3D-CNN-Based Sky Image Feature Extraction for Short-Term Global Horizontal Irradiance Forecasting

Abstract

The instability and variability of solar irradiance induces great challenges for the management of photovoltaic water pumping systems. Accurate global horizontal irradiance (GHI) forecasting is a promising technique to solve this problem. To improve short-term GHI forecasting accuracy, ground-based sky image is valuable due to its correlation with solar generation. In previous studies, great efforts have been made to extract numerical features from sky image for data-driven solar irradiance forecasting methods, e.g., based on pixel-value color information, and based on the cloud motion detection method. In this work, we propose a novel feature extracting method for GHI forecasting that a three-dimensional (3D) convolutional neural network (CNN) is developed to extract features from sky images with efficient training strategies. Popular machine learning algorithms are introduced as GHI forecasting models and corresponding forecasting accuracy is fully explored with different input features on a large dataset. The numerical experiment illustrates that the minimum average root mean square error (RMSE) of 62 W/m2 is achieved by the proposed method with 15.2% improvement in Skill score against baseline forecasting method.