Low-temperature desulfurization forecasting using soft computing models

Abstract

Flue Gas Desulfurization (FGD) is pivotal in reducing Sulfur Dioxide (SO2) concentrations through neutralization. This study explored dry FGD modeling using Artificial Neural Networks (ANN) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS). The independent parameters used were diatomite to Ca(OH)2 ratio, hydration time, hydration temperature, SO2 concentration, and sulfation temperature, while the output responses incorporated were sulfation efficiency () and sorbent conversion (). ANN simulations employed the Levenberg-Marquardt (LM), Bayesian Regularization (BR), and Scaled Conjugate Gradient (SCG) algorithms with 7 to 10 hidden cells. The sigmoid and linear functions served as trigger mechanisms. ANFIS models, utilizing grid partitioning and subtractive clustering, were trained with hybrid and backpropagation methods. Seven ANFIS membership functions were compared for the best-fit model. The computing models were critiqued using RMSE, MSE, and R2 statistical metrics. Numerical error analysis favored the ANN program, with BR exhibiting the highest R2 values (0.9987 for , 0.9986 for ). However, the SCG algorithm emerged as the most dependable model due to its lowest RMSE and MSE values. In contrast, the ANFIS model demonstrated inferior R2 values and forecasting capabilities. This investigation provided nuanced insights into dry FGD modeling, elucidating the interplay between computational methodologies and process parameters.

Keywords

• deep learning
• desulfurization
• neural networks
• fuzzy logic systems
• emission control