Exploring an Alternative Approach for Predicting Relative Permeability Curves from Production Data: A Comparative Analysis Employing Machine and Deep Learning Techniques

Abstract

Abstract This study aims to present a novel approach for estimating relative permeability curves using Machine Learning (ML) and Deep Learning (DL) techniques based on production data. This method aims to overcome the shortage in core data availability, which is much needed for reservoir simulation. The adopted approach involves devising an algorithm utilizing different methodologies employing synthetic simulation data. The water-oil relative permeability curves correlated with various parameters such as production data and reservoir parameters. Subsequently, the model was applied to real field data to predict relative permeability curves. The procedural framework involves data compilation, model training, application to real field data, and integration with reservoir simulation. In data compilation, synthetic simulation outcomes, including water-oil relative permeability curves, oil and water rates, flowing pressures, and diverse reservoir characteristics are gathered. Model training focuses on developing an ML model proficient in discerning complex relationships between production parameters and relative permeability curves. Following this, the model’s accuracy is rigorously assessed through validation, involving data not used during the training or the validation process. In the application to field data, the trained model is employed to estimate relative permeability curves using authentic production data. Finally, integration with reservoir simulation entails assimilating the estimated curves into simulations to enhance the process of initial history matching. The methodology encompasses the generation of synthetic data to train the Machine Learning model, validation of its performance, application to real field data, and the utilization of the estimated relative permeability curves in reservoir simulations. This process is particularly crucial in situations where core data is non-existent or only limited data points are available. The use of such data contributes significantly to the improvement of history-matching accuracy. The outcomes demonstrate the effectiveness of the employed methodology, highlighting the successful utilization of ML and DL to precisely predict relative permeability curves from production data. The integration of these curves into reservoir simulations yields a notable improvement in the accuracy of the history-matching process in a short time compared with the traditional approaches. The observations elucidate the adeptness of ML or DL in capturing intricate relationships between production parameters and relative permeability curves. The estimated curves serve as a pivotal initial step for history matching, playing a crucial role in substantially mitigating uncertainty within reservoir simulations. In conclusion, this study introduces a new approach for predicting relative permeability curves by leveraging ML and DL integrated with production data. This method contributes to addressing the uncertainties linked to traditional core-based measurements by furnishing a more accurate initial prediction for relative permeability curves. The successful integration of this approach into reservoir simulations holds the promise of streamlining and enhancing the accuracy of reservoir management practices. The emphasis on utilizing the available production data enhances our capability to mitigate the scarcity of Special Core Analysis (SCAL) data. Consequently, this methodology contributes to refined reservoir simulation outcomes and more efficient history-matching processes.