Development of Advanced Hybrid Modeling Techniques for Liquid Production Forecasting in Waterflooded Reservoirs Using Historical Injection and Production Data

Summary In this study, we examine interwell connectivity (IWC) in waterflooding operations, a critical factor for assessing the efficiency of production and injection well connections. Understanding IWC is essential for evaluating flooding effectiveness and predicting future production rates. While reservoir simulation provides detailed insights, it requires an extensive input data set. Forecasting future production rates using only historical injection and production data as sole inputs, without incorporating information on reservoir characteristics (such as porosity and permeability), fluid pressure/volume/temperature properties (density, viscosity, formation volume factors, and solution gas/oil ratio), or rock-fluid interaction data [like special core analysis (SCAL) data] is a formidable challenge in reservoir engineering. 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 such as a constant productivity index (PI) and pseudosteady-state flow. We performed a comparative analysis of various models in this study, with the aim of predicting liquid production rates based solely on historical water injection rates data. 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 and implementing various modifications. A variety of pure data-driven methods (models that rely entirely on data without incorporating any physics-based constraints) were utilized, specifically modern deep-learning time-series techniques like recurrent neural networks (RNNs), to predict liquid production rates by training the model to learn the relationship between injectors and producers purely from the historical injection and production data. Furthermore, physics-informed data-driven methods—specifically sparse identification of nonlinear dynamics (SINDy) and physics-informed neural network (PINN) methods (modified from their actual forms to suit specific problems and niche applications)—were trained with physics-based constraints to control the degrees of freedom in the optimization. The observations of this study indicate that the use of augmented PINNs is the best-performing method in terms of accuracy, leading to its selection as our final solution. It demonstrated superior accuracy compared with conventional CRM approaches, offering simplicity, computational efficiency, and scalability for large data sets. In addition, we developed an algorithm to screen outliers in injection and production rates during training, improving model health checks. This research significantly contributes to enhancing the understanding of IWC and improving the precision of liquid production rate forecasts in waterflooding scenarios.