基于小波神经网络的地下水流数值模拟模型的替代模型研究

以吉林西部为研究区,建立地下水流数值模拟模型,分别应用蒙特卡罗方法和拉丁超立方方法在研究区10个县(市)开采量的可行范围内进行采样,经对比选择拉丁超立方抽样结果得到输入(开采量)—输出(水位降深)数据集,建立小波神经网络模型作为地下水流数值模拟模型的替代模型,而后对替代模型有效性作误差分析,并与多元非线性回归替代模型进行对比.结果显示,2种替代模型在功能上都能逼近地下水流数值模拟模型,但小波神经网络模型得到的水位降深均值和水位降深剩余标准差与模拟模型计算结果的相对误差分别低于多元非线性回归模型76%和45%,说明小波神经网络模型更适合作为地下水流数值模拟模型的替代模型,这为减少优化模型求解过程中直接调用模拟模型所造成的计算负荷提供了一种有效的替代方法.

Surrogate model of numerical simulation model of groundwater based on Wavelet Neural Network

Western Jilin was selected as the study area, and the groundwater numerical simulation model for this area was established. Monte-Carlo and Latin Hypercube method were applied to sample exploitation of 10counties (cities) as practicable range in the study area, Latin Hypercube Sampling was selected to obtain input (pumping) and output (water level drawdown) data sets, and wavelet neural network was proposed to establish an surrogate model of the groundwater numerical simulation model, Compared the fitting mean relative error of wavelet neural network model with that of multivariate nonlinear regression model. Two surrogate models both could approach the function of numerical simulation model of groundwater, however, the relative error of mean groundwater level drawdown and the remaining average relative standard deviation of groundwater level drawdown between wavelet neural network model and simulation model was smaller than the multiple nonlinear regression model 76% and 45%, which indicated that the wavelet neural network model can effectively replace groundwater numerical model. This study will provide an effective surrogate method to reduce computational load resulted from multiple invocation of the numerical simulation model of groundwater in the processes of iteration solution by optimization model.