.png){kind=icon}
.png){kind=icon}
Return
Early kick detection based on multi-scale temporal modelling and interpretability framework
.png){kind=icon}
PRE
AI
DOI：10.1016/j.psep.2025.108177.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The early characteristics of kick are hidden. Historically, data-driven methods exhibited constraints in jointly modeling various feature levels and time granularities, primarily due to their reliance on a singular pattern. Moreover, the model's black box nature resulted in a lack of transparency in its decision-making process, raising concerns about its credibility. This study presents a model (TCN-BiGRU-AT) designed for the monitoring of well kick events. The model effectively captures local fluctuations in parameter characteristics and models long-term state evolution, while integrating attention mechanisms to dynamically focus on key abnormal signals. The SHAP framework is introduced to elucidate the influence mechanism of drilling parameters on model decision-making, focusing on global feature importance and local feature interaction. The findings indicate that this method effectively minimizes the risk of missed detections while achieving high accuracy, with a recall of 98.7 % and precision of 97.6 %. The system demonstrates elevated recognition rates for kick testing across various reservoir conditions, offering kick detection 20 and 14 min in advance for two blind wells, respectively. The SHAP global explanation identified common risk factors associated with kicks, such as Hook Load, Standpipe Pressure, and Outlet Flow Rate, and elucidated the dynamic changes in characteristics during kick events through localized case studies. This method effectively captures complex kick risks and elucidates the contribution and interaction effects of various characteristics, thereby serving as a guide for kick emergency decision-making and engineering optimization.
Journal
IF:
7.8
Papers:
9.5K
Citations:
3.8W