Prediction of NOx emission from a 600 MW coal-fired boiler with deep learning: Plant data learned by Gini index and lightweight convolutional neural network

Abstract

Selective catalytic reduction (SCR) is a key technology for controlling pollutant emissions from coal-fired boilers. Accurate prediction of nitrogen oxide emissions at the SCR inlet helps improve denitrification efficiency. However, traditional prediction models are unable to efficiently establish dynamic mapping relationships between nitrogen oxide emissions and combustion control variables. To improve the prediction accuracy of NOx emissions, this study adopts a Deformable Convolutional Neural Network for Multimodal (DFC-CNN) to build a dynamic model. DFC-CNN improves the model ability to extract hidden features by modulating the deformable receptive field. To extract the key variables, the Gini index was used to measure the importance of the variables. To evaluate the prediction capabilities of DFC-CNN, a simulation experiment was executed using real historical data from a 600 MW coal-fired boiler. Comparative experiments between Back Propagation (BP) and Long Short-Term Memory (LSTM) showed that the prediction accuracy of DFC-CNN under transient conditions significantly outperformed the traditional models (such as BP and LSTM). Compared to the baseline CNN, the expanded receptive field and introducing the modulation factor significantly improved the prediction performance of the DFC-CNN. The Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) were 12.077 mg/m³, 8.628 mg/m³, and 0.018, while the coefficient of determination (R²) was 0.973. Compared to Computational fluid dynamics (CFD) models, which require several days to predict a working condition, DFC-CNN models only need a few seconds. The DFC-CNN-based modeling method can meet the demands of industrial production and support cleaner combustion coal-fired boilers.