Spatiotemporal forecasting for groundwater level using a WT-LSTM model

Accurate groundwater level forecasting models is essential to ensure the sustainable utilization and efficient protection of groundwater resources. In this paper, a novel method for groundwater level forecasting is proposed on the basis of coupling discrete wavelet transforms (WT) and long and short term memory neural network (LSTM) . In this model, the wavelet transform is used to decompose the cumulative displacement into the term of trend and term of periodicity . The trend term reflects the long-term tendency of groundwater level variation, which is simulated by a linear regression method. The periodic term driven by external factors such as rainfall, the river stage and the distance from river, is modelled using a LSTM method. The distance from river and the distance from observation wells are used for spatiotemporal model interpretation. Finally, the trend term and periodic term are superposed to achieve the cumulative spatiotemporal prediction of groundwater level. A typical study area located in Haihe basin is taken as an example to validate the performance of the proposed model. The proposed mode (WT-LSTM) is compared with the regular artificial neural network (ANN) model and autoregressive integrated moving average (ARIMA) model. The results show that the prediction accuracy of WT-LSTM model is higher than ANN model and ARIMA model, especially during the flood period. Furthermore, the spatiotemporal groundwater level forecasting is not only included the observation of groundwater and precipitation, but should also take the influence factors of surface water into consideration. The proposed model gives a new sight in the prediction of groundwater level.