Advance prediction of coastal groundwater levels with temporal convolutional network

Abstract

Abstract. Prediction of groundwater level is of immense importance and challenges for the coastal aquifer management with rapidly increasing climatic change. With the development of artificial intelligence, the data driven models have been widely adopted in predicting hydrological processes. However, due to the limitation of network framework and construction, they are mostly adopted to produce only one-time step in advance. Here, a TCN-based model is developed to predict groundwater level variations with different leading periods in a coastal aquifer. The historical precipitation and tidal level data are incorporated as input data. The first hourly-monitored ten-month data were used for model training and testing, and the data of the following three months were predicted with 24, 72, 18 and 360 time steps in advance. For one-step prediction of the two wells, the calculated R2 are higher than 0.999 in the prediction stage. The performance is meanwhile compared with a powerful network in the field of time-series prediction, long short-term memory (LSTM) recurrent network. The corresponding R2 of the LSTM-based model are 0.996 and 0.998. While the RMSE values of TCN-based model are less than that of LSTM-based model with shorter running times. For the advanced prediction, the model accuracy greatly decreases with the increase of advancing period from 1-day to 3-, 7- and 15-days. Overall, the TCN- and LSTM-based models show great ability to learn complex patterns in advance using historical data within the time series. Considering the simulation accuracy and efficiency, the TCN-based model outperforms the LSTM-based model and has been proved to be a valid localized groundwater prediction tool in the subsurface environment.