DEVELOPMENT OF TIME SERIES FORECASTING MODELS FOR AIR POLLUTION BASED ON DEEP SPARSE TRANSFORMER NETWORKS

Abstract

This study investigates the application of fractal analysis and deep learning methods for forecasting pollutant emissions from the Ekibastuz coal-fired power plant. The research is based on time series of NO, NO₂, and PM₁₀ concentrations collected by industrial sensors during 2023–2024. To assess long-term dependencies, an R/S analysis was performed, and the results demonstrated stable persistence with average Hurst exponent values exceeding 0.67. This confirmed the appropriateness of employing models capable of capturing long-range memory in the data. In the second stage, a Deep Sparse Transformer Network (DSTN) architecture was implemented and adapted to the task of emission forecasting under different boiler operating modes. DSTN combines the advantages of transformer-based models with a sparse attention mechanism, which reduces computational complexity and enables efficient handling of long sequences. The model was trained using the PyTorch framework on a dataset of more than 67,000 records, with forecasting performed at horizons of 1, 6, 12, and 24 steps. The highest accuracy was achieved for short-term forecasts: the coefficient of determination for NO₂ reached 0.95 at a one-step horizon and decreased to 0.38 at 24 steps. For NO and PM₁₀, R² values ranged from 0.93 to 0.26. These findings indicate that DSTN is a highly effective tool for short-term forecasting but less accurate at longer horizons due to error accumulation. The results confirm the practical value of integrating fractal analysis with transformer architectures for emission monitoring and coal power plant operation management. The proposed approach can be embedded into industrial control systems to enable timely responses to peak emissions, optimize combustion modes, and mitigate environmental risks.