Predicting production behavior plays an important role in oil and gas production and aids engineers to perform field management. However, this can be challenging in heterogeneous reservoirs using traditional models that require expensive computational time and various types of formation and fluid data. Besides, frequent manual operations are always ignored in the production history because of their cumbersome processing. To overcome this limitation, a supervised deep neural network (DNN) model is established in this paper to forecast hydrocarbon production that considers the nonlinearity as well as the impact of manual operations.
Production history was obtained from 3 wells of the Eagle Ford shale (Frio, La Salle, and Zavala County) where a single well data was split into a 75:25 ratio for training and testing. Deep learning (DL) algorithms, long short-term memory (LSTM) and Bi-LSTM models were built, and an optimization algorithm was used to determine the hyperparameters of the optimal model in production prediction. Finally, machine learning (ML) classifier, Random Forest (RF) were used to investigate the efficacy of forecasting.
The results reflect promising predictions using DL models. Among three wells, LSTM and Bi-LSTM perform relatively closer in three counties with non-linear production history and reported as Frio with (R2 ± 84%, Error ± 9), La Salle ( R2 ± 97%, Error ± 4) and Zavala (R2 ± 91%, Error ± 5). Among all tested models, Bi-LSTM showed superior prediction performance for La Salle county’s well with ± 98% R2 and less than ± 5 of RMSE. While datasets have missing values, these results demonstrate an excellent ability to generate a realistic forecast.
DL models will likely be able to fill the information gap by making an accurate prediction. The adoption of DL models for predicting well production could have extensive impacts on the role of production engineers where well production history has significant non-linearity due to manual recording. Operators could leverage this promising DL model by utilizing it as an automated handy tool to audit the production prediction quickly where datasets have missing values.
