Intra-Hour Solar Irradiance Estimation Using Infrared Sky Images and MobileNetV2-Based CNN Regression

In recent years, the rapid growth in both the development and use of solar energy has spurred intensive research in solar power estimation and forecasting in energy system integration. Solar irradiance, the sun's energy incident on Earth, is subject to fluctuations caused by dynamic atmospheric particulates such as clouds, significantly impacting energy pro-duction. Ground-based sky imagers, such as the Total Sky Imager (TSI), have substantially improved forecasting accuracy by cap-turing essential cloud data. Deep learning models, particularly Convolutional Neural Networks (CNNs), have become pivotal in this field. This paper introduces a novel method for intra-hour solar irradiance estimation using infrared sky images and a CNN-regression model. The primary objective of this research is to achieve real-time accuracy, addressing the critical need for precise solar energy forecasting. The proposed method advances the state-of-the-art by reducing the Root Mean Square Error (RMSE) to an impressive 16.18 W/m
²
. In our study, we leveraged a pre-trained and lightweight MobileNetV2 model, underlining its exceptional effectiveness in enhancing our approach for solar irradiance estimation from infrared sky images. This utilization of advanced pre-trained CNN models underscores the potential for accelerating progress in the field of solar energy forecasting.