Enhanced Short-Term GHI Prediction using Cloud-Aware Deep Autoencoder Network

Accurate short-term prediction of Global Horizontal Irradiance (GHI) is vital for reliable solar energy integration. This paper proposes a site-specific, data-driven forecasting framework combining clear-sky modeling with deep learning to capture cloud-induced variability. Visible-band satellite imagery is used to generate spatiotemporal cloud mask cubes, which are encoded via a Convolutional LSTM Autoencoder to learn cloud evolution patterns. These latent features, combined with clear-sky GHI estimates and temporal inputs, support multi-step forecasting. Cloud indices derived from normalized cloud masks help reconstruct ground truth GHI. Using a sliding-window approach, the model predicts four future steps from six past observations at 30-minute intervals. The proposed model is trained using a sliding-window approach, the model predicts four future steps from six past inputs. It achieves an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.67, indicating good accuracy and the ability to learn meaningful patterns, making it suitable for operational solar nowcasting. The integration of cloud dynamics, radiative physics, and sequence learning yields accurate and interpretable GHI forecasts, suitable for operational nowcasting in data-limited settings. The resultant framework shows high potential for operational use in nowcasting systems, especially in data-scarce or satellite-only settings.