Research on Photovoltaic Power Forecasting Based on Ground-Based Cloud Images

Abstract

: To address the challenges of photovoltaic (PV) power forecasting, this study proposes an improved forecasting model that aims to enhance prediction accuracy and cope with the uncertainty of the climate system. The core of the model consists of three parts: (1) A Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) network is employed to predict ground-based cloud images, demonstrating superior detail prediction capability over Generative Adversarial Networks (GAN). (2) An algorithm is designed to track the sun’s position, partition the cloud image region, and construct quantitative indicators, which are verified through statistical tests to be highly correlated with direct solar radiation. (3) Top-of-atmosphere irradiance is predicted by incorporating meteorological parameters, and a Classification and Regression Tree (CART) regression tree model is subsequently used for power forecasting, achieving a Mean Absolute Error (MAE) of 4.43 and a Root Mean Square Error (RMSE) of 6.08.