Research on Prediction Methods of Mud Loss Risk Based on Different Seismic Attributes

ABSTRACT: Seismic data is often used to predict the risk of leakage. Seismic attributes are data derived from original seismic data. Different seismic attributes have different responses to the risk of leakage. In this paper, a prediction model of mud loss probability based on different kinds of seismic attributes is established. Firstly, 25 kinds of seismic attributes are calculated, the correlation degree of 25 kinds of different seismic attributes to the leakage point is analyzed, and three kinds of seismic attributes with correlation degree [0,1), [0.2, 1), [0.4, 1) are used as training data sets to establish a deep neural network model. Finally, the obtained mud leakage probability model is applied in BZ block. The research shows that the accuracy of the model trained by training data with correlation degree [0.2, 1) is 78.2%, and the prediction accuracy of this model is the highest. Selecting the appropriate seismic attributes can improve the prediction accuracy of the model. This paper provides some guidance on how to select the seismic attributes required by machine learning, and provides scientific suggestions for the prevention of mud loss. 1. INTRODUCTION In the context of the high cost of offshore oil drilling, loss accidents often occur, which will not only prolong the drilling cycle, but also increase the probability of blowout, kick, wellbore instability and borehole abandonment (Zhang et al.,2021). In particular, in complex formations such as induced fracture development, leak-collapse, leak-flow coexistence, large fractures and caverns, and faults, malignant leakage occurs frequently, which seriously restricts the process of exploration and development (Zhang Qian et al.,2021). One of the keys to prevent and stop leakage is pre-drilling prediction and rapid determination of leakage position during drilling (Yang et al.,2021). At present, the commonly used drilling accident identification methods include manual judgment, expert system, etc., but these methods have the problem of inflexible knowledge acquisition in drilling accident identification (Sun et al.,2022). With the development of comprehensive logging technology, computer information technology and other intelligent technologies (Liu et al.,2019), informatization and intelligence have become the new construction direction of oil field enterprises (Li et al.,2023), and the intelligent early warning system applied in oil drilling engineering has gradually developed and applied (Li Qi et al.,2008). Traditional methods have low accuracy and poor timeliness in identifying well losses, while artificial intelligence technology can well solve multi-parameter, non-linear complex problems (Chen et al.,2022). Neural network is one of the core models in the field of artificial intelligence (Fan et al.,2022), and most scholars use neural network model (Li et al.,2013), because neural network is suitable for complex nonlinear data modeling, and big data technology is a method for data analysis and prediction based on non-mechanism models. Therefore, it is of great significance to carry out research on the application of such technologies in the drilling engineering field, which is especially complex and accident-prone in the petroleum industry (He et al.,2019). There are many reasons leading to well loss. This paper mainly gives early warning of well loss from complex formations such as faults, fractures and geological fracture zones. In order to judge complex structures in these unknown formations before drilling, seismic data can be used to effectively interpret the complexity of the strata (Lu et al.,2019). Many scholars have used machine learning to build a well loss prediction model (Meng et al.,2022), but the selection of data set required by artificial intelligence learning and how to make labels for the probability of loss are fuzzy, thus reducing the accuracy of the prediction results. This paper gives some guidance on how to select the data set required for machine learning and how to make labels suitable for prediction of leakage risk.