融合LightGBM和SHAP的井漏类型判断及主控因素分析.

In the Kuche piedmont structure in the Tarim Basin where complex geological conditions prevail, frequent mud losses into the salt/gypsum formations and the target zones cause huge economic losses. To identity the types of the mud losses, a judgement model is established using the LightGBM algorithm. The LightGBM model, with good discriminative performance, has average recall rate of 85%, precision of 91% and F1-Score of 86.7%. In analyzing the types of mud losses, the interpretable machine learning techniques based on SHAP values are adopted to analyze a single mud loss event and all mud loss events as a whole. The SHAP value method, which is based on Cooperative Game Theory, breaks down the occurrence of mud loss events into contribution values of different features, and explains the effects of each feature on the mud loss event. Studies show that the main factors affecting mud losses include the difference between the mud density and the equivalent density calculated from the fracture pressure of the formation, the flow rate of mud, the well depth and the formation drilled. For the geology of the salt/gypsum formations and the target zones in the Kuche piedmont structure, the effects of the formation geology and the vertical distribution of the interlayer are in depth analyzed. This study enables the field engineers to fast and accurately determine the types of mud losses, and provides a strong support to the design of measures for preventing and controlling mud losses. [ABSTRACT FROM AUTHOR]