Two-stage sampling scheme for a CNN-based well placement optimization of 3D benchmark reservoirs

Abstract

Well placement is crucial for optimal oilfield development. Since global search algorithms require lots of computational costs, proxy models have been developed to replace full reservoir simulation. However, many proxy models have difficulties for predicting dynamic behaviors of complex reservoirs. Some may utilize extensive amount of training data, apply different machine learning techniques, incorporate parameters for dynamic behaviors, or keep retraining proxy models as needed. For the retraining of a proxy model, its time and frequency are arbitrary.In this study, we propose an improved sampling scheme for initial training data to achieve good performances of the proxy model without the retraining process. The proposed method performs sampling over two stages. Quality information on the well placement scenarios, which indicates the potential productivity of each option, is evaluated from the samples in the first stage. After that, the second samples are obtained by reflecting this quality information as probability distribution. To validate the effectiveness of the proposed method, we perform well placement optimization for 3D benchmark reservoirs of PUNQ-S3 model and Egg model. The proposed proxy model yields optimization results with good predictive errors less than 2.1% without the retraining.