| description abstract | Abstract. Accurate prediction of flowing bottomhole pressure (FBHP) is essential for the effective design and optimization of gas-lift systems in unconventional wells. Conventional methods for FBHP estimation often prove inadequate for unconventional wells, either being overly simplistic or computationally expensive. Accordingly, we develop and evaluate several deep learning architectures for predicting FBHP in unconventional shale wells under gas-lift operations. A comprehensive dataset is compiled from 21 oil wells in the Texas Permian Basin Shale, incorporating readily available parameters such as well depth, operating valve depth, and production/injection data. The predictive capabilities of an artificial neural network (ANN), a long short-term memory (LSTM) network, a hybrid LSTM–ANN model, and a transformer model are evaluated. Five of the 21 wells are used to assess the deep learning model performance. The analysis of the results revealed factors that influence the deep learning model's performance. The results show that the transformer model's predictive performance is most reliable across different scenarios and among the four deep learning models, with an error of around 10%. Additionally, selection bias in the training set can significantly affect the model's predictive performance. Furthermore, our analysis reveals that hyperparameter tuning reduces residual errors by refining model parameters, leading to architectures that better capture the patterns in the training data and deliver enhanced predictive accuracy. This work highlights the significant potential of advanced deep learning models as practical tools for optimizing gas-lift operations across a wide range of fluid and reservoir conditions. | |
