Short-term GHI Forecasting based on All-Sky Images using a CNN-LSTM Architecture

Accurate short-term forecasting of solar irradiance is essential for reliable grid integration and energy planning. We present a deep learning approach combining convolutional and recurrent neural networks (CNN-LSTM) to predict Global Horizontal Irradiance (GHI) from all-sky fisheye images. Using sequences of recent images, the model captures both spatial cloud features and their temporal dynamics to forecast GHI up to 15 minutes ahead. Evaluation shows an MAE of 66.25 W/m2 and RMSE of 119.55 W/m2, representing a 16% improvement compared to a persistence baseline. This demonstrates the potential of sky-image-driven models for enhancing short-term solar forecasting.