College of Artificial Intelligence, China University of Petroleum(Beijing)
College of Petroleum Engineering, China University of Petroleum(Beijing)
National Key Laboratory of Petroleum Resources and Engineering, China University of Petroleum(Beijing)
Jianghan Machinery Research Institute Limited Company of CNPC
CNPC Engineering Technology R&D Company Limited
College of Mechaical and Transportation Engineering, China University of Petroleum(Beijing)

钻井管柱力学的准确表征和动态分析是保证安全高效钻井的关键。现有经典的钻井管柱摩阻扭矩软杆/刚杆模型中，钻柱摩阻系数通过经验估算或钻后反演来确定，其准确性和时效性有待提升。本研究基于人工智能技术在复杂非线性映射问题中的有效性，通过预测摩阻系数建立机理−数据融合的钻柱摩阻扭矩预测方法。首先利用已钻录井数据和软杆模型反演摩阻系数，为摩阻系数智能预测提供数据基础，通过对74口井数据处理和特征量化分析，建立考虑数据序列特征的LSTM(Long Short-Term Memory)网络，并通过摩阻扭矩预测和SHAP(SHapley Additive exPlanation)可解释性分析验证模型合理性。结果表明：摩阻系数预测误差为5.89%，摩阻扭矩预测误差降低了4.41%，模型表征的输入特征与摩阻系数的映射关系符合管柱力学机理，具备较强稳定性和可解释性。该方法可为钻井管柱力学的准确表征与动态分析提供理论与技术支撑。

The accurate characterization and dynamic analysis of drilling string mechanics are essential to ensure the safe and efficient drilling. In the classical soft/rigid string model for drag & torque of drilling string, the friction coefficient of the drilling string is determined by empirical estimation or post-drilling inversion, of which the accuracy and timeliness needs to be improved. Based on the effectiveness of artificial intelligence technology applied in complex nonlinear mapping, a drag and torque prediction method of drill string with mechanism-data fusion was proposed by predicting the friction coefficient. Firstly, the friction coefficient was inversed using the drilled and logged well data and the soft-string model, which provides the data basis for intelligent prediction of friction coefficient. As a result of the data processing and quantitative feature analysis of 74 wells, a Long Short-Term Memory (LSTM) network considering data series features was established, and the reasonability of the model was verified through drag & torque prediction and the interpretability analysis by Shapley Additive explanation (SHAP). The results show that the prediction error of the friction coefficient is 5.89%, and the prediction error of drag & torque is reduced by 4.41%. The mapping relationship between the input features of the model characterization and the friction coefficient is consistent with the mechanical mechanism of drilling string, which indicates that the model has strong stability and interpretability. Generally, this method could provide the theoretical and technical support for accurate characterization and dynamic analysis of drilling string mechanics.

国家重点研发计划项目(2019YFA0708300X)；NSFC国家杰出青年科学基金项目(52125401)；中国石油天然气集团有限公司−中国石油大学(北京)战略合作科技专项项目(ZLZX2020-03)；中国石油科技创新基金项目(2022DQ02-0308)

刘慕臣，宋先知，李大钰，等. 钻柱摩阻扭矩智能预测模型与解释[J]. 煤田地质与勘探，2023，51(9)：89−99.

LIU Muchen，SONG Xianzhi，LI Dayu，et al. An intelligent prediction method and interpretability for drag and torque of drill string[J]. Coal Geology & Exploration，2023，51(9)：89−99.