A predictive model with time-varying delays employing channel equalization convolutional neural network for NOx emissions in flexible power generation

This study aims to predict NOx (i.e. NO + NO2) emissions at the selective catalytic reduction inlet from coal-fired boilers. It develops a lightweight predictive model that can be adapted to the generation system dominated by renewable energy with fluctuating power. The random forest algorithm was first used to select the input variables from the auxiliary variables. The maximum information coefficient algorithm was then used to estimate the time delay between the input variables and the NOx emissions. Afterwards, to overcome the “channel collapse” issue in convolutional neural networks (CNNs), a Channel Equalization block (CE-Block) was proposed to activate the suppressed channels, and a lightweight network, referred to as CE-CNN, was designed by fusing a CNN with the CE-Block. Finally, a dynamic NOx emissions prediction model was developed employing the CE-CNN, and the model was evaluated using real data from a 600 MW down-fired boiler. The results showed that the dimensionality reduction policy decreased the model training time by 21.74 %, and the analysis of the time delay decreased the RMSE by 7.31 %. Compared to the baseline model 3-layer CNN, the proposed model achieves performance improvements of 20.97 %, 22.11 %, and 3.48 % in terms of RMSE, MAE, and R2, respectively.