Accurate short-term cloud forecasting is essential for solar radiation and solar energy applications, where rapidly evolving cloud fields introduce substantial variability in surface solar irradiance. In this study, we develop a hybrid ConvLSTM-UNet deep learning model for 15-min-ahead forecasting of cloud optical thickness (COT) from Meteosat/SEVIRI observations over Europe. The model combines multiscale spatial feature extraction with recurrent spatiotemporal learning and employs a combined Charbonnier–SSIM loss to improve both numerical accuracy and the preservation of cloud structures. The proposed forecasting model is evaluated against real COT satellite retrievals and benchmarked against persistence and a dense optical flow cloud motion vector (CMV) approach. For 15-min-ahead forecasts, ConvLSTM-UNet achieves lower COT prediction errors and higher perceptual quality than both benchmarks, with MAE and RMSE of 1.25 and 9.52, respectively, and SSIM of 0.95. The impact of COT forecast errors on solar radiation is further assessed through the cloud modification factor (CMF), for which the proposed model achieves a forecast skill of approximately 30 % relative to persistence. Its performance advantage is particularly evident under cloudy and rapidly evolving conditions, and remains evident under recursive forecasts extending to approximately 165 min ahead. These results demonstrate the potential of deep spatiotemporal learning for satellite-based short-term cloud and solar radiation forecasting.