Concatenating data-driven and reduced-physics models for smart production forecasting

Production forecasting is vital for petroleum reservoir management but remains challenging. This study combines the Capacitance-Resistance Model (CRM) (a reduced physics model) with machine learning (ML) (or data-driven) approaches - dubbed CRM-ML hybrids - to enhance production forecast accuracy in petroleum reservoirs. Using both synthetic field (synfield) and real field data, four ML approaches (Nu-Support Vector Machine, NuSVM, Extreme Gradient Boost, XGB, Extreme Learning Machine, ELM, and Multilayer Perceptron, MLP) were tested. Considering all 560 evaluations, the CRM-ML hybrids generally outperformed standalone ML approaches, with the CRM-XGB hybrid achieving the lowest mean absolute error of 7.2 barrels per day. The findings reveal that hybrid models improve production forecasts, with performance influenced by well-specific operational and reservoir factors. Despite possible challenges with interpretability and computational costs, this integration demonstrates the potential for leveraging reduced-physics models and ML for better reservoir predictions.