On Sky Imaging Analysis and Deep Learning for Photovoltaic Output Nowcasting

Abstract

In nowcasting of photovoltaic (PV) output, all-sky images can be utilized as an exogenous data input to improve the accuracy of the prediction. In this paper, we first investigate the correlations between sky images and power outputs of three different types of PV panel installations: fixed array, single-axis tracking array, and dual-axis tracking array. This has not been sufficiently addressed in the literature. Based on the correlation analysis, we conduct an image processing-based PV output nowcasting. Instead of directly using the original sky images as input of convolutional neural network (CNN) to learn features, we propose a pre-processing step to extract the statistical features embedded in the sky images. Then PV output prediction is implemented by a recurrent neural network (RNN)-based model. The experiments results show that the proposed light-weighted deep learning model can attain a promising forecast accuracy.