Machine learning-based prediction of mercury and arsenic emissions under real operating conditions in pulverized coal boilers

Abstract

Mercury (Hg) and arsenic (As) emissions from coal-fired power plants remain a major environmental challenge due to their toxicity and complex partitioning behavior during combustion. This study presents a combined experimental and machine learning analysis of Hg and As emissions from two 225–240 MW pulverized coal boilers operating under variable load and 10% biomass co-firing conditions. A dataset comprising 73 operational variables and 497 measurement cases was used to develop regression and classification models. For mercury, multilayer perceptron (MLP) networks achieved good predictive performance, with the optimized 49–17-1 architecture reaching R2 = 0.92 and MAPE = 10.5% on the validation dataset. Feature reduction from 67 to 49 predictors improved generalization while preserving physical interpretability. Sensitivity analysis indicated that Hg emissions are influenced by combustion temperature, excess air ratio, NO2 concentration, and gas flow dynamics, highlighting the role of thermal and oxidative conditions. For arsenic, this work represents, to the best of our knowledge, one of the first applications of artificial neural networks for predicting flue gas As concentrations in pulverized coal boilers. The MLP 15–15-1 model achieved R2 = 0.971, with fuel arsenic content and particulate matter concentration identified as the most influential predictors. The developed approach enables both continuous emission prediction and threshold-based monitoring, supporting improved analysis of trace element behavior in coal combustion systems.