Using artificial neural networks in the modeling of the hydrocracking process

A feed-forward neural network that models the hydrocracking process of Arabian light vacuum gas oil is presented. The input-output data to the neural network were obtained from actual local refineries. Several network architectures were tried and the network that best simulates the hydrocracking process was retained. The network is able to predict yields and properties of products of the hydrocracking unit (e.g. iC(4), nC(4), light and heavy naphtha, light and heavy ATK, Diesel, etc.). The predictions of yields and properties of various desired and undesired products at different conditions are required by refineries for process optimization, control, design, catalyst selection, and planning. The predictions of the prepared neural network have been cross validated against data not originally used in the training process. The network compared well against this new set of data with an average percent error always less than 6.76 for the different products of the hydrocracking unit.