ПРЕПРИНТ
О результатах, изложенных в препринтах, не следует сообщать в СМИ как о проверенной информации.
Intermittent solar energy must be correctly predicted on short-term time scales to maintain the stability and efficiency of the power grid it interacts with. This work presents a deep learning framework that is effective in predicting solar irradiance, a direct proxy for energy generation from the input data, which consists of time-ordered sky images. We establish an end-to-end pipeline to process and synchronize high-resolution sky image data from a fisheye camera with colocated direct normal irradiance measurements from pyranometers. The essential part of our model is a Convolutional Neural Network (CNN) that extracts important spatiotemporal identifiers from the images and translates them into predicted values for future irradiance. Our model shows a significant predictive power trained and assessed on a real-world dataset, specifically showing a Sunny Days Mean Absolute Error (MAE) of 0.353, a Root Mean Squared Error (RMSE) of 0.463, a Cloudy Days Mean Absolute Error (MAE) of 0.964, and a Root Mean Squared Error (RMSE) of 2.029, noticed within the held-out test set. The amount of variance in irradiance explained by the model based on visual sky conditions also represents a large jump from baseline, suggesting that this model captures almost 50% of variability in irradiance. Model stability and generalizability are confirmed by the evaluation of training progress. The proposed work lays the groundwork for a highly effective and reproducible deep learning framework for short-term solar forecasting, which serves as a robust paradigm suitable for future incorporation into smart grid management systems to improve solar power generation reliability.
Rahman Sh., Akber M. A., Islam M. A., Miah Sh., Islam M. S. 2026. Proposing A Robust Deep Learning Framework for Short-Term Solar Power Forecasting Using Sky Images. PREPRINTS.RU. https://doi.org/10.24108/preprints-3115069
