Abstract
The use of machine learning (ML) techniques for forecasting in the oil and gas sector is growing rapidly, driven by advancements in data processing and computational power. Traditional methods like decline curve analysis and numerical simulations have long been the preferred approach for single-phase fluids or when large amounts of high-quality data were available. However, these methods often fall short when faced with changing operating conditions, limited data, and uncertainties in surface and subsurface parameters. This paper introduces an integrated machine learning framework, based on established neural network models, proposing a method to combine engineering model results with data-driven, self-learning tools to predict future production. This approach proves particularly valuable when a single forecasting technique is insufficient for complex fields with fluctuating surface conditions and dominant subsurface uncertainties.
The study details the creation of this machine learning framework, evaluating traditional forecasting tools such as decline curve analysis, rate transient analysis, and numerical simulations, alongside training and testing a machine learning model using real production data from a set of wells in water and gas injection reservoirs. A subset of historical data was collected and validated for testing the model, integrating relationships between various parameters to assess the model's ability to learn patterns from time series data and model non-linear relationships for future predictions. Actual field data from selected secondary drive reservoirs were used to predict well production, while reservoir events like water breakthrough were inferred using existing models.
This case study demonstrates the superior performance of the machine learning model over conventional methods like decline curves and numerical simulations, particularly in scenarios with fluctuating operational conditions. The proposed ML technique provides better predictions in cases of water breakthrough with robust learning capabilities when additional data from field studies were included. The integration of numerical simulation data with historical data allowed the model to uncover relationships between key parameters, enhancing its predictive accuracy.
The study provides evidence that, when used appropriately, ML models offer a self-learning and dependable method for production forecasting, addressing complexities that traditional statistical methods often fail to capture. Numerical simulations alone are insufficient in secondary drive reservoirs with fluctuating surface conditions. Similarly, data driven models also lack in predicting an event which has not occurred in the past. However, combining key results from different modeling approaches—empirical, rate transient, material balance, finite element, and data-driven models—can help manage uncertainty in production forecasting. The workflow and procedures developed in this study can be adapted to other reservoir types and machine learning models, with adjustments to neural network structures and the establishment of functional relationships among well parameters. The paper highlights specific cases where ML models can be tuned with other engineering analyses to reduce prediction errors, providing a guide for petroleum engineers to identify critical hyperparameters that improve the accuracy of predictions. Ultimately, achieving forecasting precision through artificial intelligence depends on integrating domain knowledge, incorporating existing study outcomes, and leveraging advanced computational power with functional algorithms.
