Convolutional Neural Network for Short-term Solar Panel Output Prediction

Abstract

The volatility of cloud movement introduced a large amount of uncertainty in short-term solar power prediction, which complicates modern power grid's operation. This work employs a specialized CNN model SUNSET, that utilizes both sky images and solar panel output history as input to predict 15-minute ahead solar panel generation. On a full year database, the model achieves 26.2% forecast skill on the sunny test set, and 16.1% forecast skill on the cloudy dataset. Both sky images and PV output history are shown to be pivotal model input, and two minutes is shown to be a suitable sampling frequency for this application.