A novel hybrid forecasting approach for NOx emission of coal-fired boiler combined with CEEMDAN and self-attention improved by LSTM

The precise prediction of NOx generation concentration in coal-fired boilers serves as the foundational cornerstone for the judicious optimization and control of selective catalytic reduction denitrification (SCR) systems. Owing to the intricate nature of the denitrification process within SCR, there exists a temporal delay in regulating the ammonia injection rate based on the monitored data of NOx concentration at the SCR inlet. Such delays can give rise to ammonia leakage and subsequent obstruction of the air preheater. In light of this, a predictive model, CEEMDAN-LSTM-SA, is proposed as an amalgamation of data decomposition and the LSTM (long short-term memory) fusion self-attention mechanism within a deep learning network, which is introduced to forecast the NOx emission concentration at the SCR inlet of coal-fired units. To mitigate the impact of data outliers on the training effectiveness of the model, a clustering method coupled with a statistical testing strategy is initially applied to refine the dataset first. CEEMDAN data decomposition technology is leveraged to facilitate the breakdown of data, alleviating its non-stationary and intricate characteristics. Subsequently, through spectral analysis, the decomposed components are grouped and aggregated to form novel data elements, which are then subjected to prediction by the constructed LSTM-SA deep learning network. The ultimate NOx emission concentration prediction value is derived through a process of fusion. Upon scrutinizing and comparing the predictions derived from various models using coal-fired power plant data, it is evident that the performance metrics of CEEMDAN-LSTM-SA predictions exhibit a mean absolute error of 7.425, mean absolute percentage error of 2.415%, root mean square error of 9.715, R-squared (R²) value of .789, mean absolute relative error of 2.109%, and a Theil's information criterion of .016. In contrast to other models, including traditional self-attention networks, LSTM, and LSTM-SA combination networks, CEEMDAN-LSTM-SA proposed in this study demonstrates superior prediction accuracy and enhanced generalization capabilities. Consequently, this predictive model stands poised to furnish an efficacious framework for the SCR ammonia injection strategy within thermal power units.