Accurate identification and prediction of lost circulation(LC)are critical for ensuring drilling safety and reducing production costs.Traditional LC detection methods rely heavily on expert knowledge,which inherently ...Accurate identification and prediction of lost circulation(LC)are critical for ensuring drilling safety and reducing production costs.Traditional LC detection methods rely heavily on expert knowledge,which inherently suffers from limitations such as subjective bias,low operational efficiency,and insufficient precision.In this context,data-driven prediction approaches have shown promising potential.This study develops a novel LC risk pre-diction framework that integrates transfer learning with deep learning.By leveraging the mud pit volume from acquired drilling parameters and defining three distinct step intervals,LC risk categories were efficiently labeled.An optimized LSTM-based predictive architecture was constructed using three distribution-alignment transfer learning techniques.Comparative experiments under varying numbers of input features confirmed the effec-tiveness of transfer learning in improving LC prediction performance.To address the class imbalance issue commonly observed in LC risk prediction,a delayed matching verification(DMV)strategy-customized for drilling operations-was introduced.This method mitigates the impact of class imbalance and enhances the evaluation of LC risk recognition capabilities.Experimental results from five test wells demonstrate that the proposed method can effectively label LC risk categories and promptly identify risk types,thereby offering valuable insights to support safe and efficient drilling operations.展开更多
基金supported by the National Natural Science Foundation of China(No.52525402)National Natural Science Foundation of China(No.52274008)National Natural Science Foundation of China Joint Fund Project(No.U23A2026).
文摘Accurate identification and prediction of lost circulation(LC)are critical for ensuring drilling safety and reducing production costs.Traditional LC detection methods rely heavily on expert knowledge,which inherently suffers from limitations such as subjective bias,low operational efficiency,and insufficient precision.In this context,data-driven prediction approaches have shown promising potential.This study develops a novel LC risk pre-diction framework that integrates transfer learning with deep learning.By leveraging the mud pit volume from acquired drilling parameters and defining three distinct step intervals,LC risk categories were efficiently labeled.An optimized LSTM-based predictive architecture was constructed using three distribution-alignment transfer learning techniques.Comparative experiments under varying numbers of input features confirmed the effec-tiveness of transfer learning in improving LC prediction performance.To address the class imbalance issue commonly observed in LC risk prediction,a delayed matching verification(DMV)strategy-customized for drilling operations-was introduced.This method mitigates the impact of class imbalance and enhances the evaluation of LC risk recognition capabilities.Experimental results from five test wells demonstrate that the proposed method can effectively label LC risk categories and promptly identify risk types,thereby offering valuable insights to support safe and efficient drilling operations.
