Abstract: Based on in-situ stress measurement and three-dimensional (3D) inversion computing model of Huangdeng large-scale underground cavern, this study is to revearl the distribution characteristics of in-situ stress fields and further to provide accurately initial material for the design of the underground excavation and reinforcement. The inversions of in-situ stress fields are obtained by the multiple linear regression (MLR), the artificial neural network (ANN) and the stepwise linear regression (SLR) combined with ANN, respectively. Moreover, the SLR-ANN method is characterized by the nonlinear intelligent inversion analysis, which especially considers the geological history process. In addition, the results of inversion analysis are proofread and examined with measured in-situ stress results. It shows that the inversion results by these three methods are in good agreement with measured results. The results indicate that these three methods can truly reflect the distributions and characteristics of 3D in-situ stress in the whole underground cavern group. Compared with these three methods, the results by SLR-ANN algorithm are closest to measured results. This method also significantly improves the inversion efficiency by reducing inversion parameters. Therefore, it can be practically applied in realistic scenarios to achieve efficient and accurate estimations of in-situ stress in rock mass.

Key words: hydraulic engineering, in-situ stress field, stepwise linear regression (SLR), artificial neural network (ANN), quadratic back analysis