Data-Driven Intelligent Prediction of In-Situ Stress in Horizontal Wells and Its Application in Hydraulic Fracturing Design

ABSTRACT: In-situ stress prediction and distribution are key factors influencing hydraulic fracturing design to ensure balanced fracture initiation and propagation. Traditional in-situ stress prediction relies on logging data interpretation and rock mechanics experiments while drilling data help characterise the reservoir's geomechanical properties. This study proposes a new method for intelligent prediction of in-situ stress in horizontal wells using drilling data from horizontal wells, logging data, and rock mechanics data from neighboring evaluation wells (vertical wells). The process started with drilling, logging, and rock mechanics parameter integration, followed by correlation analysis to construct the input dataset. Next, machine learning technologies were used to predict the minimum horizontal in-situ stress, such as the LightGBM+XGBoost model and the Self-Attention+LSTM model. Finally, the model was validated through accurate predictions of the minimum horizontal in-situ stress in a test well from the study area, which achieved an RMSE of 0.073 MPa, demonstrating high predictive accuracy. Additionally, an unsupervised clustering model of in-situ stress is established to optimize fracturing stages and clusters placement to ensure balanced fracturing. This study introduces a new method for in-situ stress prediction and optimization of fracturing stages/clusters in unconventional reservoir horizontal wells. 1. INTRODUCTION Hydraulic fractures generally propagate in the direction of minimum in-situ stress, with the magnitude of this stress being closely related to fracture initiation pressure (Sone and Zoback 2014, Sheng et al. 2024). In most unconventional oil and gas reservoirs, the minimum in-situ stress is usually the minimum horizontal in-situ stress Shmin (Snee and Zoback 2022). Accurate prediction of Shmin distribution is essential for optimizing hydraulic fracturing designs in multistage fractured horizontal wells. A well-designed perforation cluster placement, such as selecting intervals with similar Shmin, can ensure balanced fracture initiation and propagation, ultimately enhancing well productivity and recovery efficiency. At present, the methods for determining the Shmin include direct measurement methods, indirect inference methods, numerical simulation methods, and seismic prediction methods. Direct measurement methods, such as field-scale hydraulic fracturing, small-scale fracturing tests, laboratory-scale wave velocity anisotropy tests, and Kaeser acoustic emission tests, offer high accuracy in measuring in-situ stress. However, due to the high cost of experiments, operational complexity, and the limitation of measuring stress at only a few points, it is difficult to evaluate stress along the entire wellbore. As a result, these methods are typically used as calibration values in stress modeling (Meng et al. 2022, Xu et al. 2022, Yang et al. 2022, Wang et al. 2011). Indirect inference methods use logging data and rock mechanics experiments data, combined with elastic mechanics theory, to calculate in-situ stress. This method is less accurate than direct measurement methods but provides continuous vertical stress distribution and is widely applied in stress analysis due to its lower cost (Ma and Xiang 2021, Wang et al. 2022). However, indirect inference methods heavily depend on array sonic logging data, and the calculation process is complex. Additionally, high logging costs often result in most horizontal wells not being logged, leading to a lack of essential data for interpreting stress distribution in horizontal sections. To address these challenges, this study introduces a cost-effective, data-driven approach to predict in-situ stress using machine learning algorithms.