Abstract:

During production of condensate gas, there is a gas-to-liquid phase transition along with temperature and pressure changes in wellbores. This phenomenon results in low accuracy of conventional bottomhole flowing pressure (BHFP) calculation methods such as empirical formula and the modified H-B algorithm. To address this problem, a BHFP prediction model based on machine learning (ML) algorithms including BP neural network and LightGBM was proposed. Using 1,536 measured datasets from 24 wells, followed by normalization, noise removal, and correlation analysis, the modified H-B algorithm, BP neural network, and LightGBM algorithm were employed for modeling. The results show that the histogram-based splitting and feature binding techniques of LightGBM can significantly reduce the computational complexity and improve the fitting accuracy, with the correlation coefficient up to 0.9893, and the mean absolute error (MAE) of 86.5% lower than the modified H-B algorithm. A quantitative analysis was conducted on the main factors controlling BHFP prediction, suggesting gas production as the most prominent factor. Based on the accurate fitting of BHFP, the proposed model was extended to condensate gas wells without pressure-measurement conditions, and it accurately identified two production wells at high-risk pressure points of retrograde condensation. Specifically, for Well A6, the production regime was adjusted by increasing the wellhead tubing pressure by 2.73 MPa, which ultimately led to an increase in the daily gas production by more than 2,000 m3/d, and in the gas productivity index by 31.6%. This performance validated the effectiveness of the proposed model in identifying high-risk wells and guiding optimization of production regime. It is concluded that the ML-based BHFP prediction model can efficiently quantify the multiphase flow behaviors in wellbores, providing reliable technical support for adjusting the production regime of condensate gas reservoirs and controlling retrograde condensation pollution.

Key words: machine learning, condensate gas, bottomhole flowing pressure, BP neural network, LightGBM algorithm, prediction model