Deep-Learning Technique Optimizes Sequestration, Oil Production in CCUS Projects

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 227168, “Modern Deep-Learning Framework Using Temporal Fusion Transformer To Optimize Sequestration Efficiency and Oil Production in CCUS Projects,” by Ahmed Wagia-Alla, SPE, Mohamed Alghazal, and Turki Alzahrani, SPE, Saudi Aramco. The paper has not been peer-reviewed. _ Optimization of water-alternating-gas (WAG) processes is critical for maximizing both oil recovery and carbon-sequestration efficiency in CO2 enhanced oil recovery (EOR) projects. Conventional optimization using simulation models can be cumbersome because of the vast design space and high uncertainty. In this study, a deep-learning model, the temporal fusion transformer (TFT), is applied for multivariate forecasting of oil production and CO2-sequestration efficiency across a range of WAG scenarios using field data from six legacy CO2 EOR projects. Introduction In this study, a novel, data-driven framework is presented that leverages TFT to forecast 12 months of CO2 and incremental oil production across six legacy CO2 EOR fields in the Permian Basin: East Vacuum (EV), Denver Unit (DU), Wasson San Andres (WSA), Seminole San Andres Unit (SSAU), Scurry Area Canyon Reef Operators (SACROC) Unit, and Rangely Weber Sand (RWS). These fields were selected because of their long production histories, public data availability, and diverse operational characteristics. Production and injection data were digitized from historical publications and technical papers and preprocessed to isolate the incremental oil response to CO2. For each field, the waterflood trend was extrapolated using an exponential decline curve to estimate baseline performance; the difference was attributed to CO2 injection response. To evaluate short-term operational strategies, five different WAG scenarios were created per field by varying the gas/water ratio (GWR). These scenarios were input into the TFT model, which was trained on historical data and used to forecast monthly oil and CO2 production. Performance metrics including root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) were calculated to assess model accuracy. Additionally, CO2 throughput, use, and retention efficiency were calculated using assumed formation volume factors of 0.5 for CO2 and 1.0 for water because these values are common in Permian basin carbon capture, use, and storage (CCUS) fields. This allowed computation of a novel efficiency score to rank WAG scenarios by both economic and environmental merit. The results demonstrated that moderate WAG ratios (approximately 1.5–2.5) offer the best tradeoff between incremental oil recovery and CO2 retention. To the best of the authors’ knowledge, this is the first study to apply a TFT across multiple legacy EOR fields for CCUS optimization. Methodology The workflow consisted of five major stages: data preprocessing and normalization, parsing of CO2 incremental response, model training, WAG scenario creation, and postprocessing for performance and efficiency analysis.