Product Yields Prediction of Tehran Refinery Hydrocracking Unit Using Artificial Neural Networks

In this contribution Artificial Neural Network (ANN) modeling of the hydrocrackingprocess is presented. The input–output data for the training and simulation phases ofthe network were obtained from the Tehran refinery ISOMAX unit. Different networkdesigns were developed and their abilities were compared. Backpropagation, Elmanand RBF networks were used for modeling and simulation of the hydrocracking unit.The residual error (root mean squared difference), correlation coefficient and run timewere used as the criteria for judging the best network. The Backpropagation modelproved to be the best amongst the models considered. The trained networks predictedthe yields of products of the ISOMAX unit (diesel, kerosene, light naphtha and heavynaphtha) with good accuracy. The residual error (root mean squared difference)between the model predictions and plant data indicated that the validated model couldbe reliably used to simulate the ISOMAX unit. A four-lumped kinetic model was alsodeveloped and the kinetic parameters were optimized utilizing the plant data. The resultof the best ANN model was compared to the result of the kinetic model. The root meansquare values for the kinetic model were slightly better than the ANN model but theANN models are more versatile and more practical tools in such applications as faultdiagnosis and pattern recognition.