A Hybrid Machine Learning Approach for Accurate Horizontal in-situ Stress Prediction in Complex Hydrocarbon Reservoirs Using Core and Well Log Data

Accurate estimation of horizontal in-situ stress, minimum (Sh min ) and maximum (SH max ), is crucial for subsurface planning, wellbore stability, and hydraulic fracturing in hydrocarbon and geothermal reservoirs. However, estimating horizontal in-situ stress is hindered by analytical or traditional neural network models, limited experimental data, formation uncertainties, missing tectonic stress data, and complex well-log correlations. Therefore, this study presents a hybrid model of group method of data handling (GMDH) with Levenberg-Marquardt (LM) to address these limitations. Using 10,105 well-log datasets of three wells, including nine input parameters such as (AC, Vs, Es, DEN, PP, CNL, GR, RT, and CALI) to predict horizontal in-situ stress of the "M" field of the Junggar Basin, China. The GMDH-LM model utilizes non-parametric regression techniques, providing enhanced flexibility, decreasing execution time, overcoming overfitting, and precision in addressing complex nonlinear challenges. The results show that the GMDH-LM model outperforms both GMDH and backpropagation neural networks (BPNN) models in terms of accuracy, as it efficiently estimated Sh min and SH max with a high coefficient of determination (R = 0.9948), low root mean square error (RMSE = 0.0401), and mean absolute error (MAE = 0.0241) during training. During testing, it delivered (R = 0.9949), (RMSE = 0.0502), and (MAE = 0.0237) and a shorter execution time of 0.82 seconds. Furthermore, Sensitivity analysis identified AC, Vs, ES, DEN, and PP as dominant input variables that enhance model prediction performance. The GMDH-LM was validated by applying it to the blind well (Well 004) and two additional wells, which accurately predicted horizontal stresses. Ultimately, this research advances horizontal stress modeling for optimizing the drilling process and maintaining well integrity and decision-making processes within the geotechnical and petroleum industries. • Hybrid group method of data handling with Levenberg-Marquardt (GMDH-LM) was proposed to predict horizontal stress. • Sensitivity analysis was used to examine the effect of input data on the model's performance. • The GMDH-LM shows greater precision with less computation time during predictions than GMDH and BPNN. • The GMDH-LM delivers reliable predictions of horizontal stress for other exploration wells. • This field is tectonic-lithologic reservoir influenced by tectonic forces and lithology.