II-22 Radial Basis Function - Neural Network Based Identification for Aqueous Ammonia Binary Distillation Column Yuliati 1) , Satriyo Nugroho 2) 1) Electrical Engineering Department Widya Mandala Surabaya Catholic University Jl. Kalijudan 37 Surabaya, Indonesia Phone: +62-31-3891264, ext 467 Email : yuliatheresia@yahoo.com ; yulia@mail.wima.ac.id 2) Maintenance Department PT. Petrokimia Gresik Jl. Ahmad Yani 1 Gresik 61101, Indonesia Phone: +62-31-3981811, ext 2321 Email : satriyon@petrokimia-gresik.com Abstract Ammonia is a useful chemical fluid for many sectors of industry. Many processes are involved along its production process, i.e. desulphurization process, reforming process, synthesize process, etc. One of most important is ammonia purge gas recovery system. It includes high pressure gas scrubber, low pressure ammonia scrubber and ammonia stripper. Ammonia striper is an aqueous ammonia binary distillation column. Its dynamic characteristic is difficult to be identified due to the non-linear behaviour and is easily affected by many factors. The success of model-based non-linear control technique is usually conditioned by the availability of accurate models which reflects the non-linear process complexities. Prior to implementing any control scheme, its dynamic properties need to be identified properly. Owing to their good approximating properties and their simple topological structure, Radial Basis Function- Neural Network (RBF-NN) will be applied to identify the dynamics of an ammonia stripper. An experiment for identifying the ammonia stripper from real- time data was conducted on real-time plant operation of a fertilizer plant in Gresik, East Java, Indonesia. Validation model was also performed to verify the results of identification. The performance of the procedure shows how the RBFNN algorithm produces accurate models for non-linear control design. Keywords: non-linear identification, model- based control, binary distillation column, radial basis function, neural network, real- time experiments. 1. Introduction During the past few years, the non-linear dynamic modelling of processes by neural network has been extensively studied. In standard neural network, the non-linearity is approximated by superposition of sigmoid functions. Several studies have also reported the use of neural networks for the purpose of non-linear system identification and control [Narendra and Parthasarathy, 1990, 1992]. However, there are some problems that normally associate with this method such as parameters convergence and heavy computation. Several authors (Powell, 1987; Moody and Darken, 1989) have been proved that Radial basis function neural network (RBF-NN) overcome these problems. RBF- NN was introduced as a class of single-hidden layer feed-forward neural network using radial functions as activation functions whose architectures consist of simple processing elements; each having a number of inputs and one single output, operating in parallel and connected together in a layer structure where the input feeds forward through the network layers to the output. Due to simpler topological structure and faster learning capabilities, RBF- NN offer good alternative approximation and in the growing interest and increased research activities in identification and control. However, the main problem with RBF-NN is in determining the quantity and distribution of the basis function centres in the networks input space and estimating the weights and width of the network. In this paper, an alternative approach to model the binary distillation column system using RBF-NN will be presented. As other binary distillation column, an ammonia stripper consists of a re-boiler, a condenser and a reflux accumulator. It separates the binary mixture, namely aqueous ammonia into two products, ammonia (distillate) and water (bottom product) in the desired purity. As shown in the figure 1, the feed flow comes from an ammonia scrubber, which has 14.1% wt ammonia.