(243j) PSO-Optimized BP Algorithm for Prediction of Gasoline Yield of FCC Unit

The on-line prediction of the gasoline yield is important for the catalytic cracking process in the refinery. The conventional algorithm based on the BP neural network can only deal with few data from the refinery and its ability in handling large amount of data is limited and large errors may be introduced. The present work suggested an advanced BP model that is optimized with PSO based on the conventional theory of the BP neural network. The present model is able to perform predictions based on the big data from a refinery through continually modifying the weights and threshold of the BP neural network in order to continually reduce the output error. The present work collected 160,000 groups of data and selected 17 factors (including the properties of the raw material and catalyst and some basic data relating to the operation) that may influence the gasoline yield of the MIP FCC unit from a refinery with a capacity of processing 1.2 million tons of oil per year. The above data was first pretreated and then normalized. The normalized data was set as the initial value and the actual data as the desired value. The conventional BP model and the PSO-optimized BP model were separately applied to predict the gasoline yield. Finally, part of the pretreated data directly collected from the unit was set as the predicted value, which was incorporated into the above two models in order to obtain the predicted value of the gasoline yield. The present work found that the conventional BP algorithm predicted the gasoline yield with a mean squared error of 5.667 while the PSO-optimized BP algorithm predicted the gasoline yield with a mean squared error of 4.644, which indicates that the PSO-optimized BP algorithm is better in predicting the gasoline yield of the catalytic cracking process.