Optimizing Well Control Strategies with IDLHC-MLR: A Machine Learning Approach to Address Geological Uncertainties and Reduce Simulations

Abstract

This paper presents an advanced version of the previous IDLHC-ML approach, designed to enhance life-cycle well control optimization by reducing simulations. Unlike its predecessor, this updated method, called IDLHC-MLR, uses representative models (RMs) to address the effect of geological uncertainties on production strategies. Despite presenting additional computational challenges, considering uncertainties in determining effective strategies is crucial, making the new IDLHC-MLR approach a valuable solution. The IDLHC-MLR combines the iterative discrete Latin hypercube optimization algorithm (IDLHC) with machine learning (ML) to robustly optimize the well's bottom-hole pressure (BHP) throughout the field management period. The method is applied to the UNISIM-I-M benchmark of Namorado Field, located in the Campos Basin, Brazil. The IDLHC-MLR method trains the initial ML model with well BHP strategies robustly applied to all RMs in the first iteration of IDLHC. In subsequent iterations, the trained ML model is used to predict the expected monetary value of the RMs, and only a subset of new strategies with the highest expected outcome is selected for simulation. In addition, the ML algorithms are retrained with newly generated strategies over the iterations to improve the model's accuracy. The IDLHC-MLR incorporates stacked ensemble learning, which leverages predictions from various base machine learning models to train a secondary algorithm. In this approach, the IDLHC-MLR employs multiple base learners such as Lasso, Gradient Boosting, and Random Forest to make predictions, which are then inputted into a multi-layer perceptron for training purposes. This integration of multiple base models results in a more robust and accurate prediction and provides a 45% reduction in the number of simulations required compared to the traditional IDLHC model while maintaining similar expected monetary value. To conclude, utilizing inexpensive ML models effectively reduces computational time by substituting costly full-physics reservoir simulations. The significant computational time required for full-physics simulations, particularly when considering multiple scenarios to account for uncertainties, can pose a challenge to meeting project deadlines. The IDLHC-MLR methodology, incorporating low-cost ML models, offers a practical solution to reduce computational time, increasing the likelihood of successful project implementation within the given timeline.