Modeling the corrosion rate of carbon steel in carbonated mixtures of MDEA-based solutions using artificial neural network

Abstract

Amine-based CO2 absorption process has the benefit of purifying produced natural gas and reducing greenhouse gas emissions; however, it is associated with corrosion issues. This work aims to develop a corrosion prediction model for carbon steel in methyldiethanolamine (MDEA)-based binary mixtures with monoethanolamine (MEA), diethanolamine (DEA), or piperazine (PZ) at various concentrations using artificial neural network (ANN). Experimental studies of Q345R steel are performed, and corrosion rate data obtained by weight loss method is used to create a database for training and testing of the ANN model. A backpropagation (BP) multilayer perceptron (MLP) network with three layers is proposed. The number of input variables for the input layer is optimized after performing a correlation analysis. Effect of neuron quantity in the hidden layer on ANN model performance is studied; increasing the neuron quantity in the hidden layer is found to enhance the training accuracy and reduce the testing accuracy. A new index, max absolute relative deviation (MARD), is introduced to quantify the performance of the ANN model. The developed 5-8-1 type ANN model is able to give a MARD of 8.66 %. The same corrosion rate database is used to develop the SVM model, in which radial basis function (RBF) is used as the kernel function, and K-fold cross-validation technique is applied to select the optimal model values. A comparison of performance in both training and testing shows that the ANN model outperforms the SVM model.