Comparing a linear transfer function-noise model and a neural network to model boiler bank fouling in a kraft recovery boiler, TAPPI Journal, July 2024

Application: This study shows that a significant amount of variability in boiler bank fouling can be modeled with linear models, which can lead to new control strategies. Additionally, those looking to apply neural networks to mill data can use the procedures in this study to properly train and benchmark neural network models.

Author: Jerry Ng; Gustavo M. De Almeida; Esa K. Vakkilainen; Yuri A. Laryshyn; and and Nikolai A. DeMartini

ABSTRACT: Boiler bank fouling reduces heat transfer efficiency in kraft recovery boilers. Here, we model the relationships between boiler parameters and boiler bank pressure drop, an indicator of fouling, based on recovery boiler operating data. We compared two models: an autoregressive integrated exogenous (ARIX) model and a feedforward neural network. The ARIX model better simulates boiler bank pressure drop compared to the neural network (R2 of 0.64 vs. 0.58). Based on the ARIX model, we identified six boiler parameters that significantly influence boiler bank fouling and their relative contributions. Finally, we demonstrate how the models can simulate boiler bank pressure drop given artificial perturbations in boiler parameters.