Cloud imagery provides critical parameters—including temperature, precipitation, and cloud properties—that are valuable for scientific research. Real-time cloud image forecasting could greatly extend the utility of such data for practical applications, yet large-scale cloud extrapolation and prediction remain challenging due to limitations in accuracy, high computational costs, and cumulative errors. To address these challenges, we introduce GeoAttX, a novel deep learning framework that integrates geographic attention mechanisms to forecast cloud movement using data from the FengYun-4B (FY-4B) geostationary meteorological satellite, which updates every 15 minutes. The framework employs two core algorithms: a Reversed Greedy Algorithm (RGA), which reduces the number of autoregressive steps by reversing the order of the standard greedy approach, and Second-order Standardization (SoS), which re-normalizes pixel distributions to maintain consistent brightness statistics across inputs, mitigating error accumulation and improving accuracy. GeoAttX demonstrates strong performance in cloud dynamics prediction, achieving R² values of 0.9938, 0.9539, and 0.7489 for lead times of 15, 60, and 180 minutes, respectively. These results represent a substantial advance in satellite-based cloud image forecasting via deep learning, offering valuable insights for atmospheric modeling and disaster management. The source code is available at https://github.com/songzijiang/GeoAttX.