Data-driven modeling for forecasting oil recovery: A timeseries neural network approach for tertiary CO2 WAG EOR

Abstract

Attaining reliable forecasts in the petroleum industry is a constant challenge, prompting ongoing research into models that empower engineers to make informed decisions. The conventional approach suffers from high uncertainties with complex input parameters resulting in convergence issues and high computational costs. The emergence of timeseries machine-learning models has shown to leverage trends and pattern recognition to achieve tractable, robust and cost-effective solutions. While previous research primarily focused on predicting existing production trends, this study uses timeseries analysis to forecast oil recovery in a CO2-EOR reservoir, focusing on an inverted five-spot pattern. Dynamic data encompassing pressures, WAG cycles, and injection volumes undergo preprocessing and chronological division to facilitate training and testing of an LSTM model. The analysis of field history calibration through the loss iteration of the training dataset shows promising results reflected in low mean-squared-error of 6.65. Model validation, including the test and blind datasets, demonstrates an approximate R-squared value of 0.88. The forecasted results unveil mounting uncertainty over time, and sensitivity analysis shows the significance of WAG cycle adjustments. This study introduces an economical oil recovery forecasting approach, devoid of complex physical models, offering a versatile framework applicable across disciplines, especially for scenarios influenced by past decisions.