Prediction of Plunger Lift Dynamics Using a Bidirectional Long Short-Term Memory Neural Network with an Innovative Forecasting Strategy

Plunger lift is commonly applied to both conventional and unconventional gas wells to address issues such as liquid loading, which can result in liquid accumulation and even cease production. By leveraging machine learning (ML) algorithms, well surveillance data can be utilized to predict the condition of gas wells utilizing plunger lifts, facilitating decisionmaking in gas well deliquification. In this paper, a recurrent neural network (RNN) based on long short-term memory (LSTM) for production process data analytics was employed to capture the temporal dynamic behaviors of time-series data. Our findings indicate that the bidirectional LSTM combined with the encoder-decoder architecture (BILSTM-AE), multiforecast average bidirectional LSTM (ABILSTM), and noisesmoothed bidirectional LSTM (SBILSTM) achieved root mean square error (RMSE) values of 1.20%, 1.09%, and 0.95%, along with R-squared (R
²
) scores of 95.49%, 96.97%, and 96.38%, respectively. A further statistical comparison revealed that the ABILSTM model demonstrated superior performance compared to the other models. With high-precision minute-level forecasting, this study has the potential to significantly improve production forecasting, operational adjustments, and hazard prevention in practical processes.