Hybrid wavelet–machine learning models for forecasting aquifer changes in storage: structural validation across aquifers in a semiarid catchment, Central Mexico

Workflow showing hydro-meteorological inputs, wavelet-hybrid models, accuracy and structural validation, with no single best model. Groundwater is a vital resource in semiarid regions, yet its sustainability is increasingly threatened by overexploitation and climate variability. Reliable prediction of groundwater storage changes is essential for effective water management, particularly in data-scarce basins. This study evaluated multivariate time-series models for forecasting semiannual groundwater storage changes in four aquifers of the Cuenca Alta del Río Laja catchment, Mexico. Five data-driven approaches were tested: multiple linear regression, artificial neural networks, support vector regression, random forest, and extreme gradient boosting, together with their wavelet-based hybrid versions. Hydro-meteorological variables, vegetation indices, and climate indices were used as predictors, while principal component analysis reduced multicollinearity among inputs. Model performance was assessed using accuracy metrics and Taylor diagrams. Results showed that no single model consistently outperformed the others across all aquifers. Wavelet-based support vector regression achieved the highest predictive performance in two aquifers (R2 = 0.585 and 0.624), while wavelet-based artificial neural networks performed best in the remaining two (R2 = 0.758 and 0.378). Corresponding mean absolute errors were 0.290, 0.242, 0.601, and 0.374 m. Following independent evaluation, this framework may be applied to other semiarid, data-scarce regions to support groundwater forecasting, water allocation, drought preparedness, and sustainable management.