Machine Learning-Enhanced Forecasting for Efficient Water-Flooded Reservoir Management

Abstract This study focuses on the crucial aspect of inter-well connectivity (IWC) in waterflooding operations, acting as a key indicator of the efficiency of connections between production and injection wells. Accurately evaluating the effectiveness, the flooding process, and predicting the future production rates heavily rely on understanding IWC. While reservoir simulation is comprehensive, it demands extensive input data. Forecasting production rates with only injection and production rates as inputs presents a formidable challenge. To address this, the Capacitance-Resistance Method (CRM) emerges as a prominent reduced-physics model. Despite its utility, basic CRM may produce inaccurate forecasts due to assumptions like a constant Productivity Index and pseudo-steady state flow. In this research, we performed a comparative analysis of various models aimed at predicting liquid production rates solely based on water injection rates. Our investigation involved scrutinizing assumptions and identifying deficiencies in each approach, offering a comprehensive understanding of the strengths and limitations associated with modeling waterflood production scenarios. We utilized pure data-driven methods and modern deep learning time series techniques, specifically Recurrent Neural Networks (RNN). Furthermore, physics-informed data-driven methods, namely Augmented Sparse Identification of Non-Linear Dynamics (SINDy) and Augmented Physics Informed Neural Networks (PINNs) method (which were modified from their actual forms to suit specific problems and niche applications) are also utilized. These modern data-driven regression methods are trained under physics-based constraints to limit the degree of freedom during optimization. Our observations indicated that Augmented PINNs is the best-performing method in terms of accuracy, leading to its selection as our final solution. Our proposed solution demonstrated superior accuracy compared to the established approaches like conventional CRM, offering simplicity, computational efficiency, and scalability for handling large field datasets. Furthermore, we established an algorithm to conduct a health-check for injection and production rates by screening out outliers during training and creating the regression model. This research significantly contributes to enhancing the understanding of IWC and improving the precision of liquid production rate forecasts in waterflooding scenarios.