Supervised deep learning-based models for HM and forecasting of Water flooding and Enhanced Oil Recovery Scenarios

Advances in new technologies particularly artificial intelligence (AI) and machine learning (ML) have now made it possible to use AI and ML-based approaches to build reservoir analytics models. These new methods have the capabilities to improve speed, efficiency and potential to eventually replace numerical reservoir simulators. The advantages of AI and ML-based simulators is significant time and cost savings.

AI and ML-based deep learning approach involve using subsurface reservoir static, dynamic, and well production data to form a fully independent deep learning-based reservoir simulator with capabilities to perform history matching and forecasting. The objective of this work is to use deep learning algorithms for performing history matching and forecasting on a series of industrial datasets, for waterflood and water alternating gas (WAG) scenarios.

The results show that deep learning based models are very good and give close to 85% accuracy in history matching. Three 3 different datasets are evaluated using deep learning achieving high accuracy. The models are extended to generate long-term forecasts and are optimized through optimization of hyperparameters. The work presented in this paper will demonstrate that AI and ML-based models have the potential to replace the conventional reservoir simulation workflow if an exhaustive data set is available.