Seismic Data and Prediction of Mud Loss Risk Based on Physical Information Constraints

ABSTRACT: The application of machine learning in the oil field has become a major trend, in order to study how to make the machine learning model of drilling site mud loss prediction more consistent with the physical mechanism. In this paper, a prediction method combining machine learning and leakage pressure calculation model is proposed. The calculation result of leakage pressure is taken as the physical constraint condition, and the deep neural network model based on machine learning is combined to predict the probability of downhole mud leakage risk. By comparing the relationship between the equivalent density of mud loss pressure and the mud density, combined with the actual mud loss point, the mud loss probability was calibrated with different sizes, and the relationship between mud loss probability and seismic properties was obtained by machine learning. A prediction model of mud loss probability considering physical constraints was established. The results show that compared with the traditional neural network prediction model which only considers data, the prediction accuracy of this model for mud loss at different depths can reach more than 80%. The advantages of this model are that it improves the reliability of the model prediction results, and has faster convergence speed and lower demand for data. 1. INTRODUCTION Loss is a common problem in the drilling process at home and abroad (Liang Wang et al.,2023). It will prolong the drilling cycle, consume a large amount of drilling fluid and leak plugging and prevention materials, and may even cause serious downhole accidents such as stuck drilling, overflow and even blowout, which may lead to hole abandonment in serious cases (Jinsheng Sun et al.,2021). With the increasingly close combination of oil and gas exploration and development with artificial intelligence technology (Lichun Kuang et al.,2021), artificial intelligence has become a development trend in drilling engineering to predict complex accidents such as leakage (Xin Zhang et al.,2022). As the most important branch of machine learning, deep learning has unique advantages in data processing, and practitioners have carried out research in various fields of oil and gas (Chuanshu Yang et al.,2021;Chao Min et al.,2020). Xiyu Tu et al. (Xiyu Tu et al.,2018) established a well loss accident warning model based on Xgbost algorithm with the support of logging observation logs and natural language processing. Leiwen Wang (Leiwen Wang et al.,2019) combined the biogeographic algorithm with BP neural network to establish a prediction model of the lost layer. The initial weight and threshold of the network were obtained by the algorithm to train the neural network, and relatively accurate prediction results were obtained. At present, the prediction of well loss risk probability is mainly based on seismic data (Zhiyuan Lu et al.,2019), but most of them are based on pure data (Pengfei He et al.,2019), which lacks credibility. Leakage pressure is an important parameter for formation leakage phenomenon and analysis of leakage accidents (Lin Shi et al.,2010). In this paper, a prediction method combining machine learning model and leakage pressure calculation model is proposed, which takes the leakage pressure calculation model as the physical constraint condition (Xiaopeng Zhai et al.,2013;Yongjie Xu et al.,2006;Yan Jin et al.,2007;Liang Zhu et al.,2008), and combines machine learning deep neural network model. Predict the probability of downhole loss risk.