Ph Prediction in Desulfurization System Based on Attention Mechanism and Tcn-Gru Modeling

To address the challenges of inaccurate and timeconsuming pH measurement in the limestone-gypsum wet flue gas desulfurization (WFGD) system at thermal power plants, this study proposes a slurry <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{p H}$</tex> prediction method for absorption towers based on an attention mechanism and TCN-GRU hybrid model. First, operational data from the plant's Supervisory Information System (SIS) database were preprocessed, and correlation analysis was employed to identify relationships between variables. Subsequently, an attention mechanism (ATT) was applied to adaptively assign weights to pH -related input data, distinguishing strong and weak feature variables based on their weights, thereby resolving low prediction accuracy issues. Temporal Convolutional Network (TCN) was then utilized to extract multi-scale dynamic features (short-term and long-term patterns) through dilated causal convolutions, with residual connections enabling feature dimensionality reduction. The optimized temporal features were fed into a Gated Recurrent Unit (GRU) network, significantly accelerating neural network training while enhancing accuracy in modeling complex desulfurization dynamics. Experimental validation using operational data from a 350 MW unit demonstrated superior prediction accuracy and stability compared to mainstream algorithms. The integration of this model with Model Predictive Control (MPC) further confirmed its practical effectiveness in industrial applications.