Predicting NOx and SO2 Emissions in CHP Systems with TEBA-Net for Sustainable Operations

This study presents TEBA-Net, an advanced deep learning model for simultaneous prediction of NOx and SO2 emissions from coal-fired combined heat and power (CHP) plants. The model integrates temporal convolutional networks (TCN), bidirectional gated recurrent units (BiGRU), and dual attention mechanisms to capture the complex nonlinear dynamics of combustion emissions effectively. High-frequency operational data collected from a Chinese CHP facility at 1-min intervals were analyzed to identify seven key predictive variables through rigorous feature selection methods. TEBA-Net performs better than eight baseline models, achieving exceptional prediction accuracy with R2 values of 0.9637 for NOx and 0.9340 for SO2. The model reduces root mean square error by 44%–62% relative to conventional approaches while maintaining robust performance across varying operating conditions. These improvements stem from the unique architecture of the model that combines local temporal feature extraction with long-term dependence modeling and intelligent feature weighting. The research further develops a dynamic chemical treatment strategy that utilizes the real-time emission predictions via TEBA-Net to optimize the injection rates of limestone slurry and ammonia for flue gas desulfurization and denitrification. This integrated approach enhances both environmental compliance and operational efficiency in CHP systems. This work provides a scalable framework for intelligent pollution management in energy systems by bridging advanced artificial intelligence techniques with practical emission control applications. The proposed solution offers significant potential for reducing the environmental impact of fossil fuel-based power generation while maintaining energy production efficiency. The methodology and findings contribute to developing cleaner and smarter energy infrastructure.