Comparison of missing value imputations for groundwater levels using multivariate ARIMA, MLP, and LSTM

Groundwater monitoring wells in this country may have missing values due to a malfunction of sensor or an abnormal power supply. Processing of missing values is essential for hydrological assessment using groundwater level time series, interaction analysis between surface and ground water, and estimation of natural recharge rate. In this study, multivariate ARIMA and artificial neural network models (MLP and LSTM) are applied to fill the one-month missing period of groundwater levels at the Gapyung-gapyung (GPGP) monitoring station, and the accuracy of the models is compared based on RMSE and MAE. The multivariate ARIMA and LSTM models, which use both groundwater level data of autoregressive properties and nearby rainfall data as input variables, are evaluated to be more accurate than the MLP model. It is found that the error of the LSTM model is hardly decreased even if the number of rainfall and groundwater monitoring stations increases. As the LSTM model depends on the training and validation data, further studies on the method of selecting input variables need to be conducted for the consistency of missing value imputations.