Application of Artificial Intelligence Techniques in Estimating Oil Recovery Factor for Water Derive Sandy Reservoirs

Abstract

Abstract Oil and gas operating companies are always concerned with evaluating the reserve of their assets. Evaluation process of hydrocarbon reserves requires a full understanding and knowledge of technical and non-technical aspects regarding the nature of reservoir, available technology and economic conditions as well as others. Recovery factor (RF) is the most important parameter in evaluating the reserve of new fields. Several techniques are currently available for estimating oil recovery factor, the accuracy of those techniques are highly affected by data availability which is mainly related to the field age. Some of the techniques are highly accurate but they require lots of production data, hence, their applicability early in the reservoir life is restricted. Others could be applied earlier, but on the other hand, they have very low accuracy. In this paper ten parameters (original oil in place, asset area, net pay, initial reservoir pressure, porosity, permeability, Lorenz coefficient, API gravity, initial water saturation and oil viscosity), which are usually available early in the life of the reservoir, are used to estimate the oil recovery factor through application of four Artificial Intelligence (AI) techniques namely: artificial neuron networks (ANNs), Radial Basis Neuron Network (RNN), ANFIS-2 (Adaptive Neuro Fuzzy Inference System, Subtractive Clustering), and SVM (Support Vector Machines). Data from 130 sandstone reservoirs were used to learn the AI models, and then an empirical correlation was developed based on the ANN model. The suggested AI models and the developed ANN-based correlation were then tested in other 38 sandstone reservoirs. The results obtained showed that ANN-based correlation successfully predicted the recovery factor based on early time data only with absolute average percentage error (AAPE) of 7.92%, coefficient of determination (R2) of 0.9417, root mean square error (RMSE) and maximum absolute percent error (MAE) of 3.74 and 24.07%, respectively. ANN-based empirical correlation over-performed RNN, ANFIS-2, and SVM models in term of AAPE, MAE, and RMSE for testing set. Comparison of the recovery factor predicted by the developed equation with three available correlations showed that the developed equation predictability is about 5 times better that the most accurate correlation (of the currently available ones) in term of AAPE for predicted RF of the tested 38 reservoirs.