A data-driven approach for flow corrosion characteristic parameters prediction in an air cooler

A data-driven soft measurement method based on a multiunit back propagation neural network (MBPNN) is presented in this study. This model aims to estimate the characteristic parameters that can reflect the flow corrosion of the reactor effluent air cooler (REAC). Flow corrosion failure during the hydrogenation process presents a serious concern to the petrochemical industry. In this paper, a safety evaluation of flow corrosion failure for a petrochemical diesel hydrogenation unit is first carried out. During the investigation, it is found that there is a risk of NH4Cl crystallization at around 187 degrees C. Then, considering flow-induced corrosion and ammonium salt deposition, main characteristic parameters are determined, including NH4Cl crystallization temperature (T-C), air cooler tube bundle minimum and maximum flow rate (V-min and V-max), air cooler inlet liquid water content (C-W), and NH4HS concentration (C-A). Finally, the data-driven model based on a multiunit back propagation neural network (MBPNN) is constructed. An improved particle swarm optimization (PSO) approach is employed to initialize the main parameters of the model. Compared with a multioutput back propagation neural network (BPNN) model and MBPNN model without an optimization algorithm, the presented data-driven model is proved to have high accuracy, a fast convergence rate, and high reliability.