|
|
| |
From Oilfield Byproduct to Energy Resource: Decline Curve and LSTM Forecasting of Produced Hot Water in the Bakken Formation (Alger Field)
Emmanuel AGYEI, Nathaniel Nimo YEBOAH, Emmanuel GYIMAH, Hamid RAHNEMA, William AMPOMAH, Benedicta VIDZRO, Kojo Acheampong BOATENG, Audrey AYENSIGNA, Godsway AKPABLI
[New Mexico Institute of Mining and Technology, USA]
Produced water, often co-produced with oil and gas, is typically hot and can serve as a valuable geothermal resource for generating energy to support field operations or supply nearby communities and industries. Like hydrocarbon production, the volumetric rate of produced hot water declines with time, making it essential to understand its decline behavior if it is to be harnessed as a sustainable alternative energy source. In this study, production data from three wells, Anderson 28-33 1-H, Charlie Sorenson 17-8 3TFH, and Ross 7-17H, located in the Mississippian/Devonian Bakken formation (Alger field), spanning August 2009 to August 2025, were analyzed to characterize water-production decline trends. The classical Arps decline models (exponential, hyperbolic, and harmonic) and post-Arps models (LGM, SEPD, PLE, Wang model, Duong model, and VDMA) were first employed to evaluate the decline performance of the produced water. The best-performing Arps and post-Arps models for each well were selected and further compared against an LSTM architecture trained on the same historical dataset. The study results showed that, for historical water production data with very little to no fluctuations, empirical DCA models performed equally well as the data-driven LSTM model. However, in instances where production data is severely noisy due to field operations, the data-driven LSTM outperforms the empirical DCA models, both qualitatively and quantitatively. Despite the improved predictive ability of the LSTM model in capturing non-linear dependencies inherent in field data, compared to empirical decline curve analysis models, the mean predictive errors were within 10 - 15%. This error could be further reduced by integrating variables such as well shut-ins, operational interventions, changes in choke settings, artificial lift adjustments, workover activities, water breakthrough events, reservoir pressure variations, and production constraints into the dataset used for training the LSTM model. Including these factors would enable the LSTM model to better distinguish between transient operational effects and underlying reservoir-driven decline behavior.
Topic: Production Engineering