Development of a neural network model for Sherwood number prediction for flow-accelerated corrosion in power boiler headers

Flow-accelerated corrosion (FAC) has regained attention due to recent findings of its occurrence in the harps of horizontal flue gas flow heat recovery steam generators (HRSGs), leading to unplanned outages and significant financial losses. Conventionally, a mass transfer-based corrosion rate model can accurately predict FAC rates. In this study, data generated from validated computational fluid dynamics (CFD) simulations, with mean absolute error of 0.108 mm/year, were used to train an artificial neural network (ANN) model to predict FAC rates in HRSG harps. A total of 249 data points were collected including both operating conditions and geometrical parameters and one output value which is the ratio of maximum Sherwood number observed in a harp to that of a straight pipe under the same operating conditions. First, the dataset was normalized to enhance training efficiency and split into training, validation, and testing sets. A single hidden layer ANN architecture was adopted, with the optimal configuration determined to include 16 neurons. The final model attains a root mean square error (RMSE) of 0.996 and a correlation coefficient (R) of 0.940, indicating strong predictive performance. This is the first model to provide as valuable tool for preliminary design checks and supporting proactive maintenance planning. • Artificial neural network for FAC prediction in headers of boilers is developed. • Total RMSE and the R value of the model are 0.996 and 0.940. • Model can be applied to the increasingly popular horizontal flue gas flow HRSGs.