Nonlinear Intelligent Inversion Method and Practice for In-situ Stress in Stratified Rock Masses with Deep Valley

Accurate initial stress conditions, including their magnitude and orientation, are a prerequisite for evaluating the deformation and failure of surrounding rock masses in underground caverns, as well as for assessing dynamic failure risks and optimizing the layout of the cavern’s axes. Yet, the general stress inversion method based on limited measured data points and numerical back analysis always faces the challenge of non-uniqueness in the model’s boundary conditions, which results in inaccuracies of the obtained stress field. This study established an intelligent inversion method for the stress field in stratified rock masses with deep valleys, considering the effects of river erosion. It proposed a nonlinear inversion algorithm based on Long Short-Term Memory Networks-Hybrid Optimization (LSTM-HO), which leverages the structural advantages of LSTM to construct a precise nonlinear surrogate model between measured geo-stress and the model’s boundary conditions, and by utilizing the fitness function of HO to search the reasonable boundary conditions. By inputting measured geo-stress, the HO can rapidly generate several potential boundary condition vectors of the numerical model. The cavern’s excavation damage zone (EDZ) is then employed for secondary screening, refining these possibilities to a unique, reliable boundary vector. This approach effectively addresses the non-uniqueness issue of the model’s boundary conditions in stress field inversion processes. The method is applied in the engineering site of Kala Hydropower Station. A reliable regional stress field consistent with measured stresses and the cavern’s EDZ depth is obtained in steeply inclined layered rock masses with V-shaped river valleys, which demonstrates the reliability and applicability of the presented method.