Deep Learning Multi-Horizon Irradiance Nowcasting: A Comparative Evaluation of Three Methods for Leveraging Sky Images

A team of researchers led by Erling W. Eriksen has published a comparative study evaluating three distinct deep learning (DL) approaches for short-term solar power forecasting, known as irradiance nowcasting. The core challenge is predicting global horizontal irradiance—the total solar radiation received—up to 15 minutes ahead using sequences of all-sky imager (ASI) pictures. The three methods tested were: using a Convolutional Neural Network (CNN) on raw RGB images; feeding a CNN with pre-engineered 2D feature maps based on domain knowledge like cloud motion and solar position; and aggregating those engineered features into a time-series format for the model.

The models were trained on a high-frequency, 29-day dataset and evaluated on seven selected days using root mean squared error and skill score metrics. Surprisingly, the most effective approach was not the raw image CNN, but the method that aggregated the engineered 2D feature maps into a time-series input. This technique yielded superior forecasting performance, demonstrating that integrating complex sky images into DL models for practical nowcasting can be achieved without resorting to intricate, spatially-focused neural network architectures.

This finding is significant for the renewable energy sector. It underscores that alternative image processing and feature engineering methods, informed by physical domain knowledge, can rival or surpass pure end-to-end deep learning for this specific task. The research, detailed in the arXiv preprint 2603.26704, points toward more efficient and potentially more interpretable pathways for building robust AI systems that manage solar grid integration by providing crucial, ultra-short-term power predictions.