FEATURE ENGINEERING AND HYBRID MODELING IN MACHINE LEARNING: A UNIFIED APPROACH FOR REDUCING COMPUTATIONAL COMPLEXITY IN FLOW PATTERN IDENTIFICATION AND HYDROCARBON PRODUCTION FORECASTING

Abstract

Engineering applications require predictive models that not only capture complex, nonlinear relationships but also remain physically consistent and generalizable across varying operational conditions. Achieving this balance is challenging, as existing models each have inherent limitations. Many deterministic models, including empirical and mechanistic approaches, assume fixed conditions and struggle to adapt to real-world complexity. Probabilistic models introduce computational challenges and often lack direct ties to physical laws. Machine learning (ML) approaches also face challenges in petroleum engineering due to limited physical interpretability and difficulty generalizing across diverse geological and operational conditions.Empirical models, often based on curve fitting or historical data, are fast but tend to assume homogeneous reservoirs and fixed fluid properties. This makes them unreliable under changing reservoir and operational conditions. Mechanistic models, grounded in first-principles physics, improve interpretability but typically require extensive calibration and high-fidelity input data, making them less practical for real-time applications. Both types of models struggle with complex flow behavior, particularly in unconventional plays. Accurately predicting pressure profiles along wells is essential for production optimization and well integrity analysis. The calculation is predicated on the accurate classification of flow patterns. However, empirical and mechanistic models often fail to capture complex flow patterns, such as slug, churn, and annular flow, that cause significant pressure fluctuations. These flow regimes introduce operational inefficiencies and risks that are difficult to model without sacrificing accuracy or computational feasibility. Probabilistic models, such as Monte Carlo simulations, Bayesian inference, and stochastic modeling, address uncertainty by generating distributions of possible outcomes rather than a single forecast. While useful for risk assessment, they often lack physical constraints, and their accuracy depends on well-defined input distributions and high-quality data. In applications like probabilistic reserve estimation, geological heterogeneity and parameter uncertainty make it difficult to validate input assumptions. As a result, probabilistic models alone are typically insufficient for robust forecasting and are often used in combination with empirical or physics-based methods. To overcome these limitations, this study proposes a unified machine learning framework that integrates feature engineering, dimensional analysis (DA), transfer learning (TL), and decline curve analysis (DCA). This hybrid approach enhances both flow pattern classification and hydrocarbon production ML forecasting while maintaining computational efficiency and physical consistency. Conventional mechanistic models often achieve less than 85% accuracy in predicting pressure profiles across different flow regimes. The study applied DA to derive three dimensionless predictors using the Buckingham Π Theorem. Embedding these predictors into the ML model ensures consistency with underlying physics and improves generalization. Among the evaluated models, the Extreme Gradient Boosted Random Forest (XGB-RF) achieved the highest accuracy of 93% (Kappa = 0.90), outperforming mechanistic models. This highlights the effectiveness of physics-guided feature engineering in refining ML predictions for two-phase flow pattern classification. For production forecasting in multi-fractured horizontal wells (MFHWs), traditional rate-based models often fail due to early-time data sparsity and operational variability. The Machine Learning Assisted – Decline Curve Analysis (MLA–DCA) framework addresses this by integrating DA, TL, and DCA. DA simplifies input complexity while preserving physical meaning. TL enables knowledge transfer from a pre-trained scalar model to a more adaptive vector-based ML model, optimizing both algorithm and hyperparameter selection. DCA complements the ML model for late-time forecasts, where data becomes increasingly sparse. The MLA–DCA framework achieves R² ≥ 0.80 in cumulative production forecasting, demonstrating its effectiveness in balancing accuracy, generalizability, and efficiency. By minimizing feature space, training requirements, and computation time, the framework is scalable for real-world deployment. Feature engineering ensures physical alignment, while hybrid modeling improves performance across diverse flow regimes, reservoir types, and operational conditions. This study presents a physics-informed, data-driven framework that unifies machine learning with engineering fundamentals. The proposed hybrid approach enhances flow pattern identification and production forecasting, advancing the role of ML in petroleum engineering decision-making. By ensuring that models are accurate, interpretable, and efficient, this work provides a robust solution for tackling complex challenges in petroleum engineering applications.