Transformers-based time series forecasting for piezometric level prediction

Abstract

Over the recent years, Neural Networks have been used as an alternative for classical numerical approaches in order to facilitate tasks like classification or prediction. The purpose of our study is to predict the groundwater levels for 18 stations scattered in France and explore the applicability of different approaches for prediction. The provided meteorological and hydrological data and the sensors measurements were used as our principal data source. A correlation between piezometric data, precipitation, evaporation and stream flow are used as input to train the compared models which are characterized by employing sequential data. The models have shown promising results as we mention the application of both GRU and Transformer architecture which is newly GW forecasting. The findings of this study can be used to give direction for future use of GW. Furthermore, this work can be a useful reference for other time series (TS) problems in geotechnical fields.