Prediction control of CO2 capture in coal-fired power plants based on ERIME-optimized CNN-LSTM-multi-head-attention

Abstract

Predicting CO₂ concentration in post-combustion carbon capture (PCC) systems is challenging due to complex operating conditions and multivariate interactions. This study proposes an enhanced RIME algorithm (ERIME) optimization-based convolutional neural network (CNN)-long short-term memory (LSTM)-multi-head-attention (ECLMA) model to improve prediction accuracy. The local outlier factor (LOF) algorithm was used to remove noise from the data, while mutual information (MI) determined time lags, and the smoothed clipped absolute deviation (SCAD) method optimized feature selection. The CNN-LSTM-multi-head-attention model extracts meaningful features from time series data, and parameters are optimized using the ERIME algorithm. Using a simulated dataset from a 600 MW supercritical coal-fired power plant, the results showed that after LOF outlier removal, root mean square error (RMSE) and mean absolute error (MAE) improved by 10%–13%. Post-MI delay reconstruction reduced RMSE to 0.00999 and MAE to 11.6937, with R² rising to 0.9929. After variable selection, RMSE and MAE further reduced to 0.00907 and 9.9697, with R² increasing to 0.9983. After ERIME optimization, the ECLMA model outperformed traditional models, reducing RMSE and MAE by up to 91.55% and 84.94%, respectively, compared to CNN, and by 85.91% and 69.47%, respectively, compared to LSTM. These results confirm the model's superior accuracy and stability.