In energy systems with a high share of renewable energy supply (RES), accurate forecasting methods are very important for a secure and economical energy supply. For applications such as demand-supply balancing, short-term forecast (nowcasting) of spatial RES power generation is particularly important. In the literature, most models addressing this need use satellite-derived SSI estimations as input and optical flow or block-matching algorithms to predict their motion. Recently, new families of algorithms based on deep learning have appeared with great potential for improving the performance of nowcasting systems. It is thus of relevance to assess and understand the potential and limitations of deep learning approaches. Furthermore, when maps of solar irradiance are predicted, the metrics used to evaluate the forecasts are usually computed pixel-wise and thus ignore important spatial features, such as the consistency of spatial variability and spatial resolution. In this work, we investigate the potential of deep-learning algorithms for the forecasting of maps of 15-min solar surface irradiance (SSI) over short-term horizons, from nowcasts of SSI provided by CAMS radiation. We focus on a state-of-the-art deep learning model, designed for spatio-temporal processes: the convolutional long-term short-term network (ConvLSTM). We compare it to a “classic” forecasting model, based on an optical-flow algorithm (TVL1). We first perform a pixel-wise analysis of the models’ accuracy for forecasting horizons between 15 minutes and 3 hours. We then use Fourier spectral analysis to quantify the impact of each forecasting model on the spatial features of the SSI. Finally, we investigate the impact of the loss function used to train the convLSTM. Our results show that convLSTM and TVL1 have similar pixel-wise performances for short time horizons (15 and 30 minutes ahead), whereas, for larger horizons, convLSTM has a significantly lower RMSE and higher correlation. Fourier analysis, however, reveals that this improvement in pixel-wise accuracy comes with a degradation of the spatial features of SSI. TVL1 forecasts indeed have realistic spatial variability for all tested horizons, but convLSTM produces increasingly smooth predictions: for horizons beyond 2 hours, and despite its higher accuracy, convLSTM acts as a low-pass filter and fully ignores high spatial frequencies. This shows that the gains in accuracy are obtained at the expense of the fine spatial structure, which is a well-known phenomenon in forecasting. However, highlighting and quantifying this effect is important because smoothing can be problematic in some applications such as variability or ramp forecasting. Using hybrid loss functions penalizing the lack of variability in the forecasts indeed improves the spatial behavior of the deep-learning model without significantly reducing its pixel-wise performance. At large forecast horizons, however, such hybrid loss functions cannot prevent a substantial loss of spatial variability and convLSTM predictions remain overly smooth.