Machine learning approaches for groundwater temperature prediction in Austria: Impacts of hydro-climatic variability, feature engineering, and model complexity

Seven Austrian sites, each representing unique hydroclimatic conditions. Accurate groundwater temperature (GWT) prediction is essential for water management, yet machine learning (ML) applications remain underexplored compared to groundwater level and stream temperature studies. This study introduces a novel comprehensive framework linking GWT regularity and long-term trends derived from entropy and decomposition to optimal model selection and feature engineering. Model performance was assessed across multiple algorithms, ranging from linear approaches to complex ensemble and deep learning, to determine optimal time lags under three predictor sets: (i) time only, (ii) climate only, and (iii) integrated time and climate. Predictor importance was examined using Stepwise Backward Selection (SBS) and Principal Component Analysis (PCA), followed by a direct comparison between the two methods. GWT regimes with high seasonal regularity and stable trends can be predicted accurately with simpler models, whereas irregular regimes benefit from advanced model architectures. Integrating temporal and climate predictors significantly outperforms climate-only models, achieving up to a 78 % NSE increase and a 62 % error reduction. SBS performs comparable to PCA while retaining physical interpretability, identifying air temperature and radiation as dominant predictors. These findings highlight a site-specific workflow for GWT prediction, combining hydro-climatic diagnostics with regionally tailored ML models to account for local seasonal and trend behaviors supporting improved and regional groundwater management under climate variability. • Machine learning and deep learning models predict accurate groundwater temperature. • Simple algorithms outperform the deep learning model in strongly seasonal settings. • Integrating time and climate features outperforms climate-only models. • Sequential backward selection determines the key climate parameters. • Spatial heterogeneity underscores the need for site-specific ML model selection.