基于稠密连接网络的地下水污染替代模型研究

A DenseNet Network Based Surrogate Model for Simulating Contaminant Transport in Groundwater Systems

摘要:

地下水污染溯源和含水层参数反演问题等地下水逆问题通常面临计算负荷量的制约，考虑使用替代模型作为解决方法，从而降低地下水反演问题的计算成本。借鉴卷积神经网络的图像识别过程，考虑将地下水流运动和污染物运移问题转化为输入场图像（渗透系数场、污染源信息等）与输出场图像（水头场、浓度场等）之间函数关系的图像回归问题，利用基于稠密连接网络的AR-Net-WL构建地下水流运动和污染物运移的替代模型。算例研究表明，针对替代模型的过拟合现象，尽可能选择较大的训练样本可获得约10%的精度提升；当没有条件增加训练样本时，采用最优正则项系数的AR-Net-WL在训练样本较少的情形下（训练样本500）也能够取得良好的性能，能够精确预测地下水流运动和污染物运移。

Abstract:

Groundwater inverse problems such as groundwater contaminant identification and aquifer parameter inversion problems are usually restricted by the computational load. In order to reduce the computational cost of groundwater inversion， the surrogate method is a feasible solution. In this paper， imitating the image recognition process of the convolutional neural network， the groundwater flow movement and contaminant transport problem is transformed into an image regression problem of the functional relationship between input image （hydraulic conductivity field， pollution source information， etc.） and output image （groundwater level， contaminant concentration， etc.）. The surrogate model of groundwater flow movement and contaminant transport is constructed by using AR-Net-WL based on the DenseNet network. The case study shows that， for the overfitting of the surrogate model， a 10% improvement in accuracy can be obtained by selecting training samples as large as possible. When there are no conditions to increase the training sample， AR-Net-WL with an optimal regularization coefficient can also achieve a good performance with fewer training samples （500 training samples） and can accurately predict the groundwater flow movement and contaminant transport.