Application Of Two Deep Learning Algorithms In Groundwater Simulation

Simulating and predicting water volume,water level,the distribution of pollution plume and phase saturation is an important task of groundwater research.With the development of economy and society,there are many demands for predicting groundwater accurately and considering complex processes from multi-disciplinary research.In groundwater research,Numerical simulation is the most extensive method which based on physical processes.However,it has some practical limitations that cannot satisfy the requirement well.For example,it is difficult to obtain detailed hydrogeological parameters and depict underground medium structure finely,it also takes expensive cost in uncertainty analysis and management optimization.Although traditional machine learning algorithms provide a convenient and effective alternative,their applicability to data with large dynamic changes and strong nonlinearity is low,and the problem called"disaster of dimensionality"is commonly encountered when processing high-dimensional data.In response to the above difficulties and challenges,we propose two models for groundwater forecasting which based on deep learning algorithms,namely the PSR-Bi-LSTM(Phase Space Reconstruction-Bidirectional-Long Short-Term Memory)model for time series prediction and the c DC-GAN(conditional Deep Convolutional Generative Adversarial Network)model for spatial distribution prediction.The two models were applied to actual cases respectively.The PSR-Bi-LSTM model was used to predict the dynamics of the S31 spring at Yaji Karst Test Site in Gui Lin.The c DC-GAN model was used to be surrogate model of reactive solute transport in order to uncertainty analysis.The results show that the proposed models are quite useful.The PSR-Bi-LSTM model predicts spring discharge accurately.When prediction step is single,the correlation coefficient reaches 0.986,and the Nash coefficient,relative deviation and root mean square error are 0.963,10.701%and 0.021m~3/s,respectively.The precision will gradually decrease as the prediction step size increase.The study found it is acceptable when prediction step is between 1-4.The c DC-GAN model can get corresponding distribution of uranium concentration when changes well spacing and heterogeneous fields.In the verification set,the average Structural similarity and Mean square error of the model reaches 0.92 and 0.0042respectively.The c DC-GAN model allows heterogeneous fields as variable to input model directly,which provides an image-to-image regression method,solved the problem of"disaster of dimensionality".The study also shows that it is cheap and convenient to use the c DC-GAN model as surrogate model for uncertainty analysis.The proposed models based on deep learning algorithm not only predict time series and spatial distribution effectively,but also improve the efficiency of uncertainty analysis and management optimization,which provide a new method for groundwater related research.