Prediction of water level variations in aquifers using recurrent and convolutional neural networks

The use of artificial intelligence models to simulate and forecast groundwater has become more and more popular since the early 1990ies. These models able to forecast groundwater levels are increasingly being used to support decision making for water resources management and climate change impact local assessments. This study presents a comparison of different algorithms to simulate and forecast the piezometry of the Astian sands aquifer in the Occitanie region (France) at sub seasonal to seasonal scales. The aim is to set up a transferable methodology allowing quick implementation on different aquifers, to lay the foundations for a tool for water resources management focusing on droughts.

Different models were tested in this study with prediction horizons from a few weeks and up to the hydrological year. Models used include (1) Multi-layer Perceptron (MLP), (2) recurrent models such as Long short-term memory models, Gated Recurrent Unit (GRU), Non-Linear Regressive Neural Networks (NARX), Dual Stage Attention Based Recurrent Networks (DA-RNN); (3) a Temporal Convolutional Network (TCN); (4) a Reservoir Computing Network (Echo State Network). Pre-processing methods were tested to increase performance and attempt to better reproduce higher frequencies, especially the Ensemble Empirical Modal Decomposition (EEMD). Input variables were the daily rainfall data, temperature data and piezometry. To validate the model and assess the performance of the simulations, different metrics were used (Nash Sutcliffe Efficiency, Kling Gupta Efficiency, Coefficient of persistency). The loss function was a combination of RMSE and NSE, the optimizer used was Adam, the training was done on two thirds of the time series and validation and testing on the remaining data.

Several models achieved high performances: DA-RNN, the GRU, MLP and NARX (NSE>0.91 and Cp>0.6 at 20 days). Predictions were carried out iteratively with a 20-days window, reinjecting the predicted data and re-training the model (adaptative model) until a prediction of more than 400 days was reached (using observed rainfall and temperature). The stability and robustness of the model is further investigated through additional cross validation, estimation of confidence intervals, and simulations in different hydrogeological contexts, to aim toward a transferable tool for various aquifers.