Enhancing SO2 emission concentration prediction in ammonia flue gas desulfurization systems through continual learning

Abstract

Accurately predicting SO₂ concentration at the outlet of ammonia flue gas desulfurization (FGD) systems is crucial for optimizing process parameters in power plants, thereby improving desulfurization efficiency and reducing SO₂ emissions. However, modeling the mechanism of the ammonia FGD system is challenging due to the large number of interacting components. Additionally, existing data-driven deep learning models often struggle with long-term online predictions as the forecasting horizon extends and new data is introduced. To address these challenges, we propose a novel continual learning approach, termed Experience and Model Parameters Replay (EMPR). By integrating EMPR with the iTransformer model, an online prediction model called iTransformer-EMPR is developed to predict SO₂ emission dynamics in ammonia FGD systems. This model utilizes variate tokens to capture correlations among multiple feature variables within the desulfurization system and applies the EMPR algorithm for dynamic model updates. Using actual data from the combustion boiler at the Karamay Petrochemical Co. thermal power plant in Xinjiang, a case study demonstrates the model’s effectiveness in predicting SO₂ emissions 10 minutes in advance. The experiment results show that the iTransformer-EMPR model significantly outperforms existing approaches, reducing RMSE (Root Mean Square Error) by 69.0 %, 63.1 %, and 10.2 % compared to LSTM, Informer, and iTransformer models, respectively. Furthermore, compared to continual learning algorithms MAS and DER+ +, RMSE reductions of 7.8 % and 5.9 % are achieved. Additionally, the EMPR algorithm demonstrates strong generalization capability, reducing RMSE by 6.8 % and 8.4 % when applied to LSTM and Informer models, respectively.