Advance prediction of coastal groundwater levels with temporal convolutional and long short-term memory networks

Abstract

Prediction of groundwater level is of immense importance and challenges coastal aquifer management with rapidly increasing climatic change. With the development of artificial intelligence, data-driven models have been widely adopted in hydrological process management. However, due to the limitation of network framework and construction, they are mostly adopted to produce only 1 time step in advance. Here, the temporal convolutional network (TCN) and models based on long short-term memory (LSTM) were developed to predict groundwater levels with different leading periods in a coastal aquifer. The initial data of 10 months, monitored hourly in two monitoring wells, were used for model training and testing, and the data of the following 3 months were used as prediction with 24, 72, 180, and 360 time steps (1, 3, 7, and 15 d) in advance. The historical precipitation and tidal-level data were incorporated as input data. For the one-step prediction of the two wells, the calculated R2 of the TCN-based models' values were higher and the root mean square error (RMSE) values were lower than that of the LSTM-based model in the prediction stage with shorter running times. For the advanced prediction, the model accuracy decreased with the increase in the advancing period from 1 to 3, 7, and 15 d. By comparing the simulation accuracy and efficiency, the TCN-based model slightly outperformed the LSTM-based model but was less efficient in training time. Both models showed great ability to learn complex patterns in advance using historical data with different leading periods and had been proven to be valid localized groundwater-level prediction tools in the subsurface environment.