Research On Developing Surrogate Model Of Groundwater Numerical Simulation Based On Deep Learning

Surrogate model technology refers to the establishment of a more convenient model to replace the original computationally expensive model,to avoid a large number of repeated calls to the original model,greatly saving the cost of calculation.As a huge and complex original model,the numerical simulation model of groundwater flow and solute transport needs to be called repeatedly for tens of thousands of times in the process of inversion and parameter calculation,which costs incalculable time and computer capacity.Therefore,the development of surrogate model is an important way to improve the calculation efficiency.In the area of surrogate model technique,the traditional artificial neural network(ANN)has a good surrogate ability.The traditional neural network can be used to build the surrogate model of the original model with a single time series output.However,the ability of traditional ANN to establish the surrogate of models with multiple time series output is limited.In recent years,deep learning methods have developed rapidly,and the Gated Recurrent Unit(GRU)network and Deep Belief Network(DBN)have shown good application prospects in various fields.In this paper,the deep learning method is used to establish the groundwater surrogate model,which mainly includes 1)construction of surrogate model of groundwater flow MODFLOW model and conduction of parameter auto calibration and global sensitivity analysis of MODFLOW based on the surrogate,2)utilization DBN to establish the surrogate model of groundwater transport numerical model in the Adaptive Neural Network Genetic Algorithm(ANGA),performing the optimization of groundwater pollution remediation schemes.The main achievements are as follows:(1)The surrogate model based on GRU network has a remarkable ability to approximate the simulation model with multiple time series output.The GRU surrogate model combined with particle swarm optimization(PSO)shows high efficiency and accuracy in automatic parameter calibration of models with high original calculation cost.(2)The GRU surrogate model is a good alternative to the original MODFLOW model for global sensitivity analysis.The results of global sensitivity analysis show that among the six MODFLOW parameters studied,conductivity is the most sensitive parameter of MODFLOW model,while porosity and anisotropy are not sensitive.The combination of GRU surrogate model and Sobol’ sensitivity analysis method can capture the time-varying characteristics of the original MODFLOW model.Results show that the influence of conductivity and recharge coefficient on MODFLOW output is stronger in the rainy season than in the dry season,and specific yield and maximum evapotranspiration rate have greater influence in the dry season.(3)In the parameter auto calibration and global sensitivity analysis of MODFLOW,the use of GRU surrogate model can save about 97% of the cost of calculation time.(4)The surrogate model of original MT3 DMS model based on DBN has a high precision.It is feasible to use DBN network to optimize groundwater pollution remediation scheme in ANGA framework.(5)Compared with the original model,the surrogate model based on DBN can save about 73.3% of the calculation time to optimize the groundwater pollution remediation scheme.With the ANGA framework,there is no need to design enough samples at the beginning,and the number of samples can be gradually increased during the optimization process to avoid increasing computation cost.