Data-Driven Modeling Approach to Predict the Recovery Performance of Low-Salinity Waterfloods

Low-salinity waterflooding (LSWF) has, in the past decade, attained a lot of attention to enhance oil recovery. In LSWF, diluted water is injected into an oil reservoir to improve oil recovery. The injected low-saline water changes the wettability of the reservoir, which leads to higher oil recovery. The recovery of an oil reservoir can be predicted from simulators, which are tedious, expensive, and time-consuming. Therefore, there is a need for a simple, quick, and inexpensive substitute to predict the oil recovery factor for low-salinity waterfloods. This paper presents a novel empirical correlation based on a feed-forward neural network to predict LSWF recovery efficiency in a heterogeneous reservoir at and beyond water breakthrough. The proposed model is valid for a broad range of dimensionless input parameters—degree of dilution of high saline water, mobility ratio, degree of reservoir heterogeneity, permeability anisotropy ratio, API gravity, and production water cut. The new empirical correlation was developed using 20,000 simulated data points obtained from simulation results to cover a wide range of input values. The LSWF simulation model was developed and validated with a model of a real carbonate reservoir located in the Madison formation in Wyoming. The artificial neural network (ANN) model parameters were optimized by conducting extensive sensitivities of ANN parameters (hidden layer neurons, training algorithms, and transfer functions). Moreover, an interesting trend analysis was conducted to validate the physical behavior of the ANN model, and a comparison with the unseen dataset was performed. To evaluate the performance of the newly developed correlation, three statistical indices were used, including the average absolute percentage error (AAPE). AAPE was 1.69% and 1.84% for the training and testing datasets, respectively. The proposed ANN model is limited to a single-stage, low-saline waterfloods for a 5-spot pattern.