Groundwater and gravity modelling using recurrent neural networks

Abstract

Artificial intelligence has gained a lot of interest in recent years due to its impressive performance on a variety of prediction and classification tasks. These results are in large part a result of the increase in computational capacity and the availability of large data sets. Here we apply artificial intelligence on the problem of predicting local variations of gravity due to local hydrological effects at the Onsala Space Observatory. Hydrological effects imply redistribution of mass in close proximity to the measuring equipment, and therefore contributes to the variations seen in the local strength of gravity. The objective of this thesis is to model groundwater and residual gravity levels based on data from a normal weather station, giving us the ability to construct useful simulated groundwater and residual gravity data measurements without the actual equipment in place. The modelling is performed using recurrent neural networks containing LSTM units, and with some convolutional preprocessing of the input data. The resulting models perform very well for predicting groundwater levels, reaching an RMSE between observed and predicted values of 0.046 m over the course of 2019 which is comparable to other results using neural networks to predict hourly groundwater levels ([2], [3]). The neural networks initially performed quite poorly when predicting residual gravity. However, by utilizing our good results from the groundwater predictions, we extended our available data set from about 3 to almost 8 years. The best model trained on the 8 year data set achieved an RMSE of 5.961 nms−2 on the test set, a more than 6-fold improvement over the networks trained on the 3 year data set. The conclusion is that recurrent neural networks are suitable for modelling groundwater levels and residual gravity, but their performance is highly dependent on the amount of data available and the preprocessing applied to the data.