Neural Networks Models for Estimation of Fluid Properties

Abstract

Abstract Fluid viscosity is one of the most important parameters necessary to establish reservoir production and economical potential. Until the appearance of NMR techniques in the oil industry, oil viscosity determination was limited to laboratory tests and correlations with API gravity. Nuclear Magnetic Resonance is a technique based on the magnetic behavior of hydrogen nuclei. This behavior is the consequence of fluid properties, as viscosity and density, and its interactions with its surroundings. The use of NMR signals for estimation of fluid viscosity has been based mainly on correlations with single NMR parameters as logarithmic (T2log) and geometric averages (T2geo). However, qualitative analysis of NMR T2 distributions indicate that changes on NMR patterns translate on changes on viscosity, that sometimes are not reflected on the averages used. This fact brought the idea of developing a multivariable model, which considers the use of all points of the NMR T2 distribution to enhance fluid properties estimation. The use of neural network technique was identified and several models were developed. The models were developed using the T2 distribution and the cumulative T2 distribution. The model constructed based on the cumulative T2 distribution, showed a better prediction of oil viscosity, incrementing the correlation with real values from 64% using the T2log correlation to 87% with the neural network. This work was done on 24 oil samples from different Venezuelan fields covering a range from 6 to 700 cP. An additional model was developed selecting 15 samples from the same field covering a range from 25 to 72 cP. Comparing the results of the model vs. the estimation through T2log correlations, the prediction was enhanced from 89% to 99%, creating an excellent model for fluid viscosity determinations through NMR signals. Introduction Reservoir formation evaluation is based on estimation of properties as porosity, permeability and fluid saturation. This properties together with fluid properties, as viscosity and density, are used in the determination of the reservoir productivity. Well logs have been successful in the determination of porosity and fluid saturation, however, determination of permeability, viscosity and fluid density are still limited to the use of correlations or laboratory studies. Fluid viscosity is highly important on the determination of fluid mobility inside the reservoir, which defines the reservoir productivity index. Previous studies have shown that viscosity can be estimated through NMR studies1,2. However, viscosity estimation is still limited to the analysis of few parameters of the NMR signals, as the logarithmic average (T2log). On the other hand, qualitative analysis of NMR results indicates that there is a high influence of viscosity on the NMR T2 distribution shape, since different oil components are distributed along the T2 distribution. For this reason, the idea of use non-linear multivariable techniques is proposed to use all the information content in the T2 distribution curve on fluid properties estimation. Methodology NMR data was taken on oil free samples with a Maran Ultra* equipment for an interecho time of 300µs. NMR decay signal was processed to obtain NMR T2 distribution what was the base of the present study. Oil samples were selected from different fields of Venezuela, with oil viscosities varying from 7 cp to 700 cp (24 samples). Cumulative and derivative T2 distributions were determined, to assess the benefit of preprocessing the signal before doing the neural network analysis. The non-linear study was divided in two steps, data visualization and neural network modeling. Data visualization allows two study the relationship between points. The algorithm used can project all the data in two dimensions, keeping the geometrical properties of the original space. The clustering algorithm was applied to the T2 distributions of the samples, in the original, cumulative and derivative form. This exercise was made to identify which of the curve's form can provide more information about the fluid properties studied.