracted from flame images and integrated with Distributed Control System (DCS) parameters to serve as the model's input features, while the NO_x emission concentration constitutes the model's output feature. Subsequently, the lag time between NO_x emissions and the input features is determined using mutual information (MI), followed by data reorganization to develop various predictive models for NO_x emission concentration. Finally, the extremely randomized trees (ERT) model, demonstrating superior performance, is further optimized using Bayesian optimization with tree-structured Parzen estimators (BO-TPE). Experimental results indicate that the ERT model optimized with BO-TPE outperforms state-of-the-art models, making it suitable for online optimization of industrial pollutant control and potentially contributing to cleaner production.