Prediction of Watershed Groundwater Storage Based on Seasonal Adjustment and NAR Neural Network

  Objectives  Accurate and reliable prediction of watershed groundwater storage can help ensure the sustainable use of a watershed's aquifers for urban and rural water supply.

  Methods  On the basis of precipitation information, gravity recovery and climate experiment (GRACE) and global land data assimilation system (GLDAS), the prediction of groundwater storage is performed by using the seasonal adjustment and non-linear autoregressive(NAR) neural network. The NAR model without the seasonal adjustment, the autoregressive (AR) model, and the seasonal autoregressive integrated moving average (SARIMA) model are compared.

  Results  Taking the Changjiang Basin, the Lena Basin, the Ob Basin and the Yenisey Basin as case studies, the results indicate that the deseasonalized precipitation is independent and the groundwater is good fit to AR(1), which lay the foundations of deciding the number of a time delay of the NAR neural network. The performance of the NAR neural network using the seasonal adjustment falls into the excellent category for each basin and shows superiority over AR model and SARIMA model, with root mean square error less than 1 cm and correlation coefficient more than 0.96.

  Conclusions  The integration of seasonal adjustment technique and NAR neural network can not only improve the prediction accuracy of groundwater storage, but also reduce the convergence time. The proposed method can effectively forecast the groundwater storage with improved performance due to the seasonal adjustment that reduces the data complexity.