Research on deep reservoir anti-condensate early warning method based on POA-LSTM-TCN combined model for gas reservoirs -a case study on Bozi gas reservoir

Read full text Summarise with VIRAT
Authors: ZHU SUYANG, LI YING, Peng Xiaolong, LIU WEI, Guan Wenjie
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Ultra-deep pore-fracture-fault complex condensate gas reservoirs exhibit highly heterogeneous fluid flow behaviors. During production, the fracture system often experiences locally reduced pressure, leading to retrograde condensation, while the matrix pressure and overall reservoir pressure remain above the dew-point pressure. This discrepancy makes it difficult for traditional gas reservoir engineering methods—typically based on average reservoir pressure—to accurately identify the onset and extent of local retrograde condensation. To address this issue, this study investigates the Bozi condensate gas reservoir located in the Kuqa Depression of the northern Tarim Basin. The flow mechanism and pressure response characteristics of the pore-fracture-fault triple-medium system are systematically analyzed. Based on the variation patterns of wellhead oil pressure, the production process is divided into three distinct stages: a steady-decline period, an unstable-fluctuation period, and an accelerated-decline period. Abnormal fluctuations in the gas-oil ratio (GOR) are interpreted as early indicators of phase change. By examining GOR variations across different pressure intervals, this work characterizes the dynamic evolution of complex medium gas reservoirs at various production stages. A hybrid predictive framework is proposed that integrates the Long Short-Term Memory (LSTM) network and the Temporal Convolutional Network (TCN), whose hyperparameters are globally optimized using the Pelican Optimization Algorithm (POA). A weighted fusion strategy is employed to construct the POA-LSTM-TCN combined model, enabling stage-wise fitting and prediction of GOR. The results demonstrate that the optimized POA-LSTM and POA-TCN models achieve mean absolute percentage errors (MAPE) of 3.71% and 7.73%, respectively, whereas the POA-LSTM-TCN hybrid model achieves a significantly lower MAPE of 2.40%, outperforming the single models by 1.31% and 5.33%. Numerical simulation further verifies that the traditional gas reservoir engineering approach based on average pressure fails to effectively capture retrograde condensation occurring within fractures. In contrast, the POA-LSTM-TCN model not only provides high-accuracy and efficient GOR prediction but also identifies retrograde condensation when deviations exceed the prdefined threshold. Therefore, this study overcomes the limitations of conventional engineering methods in detecting local retrograde condensation and establishes an early-warning approach based on anomaly recognition. The findings hold substantial theoretical and practical significance for production dynamics analysis, retrograde condensation mechanism identification, and development optimization of complex condensate gas reservoirs.

Fields of Study & Subjects

Reservoir Engineering and Simulation Methods Hydrocarbon exploration and reservoir analysis Hydraulic Fracturing and Reservoir Analysis

Publication details

Journal / source
DOAJ (DOAJ: Directory of Open Access Journals)
Publication date
1 January 2026
Language
ZH