Flowing Bottomhole Pressure During Gas Lift in Unconventional Oil Wells

We present artificial neural network (ANN) models for predicting flowing bottomhole pressure (FBHP) of unconventional oil wells under gas-lift operations. Well parameters, fluid properties, production/injection data, and bottomhole gauge pressures from 16 shale oil wells in Permian Basin, Texas, are analyzed to determine key parameters affecting FBHP during the gas-lift operation. For the reservoir fluid properties, several PVT models, such as Benedict-Webb-Rubin, Lee, Gonzalez, & Eakin, and Standing, among others, are examined against experimentally tuned fluid properties, i.e., viscosity, formation volume factor, and solution gas-oil ratio, to identify representative fluid (PVT) models for oil and gas properties. Pipe flow models, i.e., Hagedorn & Brown, Grey, Begs & Brill, and Petalas & Aziz, are also examined by comparing calculated FBHP against the bottomhole gauge pressures to identify a representative pipe flow model. Training and test data sets are then generated using the representative PVT and pipe flow models to develop a physics-based ANN model. The physics-based ANN model inputs are hydrocarbon fluid properties, liquid flowrate (qL), gas-liquid-ratio (GLR), water- oil-ratio (WOR), well depth (TVD), wellhead pressure (Pwh), wellhead temperature (Twh), and temperature gradient (dT/dh). A data- based ANN model is also developed based on only TVD, P_wh, q_L, GLR, and WOR. Both physics- and data-based ANN models are trained through hyperparameter optimization using genetic algorithm and K-fold validation and then tested against the gauge FBHP. The results reveal that both models perform well with the FBHP prediction from field data with a normalized mean absolute error (NMAE) of around 10%. However, a comparison between results from the physics- and data-based ANN models shows that the physics- based model accuracy is higher at the later phase of the gas-lift operation when the steady state pipe flow is well established. On the other hand, the data-based model performs better for the early phase of gas-lift operation when transient flow behavior is dominant. Developed ANN models and workflows can be applied to optimize gas-lift operations under different fluid and well conditions.