Estimation of asphaltene precipitation in light, medium and heavy oils : experimental study and neural network modeling

Abstract

Asphaltene can precipitate in oil reservoirs as a result of natural depletion and/or gas injection crippling the oil production performance. Most of the conventional models for asphaltene precipitation cannot precisely capture the asphaltene precipitation at a wide pressure range and for different oil types. To have a precise model that can be used for various oil types at a wide range of pressure conditions, a comprehensive artificial neural network (ANN) model was proposed to estimate the weight percent of precipitated asphaltene in different oil types (three oil types, namely light, medium and heavy). The dilution ratio, pressure, molecular weight of solvent, API gravity and resin-to-asphaltene ratio were considered as the model input parameters. The oil samples were thus categorized based on the differences in their API gravity and resin-to-asphaltene ratio. Five hundred and fifty experimental precipitation datapoints were obtained from our experimental apparatus in a wide range of pressure, dilution ratio and injected fluid molecular weight, and used to make a comprehensive databank for model calibration and verification. At the test stage, the coefficient of correlation (R2) was higher than 0.98 and mean square error was less than 0.04 indicating the good performance of the proposed model. Furthermore, a comparison between the prediction of ANN model and two types of alternative approaches, namely the thermodynamic and the fractal/aggregation approaches, was performed. For this purpose, the prediction of two of the widely used solubility models, Flory–Huggins and Modified Flory–Huggins and also a polydisperse thermodynamic model was compared to the prediction of the proposed ANN model. In addition to those, as a fractal/aggregation model, a scaling model was also selected and employed to compare its performance against that of the proposed ANN model. The ANN model showed a better performance as compared to the other conventional models. The results demonstrated that the proposed model provides acceptable prediction for different oil types over a wide range of pressure which is a difficult task for most of the conventional techniques.