基于DL‑ERT模型的地下水渗透系数预测方法研究

Research on Prediction Method for Groundwater Permeability Coefficient Based on DLERT Model

摘要:

针对传统方法刻画地下水含水层非均质性时面临的精度不足、预测成本高等问题，基于数值模拟和室内砂箱试验，通过残差网络优化集成卷积门控循环单元（CNN‐GRU）的强大数据学习能力和电阻率层析成像法（ERT）运用物理先验信息的优势，提出一种融合物理机理的深度学习算法—DL‐ERT 模型。将其对比多个传统反演模型， 探讨融合算法在地下水含水层的渗透系数刻画精度。结果表明：（1）模型的训练损失和验证损失都快速下降并趋近于零，且两者的收敛几乎同步，表明 DL‐ERT 模型的构建策略优秀，能快速有效的学习数据特征；（2）以某一测试集样本为例，对比 ERT、CNN‐GRU 和 DL‐ERT 对该样本的渗透系数反演云图，发现单一的算法模型均不能同时注重左右两侧的高渗区域刻画，而 DL‐ERT 则对高渗区域表现出极大的预测潜力，其拟合精度达到了 0.906；（3）制作室内砂箱试验，将融合算法与传统的克里金插值法、CNN‐GRU 以及 ERT 作对比运用，得到各模型的拟合精度值分别为 0.895、0.707、0.760 和 0.836。可以发现，DL‐ERT 确实在一定程度上弥补了单一算法的不足，相比于单一的 CNN‐GRU 和 ERT，其结果的预测精度提升了 7%~17%，表明了该模型在工程运用方面的潜力。

Abstract:

To address the issues of insufficient accuracy and high prediction costs faced by conventional methods in characterizing the heterogeneity of groundwater aquifers, this study proposed a physicsnformed deep learning algorithm—the DL-ERT model—based on numerical simulations and laboratory sandbox experiments. The model integrated the powerful data learning capability of a convolutional gated recurrent unit (CNN-GRU) optimized by residual networks with the advantage of physical prior information from electrical resistivity tomography (ERT). The DL-ERT model was compared with multiple traditional inversion models to examine the accuracy of the fusion algorithm in characterizing the permeability coefficient of groundwater aquifers. The results showed that: (1) the training and validation losses of the DL-ERT model rapidly decreased and approached zero, and their convergence was almost synchronous, indicating that the construction strategy of the DL-ERT model was excellent and that data features could be quickly and effectively learned. (2) Taking a sample from the test set as an example, the inversion cloud maps of the permeability coefficient obtained by ERT, CNNGRU, and DL-ERT were compared. It was found that individual algorithm models could not simultaneously capture the high-permeability zones on both sides, while DL-ERT demonstrated remarkable predictive potential for high-permeability zones, achieving a fitting accuracy of 0.906. (3) Laboratory sandbox experiments were conducted, and the fusion algorithm was compared with traditional Kriging interpolation, CNN-GRU, and ERT, yielding fitting accuracies of 0.895, 0.707, 0.760, and 0.836, respectively. It is evident that the DL-ERT model compensates for the limitations of individual algorithms to some extent, with prediction accuracy improved by 7%-17% compared with the individual CNN-GRU and ERT models, indicating the potential of the model for engineering applications.