Prediction of three-dimensional stress in fracture zones under multi-water-level steady-state conditions based on deep learning
摘要
The quick estimation of three-dimensional stresses in fault zones is one of the challenges associated with the monitoring of reservoir safety. Conventional thermo-hydro-mechanical coupled numerical models are too complex to calibrate parameters and incur large computational costs, and thus cannot satisfy the need to forecast rapidly. In this work, we have compared four deep learning models (EB-GRU, P-GRU, ST-GNN-GRU, and ST-Transformer) based on the numerical simulation results of 62 monitoring stations of the Maluchi Fault Zone located at the Three Gorges Reservoir. Each of the four models was tested under the same conditions of data preprocessing, training, and testing. The results showed that the ST-Transformer model achieved the best accuracy for all measurement criteria on the mixed sample, with an average RMSE of 8.394 × 105 Pa, a Global R2 of 0.9997, and a best-run RMSE of 7.356 × 105 Pa, although its calculation cost was the highest. On the other hand, P-GRU demonstrated the most balanced overall performance, with an average RMSE of 1.143 × 106 Pa and a training time of only 268 s. In addition, P-GRU achieved the best performance in both low- and high-stress areas, showing greater robustness at the ends of the stress distribution. ST-GNN-GRU has a lower mean absolute error and greater stability in iterative learning, but a small number of significant deviations still remain. EB-GRU serves as a lightweight and valuable reference model. Further analysis shows that the performance of a model cannot be assessed solely on the basis of aggregate indicators from mixed samples. In this article, we propose a layered deployment strategy by combining the results of segmented errors, representative-node sequence fitting, and the spatial distribution of errors. In areas where critical nodes and local errors remain high, the ST-Transformer model is preferred.