Rok Klančič and Klemen Kenda
Abstract
This paper presents a comparative evaluation of three distinct
categories of models applied to groundwater level data: tradi-
tional batch learning methods, time series deep learning methods,
and time series foundation models. By enriching the water level
data with weather-related features, we significantly improved
the effectiveness of simpler models. The results demonstrate that,
despite their state-of-the-art performance on univariate datasets
and the corresponding publicity, advanced models without con-
textual feature support are still surpassed by traditional methods
trained on enriched datasets.