Lake Urmia in north-western Iran has experienced one of the most severe hydrological declines of any large saline lake in the world, and the terrestrial water storage (TWS) of its basin is a key indicator for water-resources planning. This study evaluates classical machine learning (multiple linear regression, support vector regression and random forest), a deep learning model (long short-term memory, LSTM) and two hybrid models in which the hyperparameters of support vector regression were tuned by the Grey Wolf Optimizer (SVR-GWO) and the Whale Optimization Algorithm (SVR-WOA) for one-month-ahead estimation of GRACE-derived liquid water equivalent (LWE) anomalies. Monthly basin-average LWE (GRACE, 2003–2017) together with precipitation, air temperature, sea-level pressure and wind speed from the Iran Meteorological Organization and GLDAS soil moisture were used. The LWE series showed a highly significant decreasing trend (Mann–Kendall Z = -11.57, Sen's slope = -2.02 cm/year). All models were trained on 2004–2012, tuned on 2013–2014 and tested on 2015–2017 and were compared against a persistence benchmark. In the test period the LSTM gave the lowest error (RMSE = 3.16 cm, NSE = 0.76), followed closely by MLR (RMSE = 3.19 cm), SVR-WOA (3.30 cm) and SVR-GWO (3.31 cm), while the persistence benchmark reached 4.17 cm. Metaheuristic tuning reduced the test RMSE of SVR by about 2%, and all learning models improved on persistence by 17–24%. Soil moisture and air temperature were the most influential predictors. The small differences among the best models show that, for short and strongly seasonal GRACE records, simple well-regularized models remain competitive with deep learning.