An ANN model to predict oil recovery from a 5-spot waterflood of a heterogeneous reservoir

Abstract

Waterflooding is a secondary oil recovery technique in which water is injected into an underground oil reservoir to maintain the reservoir pressure and boost oil recovery. The performance of a waterflood depends on several factors such as reservoir heterogeneity, reservoir fluid properties, flood pattern, etc. Most of the models developed to predict waterflood performance are either for linear systems or involve simplified assumptions for non-linear systems. In this study, we propose a novel, artificial neural network (ANN) model comprised of two hidden layers with 256 neurons each for the performance prediction of a 5-spot pattern waterflood in a heterogeneous reservoir at and beyond water breakthrough. The proposed model can be applied to estimate movable oil recovery efficiency of the waterflood (RFM) as a function of Dykstra-Parsons permeability variation coefficient (V), mobility ratio (M), permeability anisotropy ratio (kz/kx), production water cut (fw), a simple indicator of wettability (WI), and oil/water density ratio (DR) within reasonable accuracy. The MAPE of the proposed model was ∼4% and ∼5% using training and testing data, respectively. Our ANN model recommendation is based on a detailed comparative study against other popular soft computing models, such as adaptive neuro-fuzzy inference system (ANFIS) and support vector regression (SVR). Based on the accuracy and computational efficiency, the ANN model outperforms ANFIS and SVR. AIC and BIC of the proposed ANN model were also the lowest among all applied soft computing tools. The proposed model is tested on two real field cases and compared with a semi-analytical model and an empirical correlation. The presented model shows good agreement with the real field data. The trained ANN model, proposed here, saves computational time in forecasting the waterflood performance compared to a reservoir simulator.