基于DeepSets-TD3的多障碍井约束井眼轨迹智能控制方法

Intelligent Wellbore Trajectory Control Method for Multiple Obstacle-Well Constraints Based on DeepSets-TD3

摘要:

针对密集井网条件下井眼轨迹控制面临的多邻井约束、障碍物输入顺序敏感和连续动作稳定控制问题，本文提出DeepSets-TD3多障碍井约束井眼轨迹智能控制方法。该方法以最小曲率法递推井眼轨迹，将井斜角和方位角增量作为连续控制动作；以靶点相对位移、井斜角及方位角正余弦构成基础状态，以障碍井表面净距及相对方位正余弦构成集合元素，并通过共享编码网络与均值-最大值对称池化形成置换不变表征。在固定四障碍布置测试中，DeepSets-TD3的平均目标误差为7.52 m、物理碰撞率为0。随机四障碍测试中，DeepSets-TD3的平均目标误差为11.2 m，物理碰撞率为1.2%，近失误缓冲区率为1.9%，狗腿严重度合规率为99.6%，且四项指标均优于普通TD3。结果表明，集合表征能够减弱障碍井排列变化引起的策略退化，提高复杂邻井约束下轨迹控制的精度、安全性与泛化能力。

Abstract:

A DeepSets-TD3 method is proposed for intelligent wellbore trajectory control under multiple obstacle-well constraints in dense well-pattern conditions. The wellbore trajectory is recursively propagated using the minimum-curvature method, while inclination and azimuth increments are used as continuous control actions. The base state consists of target-relative displacement and trigonometric encodings of inclination and azimuth, while each obstacle well is represented by surface clearance and relative-bearing features. A shared encoder and symmetric mean-max pooling are then used to obtain a permutation-invariant obstacle representation. In fixed four-obstacle layouts, DeepSets-TD3 achieves a mean target error of 7.52 m, and zero physical collisions. In randomized four-obstacle tests, it achieves a mean target error of 11.2 m, a physical collision rate of 1.2%, a near-miss buffer-zone rate of 1.9%, and a dogleg-severity compliance rate of 99.6%, outperforming plain TD3 on all four metrics. The results indicate that set-based obstacle encoding can alleviate policy degradation caused by obstacle-order variation and improve control accuracy, collision-avoidance reliability, and generalization under complex neighboring-well constraints.