Abstract
One of the keys to clean power production is the control of nitrogen oxide (NOx) emissions. The NOx concentration must be accurately tracked and monitored during the denitrification process in power plants. This research proposes a novel NOx emission prediction model, which innovatively establishes dynamic graph topology structures to explore the internal connections between thermal process data. It includes carefully curated and improved feature selection, delay analysis, graph attention mechanism, Transformer, and other modules. The method demonstrated commendable predictive capabilities in a 1000 MW ultra-supercritical power generation unit at the Shanghai Caojing power plant in China, achieving mean absolute error values of 0.7675, root mean square error values of 0.9800, and mean absolute percentage error values of 2.005 for NOx. When benchmarked against the six latest single and hybrid data-driven models, the model proposed in this research consistently surpassed them, demonstrating outstanding predictive performance. By weaving together advanced deep learning, powerful data preparation, and astute delay analysis, this work not only highlights the effectiveness of data-driven modeling solutions for complex problems in the power generation industry, but also lays a crucial foundation for clean power production and green energy transition.