Introduction: Despite favorable hydrological conditions, the Koohdasht plain groundwater faces stress and declining water table levels.

Objectives: This research aims to determine and predict changes in water level and groundwater storage by integrating analytical models, time series, and machine learning.

Materials and Methods: Time series models, including AR, MA, ARMA, ARIMA, and SARIMA, were combined with machine learning algorithms, including Linear Regression, Random Forest, and Gradient Boosting, to forecast groundwater fluctuations.

Findings: Spatial zonation revealed the greatest declines in the northeastern, southeastern, central, and parts of the western aquifer. The combined models achieved excellent performance, with R² exceeding 0.98 and RMSE below 0.4 m for several piezometric wells.

Conclusion: Over 15 years, groundwater levels dropped by 5.9 m, with storage deficits reaching up to 15 MCM in dry years. Among all hybrids, the time series–linear regression combination performed best across different wells. This model projected aquifer storage deficits of approximately –0.7 and –1.22 million m³ for the 2025–2026 and 2026–2027 water years, respectively. The integrated approach demonstrates strong potential for early warning and sustainable management of groundwater resources in semi-arid regions, offering a replicable framework for similar aquifers facing anthropogenic and climatic pressures.