Abstract
Accurately predicting pollutant emissions using real-time operational data is critical to enhancing the environmental efficiency of coal-fired power stations. Given the complex, non-linear relationship between emissions and plant conditions, machine learning is essential for optimizing combustion and reducing emissions. This study presents the development and validation of a Deep Sparse Transformer Network (DSTN) designed to forecast air pollutant time series with long-range dependencies. The dataset includes 67,527 records from a 4000 MW coal-fired power station in Ekibastuz, Kazakhstan, covering $\mathbf{1 4}$ variables at $\mathbf{1 0}$-minute intervals. Fractal R/S analysis revealed long-term memory patterns in NO2, NO, and PM10informing the DSTN’s input window size. The model achieved the highest prediction accuracy for NO2($\mathrm{R}^{2}$ between 0.95 and 0.38), with similar performance for NO and PM10($\mathrm{R}^{2}$ between 0.93 and 0.26). Accuracy declined sharply beyond 12 forecast steps, indicating the model is most effective for short-term predictions. These findings support the implementation of improved real-time emission control and contribute to efforts aimed at achieving carbon neutrality.