Inversion analysis of the stress field in a pumped-storage power station based on the PSO-XGBoost and the finite element model.

The inversion of geostress is key to understanding the crustal structure, predicting seismic activity, and guiding underground engineering design. A hybrid method that integrates the particle swarm optimization (PSO), extreme gradient boosting (XGBoost), and a finite element model (FEM) was proposed. A high-precision non-linear proxy model was established through global optimization of hyper-parameters of XGBoost using PSO, and the optimal boundary conditions of the FEM were obtained by inversion. In this way, the accurate distribution of the geostress was obtained, overcoming the issues arising from non-linearity and discontinuity in the inversion of deep geostress fields. The inverted results of the stress field in Hongping pumped-storage power station (Jing'an County, Jiangxi Province, China) show that the shallow part is dominated by the horizontal stress while the vertical stress gradually becomes dominant with increasing depth. The inversion results of geostress agree with the measurements.