Long-term prediction of groundwater levels for climate scenarios with machine-learning tools

Estimating the effects of climate change on water resources is an important topic for researchers in hydrological sciences. Climate model outputs can be used as inputs for calibrated numerical groundwater models to predict the effect of different climate scenarios on groundwater levels. Unfortunately, spatially explicit numerical models are spatially constrained, data-hungry, and difficult to set-up and to calibrate. Furthermore, the involved model run-times make proper uncertainty analysis of the model predictions computationally expensive.
Machine-learning (ML) tools have become recognized as a powerful alternative for numerical models for different applications in hydrological science, but come with their own challenges. They have been successfully used for the prediction of groundwater levels in single bores based on historical data, but their application for estimating groundwater levels for climate scenarios is still a matter of active research.
To identify the potential of ML techniques for this application, two different ML algorithms have been applied to predict groundwater levels for several climate scenarios at a number of groundwater wells with long-term historical data time series. The two ML algorithms are (i) multi-layer perceptrons (MLP) with a closed feedback loop and (ii) long short-term memory (LSTM) networks. For each observation well, several thousand versions of these ML models with differing setups are trained to historical, monthly data time series up to 2015 and then applied for the estimation of monthly groundwater levels. These estimations are computed by using climate model outputs for different scenarios, as well as artificially generated scenarios, as drivers for the ML models to test their sensitivity and plausibility to those input data series. The high quantity of model versions for each bore are utilized to generate mean groundwater level estimates and accompanying uncertainty bands via Bayesian Model Averaging (BMA).
Both ML techniques are able to match the historical data time series for the different bores with small uncertainty, but differ in their ability to predict long-term groundwater levels from climate change and artificial scenarios. The long-term simulations of the MLP models show believable trends and appropriate uncertainty bands, while the LSTM networks seem to underestimate the uncertainty of their future predictions.