Data-driven Modeling of an Industrial Ethylene Oxide Plant: Superstructure-based Optimal Design for Artificial Neural Networks

Abstract

Optimum selection of input variables, number of hidden neurons and connections between the network elements delivers the best configuration of an artificial neural network (ANN), resulting in reduced over-fitting and improved performance. In this study, a superstructure-oriented ANN design and training algorithm is suggested and implemented on an industrial Ethylene Oxide (EO) plant for the prediction of product related variables (i.e. EO production rate). Proposed formulation is a mixed integer nonlinear programming problem (MINLP), which takes the existence of inputs, neurons and connections of the network into account by binary variables in addition to continuous weights of existing connections. Investigations show that almost 90% of the connections are removed compared to the fully connected ANN (FC-ANN) with 50% decrease in the number of inputs of the ANN, approximately. The modified ANN delivers a better prediction performance over FC-ANN, which suffers from over-fitting. © 2021 Elsevier B.V., All rights reserved.