3D in-situ stress prediction for shale reservoirs based on the CapsNet-BiLSTM hybrid model

In-situ stress is essential in shale reservoir fracturing, driving oil and gas migration and informing wellbore stability and drilling optimization. The accurate prediction of 3D in-situ stress is inseparable from seismic data. However, existing methods predominantly rely on empirical formulas or simplified assumptions, which limit their accuracy in representing the real distribution of in-situ stress. Furthermore, these methods often predict in-situ stress from a single factor, leading to high uncertainty. To address these, we propose a method for predicting 3D in-situ stress that leverages a hybrid Capsule Network-Bidirectional Long Short-Term Memory (CapsNet-BiLSTM) model. This approach takes into account various factors, such as geological features and seismic attributes, to achieve more accurate predictions. First, we analyze the structural characteristics of shale formations and use rock petrophysical knowledge to reasonably filter input data, eliminating the impact of redundant parameters on in-situ stress prediction. Then, to overcome the limitations of traditional deep learning models in capturing correlations within complex data structures, we construct a CapsNet-BiLSTM network model. This model integrates the spatial relationship modeling capability of CapsNet and the temporal modeling capability of BiLSTM, better accounting for the anisotropic features and temporal sequence information of shale reservoirs. Applying this method to a study area in the Sichuan Basin demonstrates that the constructed CapsNet-BiLSTM hybrid model accurately predicts in-situ stress values, effectively capturing the spatial distribution patterns of complex in-situ stress within shale reservoirs, thus proving the effectiveness and potential of our method in geological engineering applications for shale oil and gas reservoirs. This hybrid model-based prediction method not only improves the accuracy of in-situ stress prediction but also provides a valuable methodological and technical support for scientific research and engineering practices in related fields.