Adaptive Weighted Deep Fusion Model for SO₂ Concentration Prediction at Desulfurization Outlet

To address the control hysteresis induced by delayed SO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> concentration measurements in desulfurization feedback systems, this study develops a predictive framework for desulfurization outlet SO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> concentration leveraging furnace operational parameters and desulfurization process variables. The proposed methodology synergistically integrates a convolutional neural network to extract localized feature interactions, a long short-term memory network to model temporal dependencies, and a LightGBM algorithm to process structured operational relationships, augmented by an adaptive weighting mechanism that dynamically optimizes model fusion based on real-time performance.Trained on historical operational data from a 330 MW coal-fired power unit, the framework achieves 120-second-ahead predictions of SO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> concentration. Experimental comparisons against standalone models demonstrate the superiority of the CNN-LSTM-LightGBM ensemble, attaining a R2 of 0.917, RMSE of 1.571, and MAPE of 5.86%. This integrated approach effectively resolves the limitations of single-model predictors in terms of insufficient prediction accuracy and weak generalization capability under fluctuating operating conditions.