NONLINEAR MODELING OF CRUDE PREHEAT TRAIN FOULING RATE VIA NEURAL-NETWORK

Fouling in heat exchangers tend to reduce the overall heat transfer coefficient. Two main impacts of fouling in Crude Preheat Train (CPT) operation are reduced heat recovery and increased pressure drop. Due to the inability of the existing monitoring tools to predict fouling rate, a heat exchanger predictive model is to be developed, whereby the predicted outlet temperatures are used to calculate the fouling. The main objectives of this study are, to develop a nonlinear model for CPT heat exchanger in a refinery using Neural-Network (NN) toolbox in MATLAB, and to optimize the model through a step by step strategy. The optimized model comprises 2 day-lagged feedback variables, with the tube integral flow normalized based on the whole range of Date Sets A and B, and Purelin-Logsig-Purelin configuration with 25-15-2 neurons in each layer. This model predicts the heat exchanger outlet temperatures with average Root Mean Square Error (RMSE) of 6.797, average correlation coefficient (R2) of 0.61 and average Correct Directional Change (CDC) of 76.84%. Furthermore, this optimized model shows even better predictive capability before the process drifting whereby the average RMSE is 3.79, average R2 0.7253 and average CDC 72.92%. To address the drift problem, adaptive training of NN is recommended to account for the changes in plant conditions.