Dynamic prediction of WFGD outlet SO2 concentration using optimal input subset generation and improved Bayesian-optimized BiLSTM

Accurate prediction of sulfur dioxide (SO 2 ) concentration at the outlet of wet flue gas desulfurization (WFGD) systems is essential for emission-risk warning, proactive control, and energy-efficient operation in coal-fired power plants. In this study, an integrated dynamic prediction framework is proposed by combining optimal input subset generation, a tree-structured Parzen estimator (TPE)-based improved Bayesian optimization algorithm, and bidirectional long short-term memory (BiLSTM). Refined inputs are constructed by integrating maximum relevance minimum redundancy (mRMR), canonical time-series characteristics (CATCH22), and SHapley additive explanations (SHAP), while the BiLSTM architecture and training parameters are optimized using the improved TPE-based strategy. Validation using 10,080 1-min samples from an industrial WFGD system showed that the proposed model achieved a coefficient of determination (R 2 ) of 0.9871, root mean square error (RMSE) of 0.628 mg/Nm 3 , mean absolute percentage error (MAPE) of 4.41 %, and mean absolute error (MAE) of 0.348 mg/Nm 3 on the independent test set. Compared with bidirectional gated recurrent unit (BiGRU), recurrent neural network (RNN), multilayer perceptron (MLP), Transformer, and extreme gradient boosting (XGBoost), RMSE was reduced by 36.39 %, 56.52 %, 64.32 %, 55.20 %, and 44.20 %, respectively. Ablation results showed that removing input subset generation, Bayesian optimization, or bidirectional encoding increased RMSE to 2.738, 1.255, and 0.950 mg/Nm 3 , respectively. External validation on datasets from different power plants yielded test R 2 values of 0.9522 and 0.9800, demonstrating cross-plant applicability after retraining. These results support early SO 2 emission warning and proactive WFGD operation adjustment, while future work will focus on online adaptive updating under sensor drift and varying loads.