Chemical Engineering Journal 147 (2009) 161–172 Contents lists available at ScienceDirect Chemical Engineering Journal journal homepage: www.elsevier.com/locate/cej Prediction of cell voltage and current efficiency in a lab scale chlor-alkali membrane cell based on support vector machines N. Shojai Kaveh a , F. Mohammadi b,∗ , S.N. Ashrafizadeh a a Research Lab for Advanced Separation Processes, Department of Chemical Engineering, Iran University of Science and Technology, Narmak, Tehran 16846, Iran b Iran Polymer and Petrochemical Institute, P.O. Box: 14965/115, Tehran, Iran article info Article history: Received 14 November 2007 Received in revised form 25 June 2008 Accepted 27 June 2008 Keywords: Chlor-alkali Membrane cell Brine Electrolysis Support vector machine abstract The main aim of this study is to investigate the impacts of operating parameters on the cell performance and predicting the same by SVM technique. This paper though introduces support vector machines (SVMs), a relatively new powerful machine learning method based on statistical learning theory (SLT), into cell voltage and current efficiency forecasting. In order to validate the model predictions, the effects of various operating parameters on the cell voltage and current efficiency of the membrane cell were experimentally investigated. The membrane cell included a standard DSA/Cl 2 electrode as the anode, a nickel electrode as the cathode and a Flemion 892 polymer film as the membrane. Each of six process parameters counting anolyte pH (2–5), operating temperature (25–90 ◦ C), electrolyte velocity (2.2–5.9 cm/s), brine concentra- tion (200–300 g/L), current density (1–4 kA/m 2 ), and run time were thoroughly studied at four levels for low caustic concentrations (5–22 g/L). The developed SVM model is not only capable to predict the cell voltage and caustic current efficiency (CCE) but also to reflect the impacts of process parameters on the same functions. The predicted cell voltages and current efficiencies using SVM modelling were found to be very close to the measured values, particularly at higher current densities. © 2008 Elsevier B.V. All rights reserved. 1. Introduction Chlor-alkali (CA) production is the major industrial scale electro synthesis; total annual capacity of about 55.6 million metric tons of chlorine world-wide [1]. The production of chlorine and caus- tic soda by electrolysis of aqueous solutions of sodium chloride or brine is one of the most important electrochemical processes, demanding high-energy consumption. The total energy require- ment is for instance 2% in the USA and 1% in Japan of the gross electric power generated to maintain this process by the chlor- alkali industry [2,3]. Significant improvement of the electrolytic process in this aspect (i.e., reduction in cell voltage) would be beneficial, both economically and environmentally. Cell voltage and current efficiency are the most important process parameters proportional to the power consumption of a CA plant. Therefore, process evaluation is important from industrial point of view in order to quantify the impact of process variables on these two parameters. At the same time, prediction of the cell voltage and current efficiency can facilitate achieving the optimum conditions ∗ Corresponding author. Tel.: +98 21 44580043. E-mail address: f.mohammadi@ippi.ac.ir (F. Mohammadi). as well as reducing the intercalary costs of trial and error experi- ments. There are different ways to predict and quantify these parame- ters such as statistical methods [4], analytical formulations [5] and non-parameter regression methods like artificial neural networks (ANNs) [26] and support vector machine (SVM). However, litera- ture does not show any published work on the application of SVM for such predictions in chlor-alkali industry, though the new SVM methods have already been applied to other fields [6–8]. Statistical methods are used to analyze the results of the exper- iments and models on response as well as to determine the contribution of each influencing factor. However, the main con- cern with statistical methods is the difficulties in fulfilling many rigid assumptions that are essential for justifying their applications, e.g. sample size, linearity, and continuity. One alternative approach for system predicting is the technique of SVM based on the struc- tural risk minimization (SRM) principle. Based on this principle, SVM achieves an optimum network structure by striking a right balance between the quality of the approximation of the given data and the complexity of the approximating function. The SVM reveals the underlying statistical relationships among variables corrupted by random error. This SVM algorithm presented by Vapnik [9], as other similar non-parametric statistical regression methods is 1385-8947/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.cej.2008.06.030