Interpolation and Prediction of Piezometric Multivariate Time Series Based on Data Augmentation and Transformers

The aim of this study is to predict groundwater levels for several stations in France using a dataset composed of piezometric data, precipitation, evaporation, and stream flow as input to train the models. One of the primary challenges addressed in this study is the data gaps, which are interpolated to fill the missing values. The study is divided into three main parts. In the first part, we analysed the data by running various statistical tests to identify trends and patterns in the data. The second part of the study focused on interpolation, which involved filling in the missing values in the dataset. We compared four approaches: XGBoost, linear regression, KNN, and our proposed approach called SMOTEINT. The third and final part of the study involved predicting groundwater levels using several approaches. We compared three existing methods: VAR, LSTM, and GRU. Our main contribution to this part of the study was the application of transformers in groundwater prediction. We evaluated the performance of each method and identified the most effective approach for predicting groundwater levels using the given dataset. Overall, the study provides valuable insights into the use of various approaches for analyzing and predicting groundwater levels, which can have important implications for resource management and environmental protection efforts.