Abstract:

Accurate prediction of collapse and fracture pressures is crucial for well trajectory design, wellbore stability control, and efficient drilling operations. Traditional numerical and analytical methods are often computationally complex and inefficient, while purely data-driven models, although faster, suffer from pronounced black-box characteristics and lack interpretability, which limits their engineering applicability. To overcome these challenges, this study proposes a hybrid-driven prediction method that integrates wellbore stability mechanistic models with a multi-task learning framework (MW-MMoE). In this approach, stress coordinate transformation is embedded as physical prior knowledge at the input stage, while the output targets are reconstructed by first predicting key stress components and then converting them into equivalent densities of collapse and fracture pressures through physical formulations. The Mohr-Coulomb criterion is further incorporated into the loss function as a physical constraint. The model architecture leverages a multi-gated mixture-of-experts network combined with the GradNorm algorithm to dynamically adjust task weights and balance gradients during training. Ablation experiments demonstrate that the proposed MW-MMoE achieves mean absolute errors as low as 0.0019 g/cm³ and 0.0033 g/cm³ for collapse and fracture pressure equivalent densities, respectively, significantly outperforming both single-task and conventional multi-task models, while achieving over a hundredfold improvement in computational efficiency compared with analytical methods. Case studies further validate its engineering applicability: the model can rapidly generate collapse and fracture pressure equivalent density curves for individual wells, produce high-resolution contour maps under arbitrary well inclinations, azimuths, and stress conditions, and perform large-scale three-dimensional predictions across the entire study area. These results highlight that the MW-MMoE model combines high accuracy, efficiency, and interpretability, providing a novel and practical solution for intelligent wellbore stability prediction with broad application prospects.

Key words: multi-task learning, hybrid-driven model, wellbore stability, collapse pressure, fracture pressure, stress coordinate transformation

摘要：

准确预测坍塌压力与破裂压力是井眼轨迹设计、井壁稳定控制及高效钻井作业的关键。传统数值与解析方法计算复杂、效率低，纯数据驱动模型虽然高效，但因“黑箱”特性显著、可解释性不足而限制了工程应用。针对上述问题，本文提出一种融合井壁稳定机理模型与多任务学习的混合驱动预测方法。该方法在输入端将应力坐标转换过程作为物理先验知识嵌入到数据驱动模型中；在输出端重构预测目标，先预测关键应力分量，再通过物理公式转换为坍塌压力与破裂压力当量密度，并在损失函数中引入Mohr-Coulomb准则作为物理约束。模型架构使用多门控混合专家网络与梯度归一化算法，实现多任务间的动态权重调节与梯度平衡。消融实验表明，所提多任务混合驱动模型MW-MMoE在坍塌压力与破裂压力当量密度预测的平均绝对误差分别低至0.0019 g/cm³和0.0033 g/cm³，预测精度显著优于单任务和传统多任务模型，预测效率较解析方法提高逾百倍。实例应用进一步验证了该方法的工程适用性，其不仅能够快速生成单井的坍塌与破裂压力当量密度曲线，还能在任意井斜、井筒方位及应力条件下生成高分辨率当量密度云图，快速预测得到三维区块的钻井品质体。研究结果表明，本文提出的MW-MMoE模型兼具高精度、高效率与强可解释性，为井壁稳定智能预测提供了一条新思路，具备良好的工程应用前景。

关键词: 多任务学习, 混合驱动模型, 井壁稳定, 坍塌压力, 破裂压力, 应力坐标转换

CLC Number: 

• TE21

• TP181