Abstract
This paper presents the research work on using a machine learning algorithm to predict the viscosity of Niger Delta oil reservoirs using formation volume factor and fluid density at bubble point pressure as correlating parameters.
Oil Viscosity stands out when considering the amount of oil recoverable from the reservoir hence it is an important input into the recovery factor computation, material balance analysis, reservoir simulation/history match, EOR evaluations and many other applications.
Laboratory techniques of obtaining oil viscosity are quite expensive and time consuming, hence the need for various mathematical correlations developed for its estimation. Majority of the correlations make use of empirical and experimental relationships developed from analyzing oil samples to obtain a trend to predict viscosity mostly for a basin. None of these has been developed for oil viscosity for Niger Delta fluids.
Viscosity has been globally defined as the resistance to shear stresses within the fluid or the resistance of the fluid molecules to deformation.
For a typical reservoir fluid system, where the liquid and gas exist in dynamic equilibrium, reservoir fluid composition along with temperature and pressure has been established to determine reservoir fluid viscosity1. Hence for an isothermal system and at a defined pressure in the reservoir the viscosity will be dependent on largely the composition. The reservoir fluid composition is also represented by the reservoir fluid density and the formation volume factor; therefore it is possible to deduce the viscosity of reservoir fluids from the oil density and formation volume factor even though a direct relationship has not been established between these parameters. Therefore, a correlation that can establish a relationship between the specific gravity (density) and FVF with viscosity will have significant value in the oil and industry.
The data used for this analysis includes viscosity, formation volume factor, oil density at 2800 sample bubble point pressure. The data was obtained by analyzing over 3500 PVT Analysis reports, extracting the data points using a python work program, cleaning up the data and removing erroneous data, performing preliminary analysis to establish baseline relationships between the data.
Supervised learning using a classification tree model was used as the machine learning approach. Seven different machine learning algorithms were reviewed, and the Random Forest Regressor was selected as the most suitable algorithm for the prediction.
The model prediction results were quiet encouraging as the model was able to predict viscosity within 10% deviation from the experimental viscosity for over 80% of the cases resulting in about 90% prediction accuracy. The analysis of the results further revealed that the model could better predict viscosity of Medium to Light oil with an R2 value of between 0.90-0.96 without adjusting some obvious erroneous data points.
Future of this research work will involve further in-depth analysis which will merge the preliminary QC plots with the results to evaluate the effect of the outlier sample points on the final predictability of the model. Also explore other machine learning models to further improve predictability and be able to predict viscosity across other pressure values other than the bubble point pressure to capture viscosity along the producing life of the reservoir.
Multiphase flow modeling of asphaltene precipitation and deposition
