The degree of success of many oil and gas drilling, completion, and production activities depends upon the accuracy of the models used in a reservoir description. Neural networks have a number of attractive features that can help to extract and recognize underlying patterns, structures, and relationships among data. However, before developing a neural network model, we must solve the problem of dimensionality such as determining dominant and irrelevant variables. We can apply principal component and factor analyses to reduce the dimensionality and help the neural networks formulate more realistic models.

We have used an intelligent software, Oilfield Intelligence (OI)1 , as an engineering tool to improve the characterization of oil and gas reservoirs. OI integrates neural networks and multivariate statistical analysis. It is composed of five main subsystems: data input, preprocessing, architecture design, graphics design, and inference engine modules. We validated OI by obtaining confident models in three different oil field problems: (1) a neural network in-situ stress model using lithology and gamma ray logs for the Travis Peak formation of east Texas1 , (2) a neural network permeability model using porosity and gamma ray and a neural network pseudo-gamma ray log model using 3D seismic attributes for the reservoir VLE 196 Lamar field located in Block V of south-central Lake Maracaibo (Venezuela), and (3) neural network primary ultimate oil recovery (PRUR), initial waterflooding ultimate oil recovery (IWUR), and infill drilling ultimate oil recovery (IDUR) models using reservoir parameters for San Andres and Clearfork carbonate formations in west Texas2 .

This paper presents the models that we developed using OI in C-4 and C-5 for the VLE 196 Lamar Field, Block V in Lake Maracaibo. We found that the best neural network permeability model should have as input variables porosity and gamma ray log. The coefficient of determination of that model is 0.9756 with an average absolute error of 18.03%. From non-linear regression and non-parametric approach methods, the absolute error of these models varies from 38.1 to 97.78%. For the same reservoir, we developed a high-quality, multivariate, and neural network model to predict the pseudo-gamma ray log from 3D seismic attributes. The data points used during the training, testing, and validation comprised approximately 11,800 points. Two dime/depth conversion tables were used to convert the arrival time of seismic to depth. We developed a Fortran-90 program to make that conversion. The confidence of that model is around 90% with an expected absolute error of 11%.

In all cases, we compared the neural network models with the non-linear regression and non-parametric approach models. The results show that one of the ways in the near future to get better reservoir characterization models is using neural networks in conjunction with the multivariate statistical techniques.

You can access this article if you purchase or spend a download.