A new correlation for predicting hydrate formation conditions for ...

( )A.A. Elgibaly, A.M. ElkamelrFluid Phase Equilibria 152 1998 23–42 33wx wx wx wx5 , Katz 3 and Kobayashi et al. 7 are accurate over a limited data range, the Makogon 6correlation gave the least accuracy for the examined data. The ANN model-A provided reasonablyaccurate results over a wide range of hydrate formation conditions.The cross plot shown in Fig. 5, compares the measured hydrate formation pressure with thatpredicted by ANN model-A. The data are presented for various phase equilibria of pure hydrocarboncomponents and their mixtures. As can be seen, relatively little scatter is obtained around the 458-line.This indicates a good agreement between the measured and calculated results.3.2. ANN compositional modelsIn the next phase of this study, three ANN models were developed to consider the relationshipsindicated in Eq. Ž 9. through Eq. Ž 11 .. Several neural network architectures were attempted to find outthe best accuracy. A one-hidden layer network was found to be suitable. The number of neurons inthe hidden layer was varied until a minimum sum squared error was obtained. Table 3 demonstratesthe final neural network properties and statistical parameters of these models. The number of neuronstabulated for the different models delivered acceptable results. They were justified by the correspondingcross plot verifications Ž Figs. 6–8 ..The ANN model-B was developed to include the composition of the hydrocarbon hydrate-formersas indicated in Eq. Ž 11 .. The model showed lower average error and sum-squares error than those ofthe ANN model-A. This may be due to the lack of data for pure n-butane, though data on its mixturewith other components were available w31 x. The model is capable to predict the hydrate formationpressure, which fitted well the experimental pressure for the methane–ethane system at differentproportions as shown in Fig. 9.Fig. 6. Prediction of hydrate formation conditions by ANN model-B for methane–ethane mixture.