| 引用本文: | 王钰豪,郝家胜,张帆,魏强,彭知南,段慕白.钻井溢流风险的自适应LSTM预警方法[J].控制理论与应用,2022,39(3):441~448.[点击复制] |
| WANG Yu-hao,HAO Jia-sheng,ZHANG Fan,WEI Qiang,PENG Zhi-nan,DUAN Mu-bai.Adaptive LSTM early warning method for kick detection in drilling[J].Control Theory and Technology,2022,39(3):441~448.[点击复制] |
|
| 钻井溢流风险的自适应LSTM预警方法 |
| Adaptive LSTM early warning method for kick detection in drilling |
| 摘要点击 1601 全文点击 464 投稿时间：2021-07-24 修订日期：2022-01-04 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2022.10662 |
| 2022,39(3):441-448 |
| 中文关键词 溢流预警 数据集扩充 自适应 滑动窗口 长短期记忆网络 |
| 英文关键词 kick detection data set expansion adaptive slide window long short-term memory |
| 基金项目 川庆钻探工程有限公司钻采工程技术研究院项目(2021ZZ03), 中国博士后科学基金面上项目(2021M700695)资助. |
|
| 中文摘要 |
| 溢流会对油田的勘探过程造成严重的影响, 众多钻井平台采取人工坐岗预警, 十分依赖坐岗人员的积累经
验, 导致误报率和成本较高, 因此迫切需要提高溢流预警的效率和质量. 研究表明, 将深度学习方法引入钻井能够有
效提高溢流预警的准确性. 然而, 溢流发生的频率较低, 不同井之间具有差异性, 可获取的训练样本有限, 这些因素
都限制了当前预警算法的应用. 针对上述问题, 本文提出了一种自适应的长短期记忆网络(LSTM)溢流预警算法, 该
算法利用滑动窗口法扩充数据集, 计算平均值增量实现不同井数据的自适应特征提取, 进而屏蔽不同井的差异性,
提高了算法的通用性. 模型在离线验证以及现场溢流的早期预警上表现优良, 与专家的经验标注保持了较高的一致
性, 均提前于操作人员的记录日志. 模型的应用减轻了坐岗人员的负担, 确保了溢流预警的效率与质量, 对提高钻井
工程溢流预警水平具有积极意义. |
| 英文摘要 |
| Kick has a significant influence on the oilfield exploration process, many drilling platforms adopt manual onthe-
job detection, which depends heavily on the accumulated experience of on-the-job personnel, resulting in false alarm
rate and high cost. Therefore, it is urgent to improve the efficiency and quality of kick detection. According to the findings,
including the deep learning technology into the drilling process can significantly increase the accuracy of kick detection.
However, because the frequency of kick is low, there are variances between wells and the available training samples are
restricted, the present kick detection algorithm’s applicability is limited. To address the aforementioned issues, this paper
proposes an adaptive long short-term memory (LSTM) kick detection algorithm that uses the sliding window method to
expand the data set, calculates the average increment, and realizes adaptive feature extraction of different well data, thereby
shielding the differences of different wells and improving the algorithm’s universality. The model performs well in off-line
verification and early detection of on-site kick, and it maintains high consistency with the expert’s experience annotation,
which is earlier than the operator’s log. The proposed method reduces the burden of on-the-job people with better efficiency
and quality, which is beneficial to improve the level of kick detection in drilling engineering |
|
|
|
|
|