On the evaluation of asphaltene precipitation titration data: Modeling and data assessment

• The amount of asphaltene precipitation during titration experiments is successfully modeled using the LSSVM-CSA approach.
• The model inputs are: temperature, type of solvent, and solvent to oil dilution ratio.
• A large data bank is used covering a wide range of thermodynamic conditions and different types of crude oils.
• The proposed approach is successfully compared to pre-existing models.
• A mathematical model is used which evaluates the quality of experimental data and detects the probable outliers.

Asphaltene precipitation causes several problems during different stages of oil production in the reservoirs. Experimental measurement of asphaltene precipitation is cumbersome, expensive and tedious. In this communication, the amount of asphaltene precipitation during titration experiments was modeled as a function of easily measureable parameters including temperature, type of solvent, and solvent to oil dilution ratio. A large data bank of asphaltene precipitation was collected from different sources, covering a wide range of thermodynamic conditions and different types of crude oils. Least square support vector machine (LSSVM) optimized with a stochastic algorithm named couples simulated annealing (CSA) was employed for the purpose of modeling. The data bank was divided into four sections based on the type of solvent and solvent to oil dilution ratio. Subsequently, for each section a model was proposed and the results showed that all of the proposed models can predict the amount of asphaltene precipitation with enough accuracy. In general, the proposed CSA-LSSVM models can predict asphaltene precipitation with an average absolute relative error of 9.46%. The proposed models were compared to pre-existing models and both graphical and statistical analyses indicated the superiority of the proposed CSA-LSSVM models over the pre-existing ones. Finally, a mathematical model was used which not only defines the applicability domain of the proposed models, but also evaluates the quality of experimental data and detects the probable outliers. The results demonstrated that all of the proposed models are statistically valid and only 3.3% of the data may be recognized as the probable outliers.