Lost Circulation Prediction in Deep Naturally Fractured Reservoirs: Integrating Theory Models with Machine Learning Algorithms

Abstract

Lost circulation is one of the most common incidents during drilling operations in naturally fractured formations. It not only results in significant losses of drilling mud but also increases non-productive time (NPT) and drilling costs. However, predicting lost circulation remains a challenging task, as the lost circulation pressure is governed by a complex array of influencing factors in intricate target formations. Therefore, different lost circulation models should be adapted to match distinct formation characteristics. In this study, we first integrate core data, well logs, and seismic data to map the locations and properties of natural fractures. Subsequently, detailed post-drilling data on lost circulation volume and rate are used to identify three possible types of lost circulation: permeable-induced lost circulation, hydraulic fracturing-induced lost circulation, and natural fracture-induced lost circulation. Based on comprehensive information about natural fractures and actual lost circulation events, appropriate theoretical models are selected for different depth intervals of the target formations, and actual drilling lost circulation data are used to validate these theoretical models. Furthermore, machine learning algorithms are employed to predict lost circulation. To enhance prediction efficiency, instead of training the algorithms using limited data from only a few wells in a specific region, a broad dataset encompassing various formation types is utilized for training. The trained machine learning models for lost circulation prediction are then evaluated by comparing their predictions with actual field data, using metrics such as mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). On the test dataset, the machine learning models achieve the following performance: For formations with permeable-induced lost circulation: MSE = 0.108, R² = 0.997; For formations with hydraulic fracturing-induced lost circulation: MSE = 0.042, R² = 0.986; For formations with natural fracture-induced lost circulation: MSE = 0.069, R² = 0.975. One field cases are also analyzed to demonstrate the practical utility of the machine learning models. The results indicate that the pre-drilling predicted values of formation lost circulation pressure show minimal deviation from actual measurements. The machine learning models exhibit high prediction accuracy, and their predictions are consistent with on-site application outcomes. The application of this machine learning-based lost circulation pressure prediction method can provide technical support for safe and efficient drilling in deep naturally fractured formations. Classify lost circulation into three types by mechanism, match three algorithms to build a system, breaking single-model limits for accurate layered prediction Field validation confirms model accuracy; optimized parameters avoid severe loss, cutting drilling cycle by 21.85 days and boosting efficiency Integrate multi-source data to clarify key indicators of each loss type, providing reliable technical support for deep fractured formation drilling