Prediction of Flowing Bottomhole Pressure in Gas-Lifted Wells Using LSTM-ANN Models

We present physics- and data-based long short-term memory neural network (LSTM-ANN) models to predict flowing bottomhole pressure (FBHP) in unconventional wells undergoing gas-lift operations. The models consider reservoir, well, and operational parameters from 16 shale wells under gas-lift operations in Texas, USA. For the physics-based LSTM-ANN model, we initially evaluate several PVT and pipe flow models to identify the representative models for FBHP calculation. The LSTM-ANN model comprises an interconnected LSTM component and an ANN component. We use six days’ injection and production data as inputs for the LSTM component, of which the output is integrated with fluid properties and well parameters as inputs for the ANN component to predict FBHP. For the data-based LSTM-ANN model, well depth, operating valve depth, injection and production data, and measured gauge pressure constitute the dataset, following a similar development procedure. The physical inputs analysis shows that hydrocarbon fluid properties, well depth and operating valve depth, temperature at the wellhead and bottomhole, wellhead pressure, and liquid and gas rates are key factors influencing FBHP during gas-lift operations. After using the 16 wells’ dataset for the training, we compare the hyperparameter-optimized physics- and data-based LSTM-ANN with physics- and data-based ANN and LSTM using four new wells as test sets. The test results from four wells demonstrate that the physics-informed LSTM-ANN model achieved the best performance, with a normalized mean absolute error (NMAE) of 8.9%. Additionally, we evaluate the model using data from a newly developed well. The results reveal that the physics-informed LSTM-ANN model maintain superior accuracy during the final three months of production, achieving an NMAE of 3.9%. Results indicate that the physics-based LSTM-ANN model performs best, particularly in long-term gas-lift operations. The proposed LSTM-ANN models utilize long-term dependent variables with static parameters, which enables them to capture the temporal variation of the FBHP while embedding physical laws. The developed models can be applied to optimize gas-lift operations under different fluid and well conditions.