BiLSTM-Multihead Attention Hybrid Model for Outlet NOx Concentration Forecasting in SCR Systems of Thermal Power Plants

Accurate prediction of NOx concentration at the exit of the denitrification process in thermal power plants is extremely important for controlling NOx concentration emissions. To address the problem of low prediction accuracy due to nonlinearity and large hysteresis in the denitrification process of power plants, this study develops a hybrid deep learning framework that synergistically combines a bidirectional long and short-term memory (BiLSTM) network and a multiple attention (MHA) mechanism for real-time monitoring of nitrogen oxides (NOx) emissions at the outlet of selective catalytic reduction (SCR) systems in coal-fired power plants. The prediction method adopts BiLSTM network to effectively explore the dynamic coupling relationship of the operating parameters of coal-fired units, and at the same time, the parallel processing capability of MHA is used to achieve the co-optimisation of the dynamic weights of the temporal features. The predicted performance of BiLSTM-MHA has been verified in a SCR denitrification system at a 350 MW thermal power plant in China. BiLSTM-MHA was compared with Backpropagation Neural Network (BP), Long Short-Term Memory (LSTM) and BiLSTM. The experimental results show that the BiLSTM-MHA model significantly reduces the prediction error and has higher accuracy. The RMSE, MAE, MAPE, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{R}^{2}$</tex> of the model for predicting the exported NOx concentration were 0.263, 0.215, 0.84%, and 0.991, respectively, and the R2 was improved by 26.81%, 4.82%, and 3.68%, respectively, with respect to the comparison model.