Data-driven modeling for the prediction of stack gas concentration in a coal-fired power plant in Türkiye

Abstract

In this research, deep learning and machine learning methods were employed to forecast the levels of stack gas concentrations in a coal-fired power plant situated in Türkiye. Real-time data collected from continuous emission monitoring systems (CEMS) serves as the basis for the predictions. The dataset includes measurements of carbon monoxide (CO), sulfur dioxide (SO2), nitrogen oxides (NOx), oxygen (O2), and dust levels, along with temperatures recorded. For this analysis, deep learning methods such as multi-layer perceptron network (MLP) and long short-term memory (LSTM) models were used, while machine learning techniques included light gradient boosted machine (LightGBM) and stochastic gradient descent (SGD) models were applied. The accuracy of the models was determined by analysing their performance using mean absolute error (MAE), root means square error (RMSE), and R-squared values. Based on the results, LightGBM achieved the highest R-squared (0.85) for O2 predictions, highlighting its variance-capturing ability. LSTM excelled in NOx (R-squared 0.87) and SO2 (R-squared 0.85) prediction, while showing the top R-squared (0.67) for CO. Both LSTM and LGBM achieved R-squared values of 0.78 for dust levels, indicating strong variance explanation. Conclusively, our findings highlight LSTM as the most effective approach for stack gas concentration forecasting, closely followed by the good performance of LightGBM. The importance of these results lies in their potential to effectively manage emissions in coal-fired power plants, thereby improving both environmental and operational aspects.