A machine learning based multi-objective optimization for flue gas desulfurization enhancement in coal power plants

Abstract

Coal-fired power plants emit large quantities of hazardous pollutants including sulfur dioxide (SO₂), oxides of nitrogen (NOx) and Mercury (Hg) that threaten environmental sustainability. Flue gas desulfurization (FGD) systems are widely deployed to reduce SO₂ emissions, yet their performance depends on large number of interacting operational variables, making real-time optimization challenging. This research aims to develop a practical, data-driven optimization framework for performance improvement of industrial-scale FGD systems. Artificial neural network (ANN) based process models have been trained for its proven capability to model complex nonlinear relationships in high-dimensional process data, and reasonable memory requirement for making excellent function approximate for real-life applications. Two years of continuous operational data from a 660 MW coal power plant were used to train ANN models that predict desulfurization efficiency, NOx, and Hg emissions based on key flue gas and slurry parameters. Monte Carlo sensitivity analysis showed that absorber slurry pH, inlet NOx concentration, and inlet dust concentration are the dominant factors for the three outputs, respectively. A Non-Dominated Sorting Genetic Algorithm II (NSGA-II) was applied to determine optimal operating settings under varying plant load scenarios, with the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) selecting the most balanced solutions. Results show that the optimized conditions improve SO₂ removal efficiency while reducing NOx and Hg emissions compared to conventional setpoints. The proposed framework offers a practical pathway for cleaner and more efficient operation of large-scale FGD systems, supporting the power sector’s net-zero objectives.

Keywords: Desulphurization efficiency, Emissions control, Coal power plant, Net-zero, Machine learning