Modeling Heat Exchanger Using Neural Networks Totok R. Biyanto, M. Ramasamy*, H. Zabiri Chemical Engineering Department, Universiti Teknologi PETRONAS Bandar Seri Iskandar, 31750 Tronoh, Perak, MALAYSIA Abstract - Tools to predict the effects caused by frequent changes in the feedstock and in the operating condition in crude preheat train (CPT) in a refinery are essential to maintain optimal operating conditions in the heat exchanger. Currently, no such tools are used in industries. In this paper, an approach based on Nonlinear Auto Regressive with eXogenous input (NARX) type multi layer perceptron neural network model is proposed. This model serves as the prediction tool in order to determine the optimal operating conditions. The neural network model was developed using data collected from CPT in a refinery. In addition to the data on flow rates and temperatures of the streams in the heat exchanger, data on physico-chemical properties and crude blend were also included as input variables to the model. It was observed that the Root Mean Square Error (RMSE) during training and validation phases are less than 0.3 o C proving that the modeling approach employed in this research is suitable to capture the complex and nonlinear characteristics of the heat exchanger. Keywords: Neural Network, heat exchanger, modeling * I. INTRODUCTION Energy minimization has taken utmost importance in process industries due to the rising oil prices and fast depleting oil reservoirs. Furthermore, the impact on the environment by CO 2 released by fossil fuels makes it necessary to reduce CO 2 release by minimizing the energy usage. Fouling in Crude Preheat Train (CPT) in oil refineries is a very serious problem that consumes additional energy and affects the plant economy. Understanding/predicting the fouling characteristics in crude preheat train is imperative to operate the CPT in an optimal manner, with minimum fouling. However, fouling is largely determined by the crude/crude blend being processed in CPT, the temperatures and flow rates. The fouling mechanism is very complex and it is difficult to develop a fundamental model to predict the fouling rate for different crude blends and different operating conditions. Recently, neural networks have been shown to approximate nonlinear functions up to any desired * Corresponding author : Tel.: +60 5 368 7585; fax: +60 5 365 6176. E-mail address: marappagounder@petronas.com.my level of accuracy [1]. The technique has been applied to many thermal problems [2], including the prediction of the steady-state [3] and the dynamic behavior of heat exchangers [4, 5, 6]. Several authors have also used neural network as an alternative modeling method for the prediction of fouling [7]. Online fouling detection and estimation of the overall heat transfer coefficient (U) [8] were reported in literature. They are also less sensitive to noise and incomplete information than other approaches such as empirical models and correlations. Quick and reliable fouling rate prediction by the neural networks model is an added advantage with less computational load. Most of the fouling prediction models reported in literature are based on the operating conditions and they do not depend on the properties of the fluid being processed. Apart from the operating conditions, crude properties play an important role in the determination of fouling rates in heat exchangers [14]. In this paper, Multi Layer Perceptron (MLP) neural networks with Nonlinear Auto Regressive with eXogenous input (NARX) structure is used to model a heat exchanger in refinery CPT. Section II explains the theory of neural networks, especially, the NARX structure. Modeling of the heat exchanger using neural networks is explained in Section III. Simulation results and discussions are presented in Section IV. II. NARX-TYPE NEURAL NETWORKS Multi Layer Perceptron (MLP) with feed forward architecture is the most used and studied neural network architecture today. It models a global approximation of a multi-input multi-output function in a similar manner as fitting a low order polynomial through a set of data points. A rich collection of different learning algorithms are available in literature [9, 10]. The MLP network is selected as the basic building block to be used in this study. The mathematical formula describing the MLP networks takes the form:                        h n j i n l j l l j j j i i i W w w f W F y 1 0 , 1 0 , , , .   (1)