A Preprocessing and Feature Extraction Method of Ground-based Cloud Images for Photovoltaic Power Prediction

Abstract

The photovoltaic system is affected by many factors, so its ultra-short-term prediction faces enormous challenges. Total Sky Imager (TSI) is used for sky monitoring, which is of great significance for photovoltaic power prediction, but its structural characteristics lead to distortion and loss of information. A method of cloud image preprocessing and feature extraction is proposed in this paper. Firstly, the template calibration method is proposed to establish the distortion model, and bilinear interpolation is used for gray reconstruction. Secondly, the method of combining Multi-Scale Retinex (MSR) method with grayscale transformation is proposed to compensate the illumination of cloud image. Then, digital image processing technology is used to extract spectral features, texture features, cloud features, shadow-band features, and other features that affect the photovoltaic power. Finally, the Gradient Boosting Decision Tree (GBDT) method is used to train and build the photovoltaic power regression model according to ground-based cloud images and photovoltaic power. Experiments show that the proposed preprocessing method of cloud image can restore the real information of the sky, and the photovoltaic power regression model constructed from the image features has high accuracy, and the R-Squared(R2) can reach 0.96. It provides an effective method for the ultra-short-term prediction of photovoltaic power.