Research On Data-driven Prediction Models Of Regional Groundwater Depth

Groundwater depth prediction is of great scientific significance for regional groundwater development and utilization,water resources evaluation and management and geological disaster management.This paper summarizes the research progress of random model,grey model,fuzzy model,artificial neural network prediction model and hybrid prediction model,analyzes and summarizes the achievements and existing problems,and selects the monthly groundwater depth data of 20 monitoring wells in Guanzhong region of Shaanxi province for the deficiencies in the current research.Six error indicators,namely average absolute error(MAE),average relative error(MRE),root mean square error(RMSE),mean square percentage error(MSPE),coefficient of determination(R~2)and Nash-Sutcliffe efficiency coefficient(NSE),as well as prediction qualification rate(QR)and accuracy classification of prediction items,are selected to carry out research and comprehensive evaluation on prediction model of groundwater depth sequence in the study area.The main research contents and conclusions of this paper are as follows:(1)Three single artificial neural network prediction models,namely BP network,support vector machine(SVM)and kernel extreme learning machine(KELM),were established,and the prediction models based on single artificial neural network were used for groundwater depth prediction in the study area.The study found that during the training period,the three prediction models show good fitting results for all monitoring wells.During the verification period,the prediction performance of SVM model and KELM model is better than that of BP network model.The prediction results of BP network model are chaotic and have poor regularity.However,SVM model and KELM model can better predict the trend change and periodic fluctuation characteristics of groundwater depth sequence.It may be because SVM model and KELM model introduce regular penalty factors,which can effectively avoid over-fitting of data in training period and further improve the generalization ability of the model.At the same time,the network search cross-validation parameter optimization method and PSO optimization algorithm can prevent the influence of trial and error method and subjective determination,and provide a reasonable,scientific and effective way for the determination of model parameters.At the same time,SVM model and KELM model have different prediction capabilities,but the two models have the common disadvantage that they have poor prediction capabilities for extreme points and there is a one-month delay in the prediction sequence.(2)Based on wavelet analysis(WA)denoising method and variational mode decomposition(VMD)denoising technology,a new hybrid artificial neural network prediction model(WA-BP,WA-SVM,WA-KELM,VMD-BP,VMD-SVM,VMD-KELM model)was constructed.The prediction results of groundwater depth in the study area show that the prediction ability of the hybrid prediction models of BP network model are worse than that of SVM and KELM.The hybrid model of SVM and KELM can improve the prediction ability of the trend term,periodic fluctuation term,extreme points of groundwater depth and the one-month delay error in prediction,but it is still difficult to accurately predict the extreme points.Among the hybrid prediction models based on wavelet denoising,WA-SVM has higher prediction accuracy.Among the hybrid prediction models based on VMD denoising technology,VMD-KELM has higher prediction accuracy.This shows that wavelet denoising technology is more suitable for SVM model combination,while VMD denoising technology is more suitable for combination with KELM model to form a hybrid prediction model with high accuracy and strong applicability.(3)Optimization of groundwater depth prediction model were given in the study area.The results show that the optimal models are all improved hybrid models.The optimal model of monitoring well W19 is WA-SVM model,the optimal models of#100,K423 and W15-1are VMD-SVM model,and the optimal models of monitoring wells K110,K214,K106,#85-1,J16,589,N16,E12-1,232,261,B9,267,B557,CQ19,W25-2 and W15 are VMD-KELM model.These results show that the hybrid model can reduce the prediction error and improve the prediction accuracy.This may be because the denoising method applied can effectively extract the trend component and periodic component in the groundwater depth sequence,reduce the noise component in the sequence,and improve the prediction accuracy.At the same time,the optimal model of 19 monitoring wells in the study area is a hybrid prediction model based on VMD decomposition and denoising technology,and only one monitoring well is a hybrid prediction model based on WA denoising.This indicates that the new VMD decomposition denoising technology proposed is more suitable for coupling with the artificial neural network prediction model to predict the groundwater depth in the study area than the wavelet threshold denoising method.To sum up,through the construction of the data-driven model of the monitoring wells in the study area and the model evaluation based on the selected error index and accuracy evaluation index,it can be found that the improved hybrid model can better utilize the information provided by the data,improve the prediction accuracy and enhance the applicability of the model.