The Chah-Nimeh reservoirs store the water that supplies the Sistan plain in south-eastern Iran, but their storage is not publicly monitored in real time. This study analyses a monthly satellite-derived record of the Chah-Nimeh 3 reservoir (December 2018 – December 2025, 84 months) that contains water-surface area, water-surface elevation, satellite-derived depth and four storage estimates (linear, quadratic, spectral and hybrid area–volume models), and develops one- and three-month-ahead forecasting models for the hybrid storage. Multiple linear regression, support vector regression (SVR), random forest (RF), an LSTM network and two hybrid models tuned by the grey wolf optimizer (SVR-GWO) and Harris hawks optimization (XGB-HHO) were compared with a persistence benchmark under a chronological split (training 2018–2022, validation 2023, test 2024–2025). Storage declined significantly over the record (Mann–Kendall Z = -4.87, Sen's slope = -23.3 MCM/year), from an annual mean of 311 MCM in 2020 to 120 MCM in 2023, before recovering to 245 MCM in 2025. The four volume estimators were highly correlated (r ≥ 0.96) but differed by more than a factor of two in magnitude, the spectral estimator being the highest. At a one-month horizon the LSTM was the most accurate model (RMSE = 34.4 MCM, NSE = 0.72), marginally better than persistence (37.4 MCM). At three months the advantage of learning models became clearer: RF reached RMSE = 62.1 MCM and NSE = 0.07 against 68.4 MCM and -0.12 for persistence. Lagged storage, surface area and the seasonal terms were the dominant predictors. The results show that satellite storage records can support short-term reservoir operation, and that the benefit of data-driven models grows with the forecast horizon.