Asst. Professor, Chemical and Biological Engineering, Koç University & Koç University-TÜPRAŞ Energy Center (KUTEM)
Data-driven Modeling of an Industrial Ethylene Oxide Plant: Superstructure-based Optimal Design for Artificial Neural Networks
Ethylene oxide (EO), a captive product used for the production of ethylene glycols (MEG, DEG and TEG), is produced through elective oxidation of ethylene and oxygen in the Socar Turkey/Petkim Petrochemical plant in Turkey. The EO process takes place in fixed-bed multi-tubular reactors at a temperature and pressure range of 240–260 °C and 17–18 bar, respectively. Side reactions result in the reduction of EO yield and an increase in CO2 emission from the plant. From the points of both safety and optimal production, monitoring of the concentrations of feed and product streams is of great importance. The operating strategy of the reactors is to maximize the EO yield under various operational constraints.
Real-time optimization of such a complex process is a challenging task, requiring significant effort for first principles modeling. Alternatively, Artificial Neural Networks (ANNs) are promising empirical models to calculate the product related variables from measurable variables (i.e. temperatures and pressures). The standard application of ANNs includes using fully connected networks, where all inputs, neurons and outputs are entirely connected. Fully connected ANN architectures (FC-ANNs) have high number of parameters. The number of these parameters increases extremely when more hidden neurons are introduced. Once the number of hidden neurons is high, the identification of corresponding parameters becomes even more challenging due to multiple solutions and over-fitting (). Introducing more data is usually not a satisfactory effort as new data do not carry additional statistical information unless they are collected from a different location in the plant.
Optimum selection of input variables, number of hidden neurons and connections of the network elements gives the best configuration of an artificial neural network, resulting in reduced over-fitting (). In this study, a superstructure-oriented ANN design and training algorithm is suggested and implemented on an industrial Ethylene Oxide plant for the prediction of product related variables (i.e. EO production rate). The formulation is a mixed integer nonlinear problem (MINLP), taking the existence of inputs, neurons and connections of the network into account by binary variables in addition to continuous weights of existing neurons. This problem is solved via a quasi-decomposition algorithm using an evolutionary integer programming algorithm and Ipopt (reference).
Investigations show that almost 80% of the connections are removed compared to the FC-ANN with 30% decrease in the number of inputs and 50% decrease in the number of hidden neurons. The modified ANN’s performance delivers better prediction performance over FC-ANN which suffers from over-fitting. Also, the modified ANN architecture parameters are easier to update due to less number of parameters. Furthermore, this decrease in the model structure results in less computational load for real-time applications. Our future work includes the application of the obtained empirical ANN models for real time optimization of this industrial plant.
Erdal Aydin is an assistant professor of Chemical and Biological Engineering at Koç University and Koç University-TÜPRAŞ Energy Center (KUTEM). He is also the recipient of the Tubitak 2232: International Fellowship for Outstanding Researchers grant. His research interests are in the areas of energy systems engineering, decision making under uncertainty, artificial intelligence, dynamic optimization and process optimization. He joined Koç University in 2021 after working as an assistant professor at Bogazici University and as a postdoctoral research associate at MIT. He received his Ph.D from the Max Planck Institute Process Systems Engineering department. He holds a BS diploma in Chemical Engineering from METU and an MS diploma in Chemical and Biological Engineering from Koc University.