基于井下环空参数的溢流智能预警技术研究

Research on Overflow Intelligent Warning Technology Based on Downhole Annulus Parameters

葛亮 滕怡 肖国清 肖小汀 邓红霞

GELiang TENGYi XIAOGuoqing XIAOXiaoting DENGHongxia

西南石油大学机电工程学院, 四川 成都 610500 西南石油大学人工智能研究院, 四川 成都 610500 西南石油大学化学化工学院, 四川 成都 610500 西南石油大学电气信息学院, 四川 成都 610500

School of Mechatronic and Electrical Engineering, Southwest Petroleum University, Chengdu, Sichuan 610500, China Institute of Artificial Intelligence, Southwest Petroleum University, Chengdu, Sichuan 610500, China College of Chemistry and Chemical Engineering, Southwest Petroleum University, Chengdu, Sichuan 610500, China School of Electrical Engineering and Information Science, Southwest Petroleum University, Chengdu, Sichuan 610500, China

随着油气勘探开发向复杂地层发展，钻井过程中发生井喷的风险增加，而溢流是井喷的前兆，所以早期溢流预警是实现安全井控预防的关键方向。针对传统预测算法在基于地面参数进行溢流预警时未分析溢流严重程度以及预测准确度不高的问题，通过对溢流征兆及溢流发生机理的研究，在利用环空电磁流量系统及其他系统直接测量井下近钻头处的环空流量和其他环空参数的基础上，建立了一种基于人工智能算法——随机森林的溢流智能预警模型来对溢流严重程度进行分类预测。为了验证该预警模型的可行性，通过搭建模拟实验平台进行测试，并与常规的BP神经网络相比较，结果显示该方法正确率高达92.68%，其分类预测的准确性明显高于BP神经网络。研究结果验证了随机森林模型进行井下溢流预警的可靠性，很好地实现了溢流的早期预警，为钻井提供了安全技术保证，具有较好的应用前景。

With the development of oil and gas exploration and development toward complex formations, the risk of blowout during drilling has increased, and overflow is the precursor of blowout, so the early warning of overflow becomes a key direction to well control and safety prevention. Aiming at the problem that the traditional prediction algorithm fails to analyze the severity of overflow and the prediction accuracy is not high when performing overflow warning based on ground parameters, through the study of overflow symptoms and the mechanism of overflow, the annulus electromagnetic flow system and other systems are used to directly measure the underground near the bit, and an overflow intelligent early warning model was established based on artificial intelligence algorithm—Random Forest to classify and predict the severity of overflow. In order to verify the feasibility of the early warning model, a simulation experimental platform was built for testing, and compared with the conventional BP neural network. The results show that the accuracy of this method is as high as 92.68%, and the accuracy of classification prediction is significantly higher than that of the BP neural network. The research results verify the reliability of the random forest model for downhole overflow early warning, which well realizes the early warning of overflow, and provides a safety technical guarantee for drilling, and has good application prospects.

10.11885/j.issn.1674-5086.2021.03.12.03