A Diffusion-Based Probabilistic Ultra-Short-Term Solar Power Prediction Using the Sky Image Sequences
<p dir="ltr">The inherently unpredictable nature of solar power generation, primarily due to rapidly changing cloud cover, poses a significant challenge to the operation of solar-integrated energy systems. Accurate ultra-short-term PhotoVoltaic (PV) power forecasting is crucial for t...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| مؤلفون آخرون: | , , |
| منشور في: |
2025
|
| الموضوعات: | |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | <p dir="ltr">The inherently unpredictable nature of solar power generation, primarily due to rapidly changing cloud cover, poses a significant challenge to the operation of solar-integrated energy systems. Accurate ultra-short-term PhotoVoltaic (PV) power forecasting is crucial for the efficient and reliable operation of these energy systems. Recently, cloud images, providing more direct and comprehensive information about cloud patterns than on-site or off-site numerical weather prediction data, have become increasingly accessible for analyzing cloud dynamics, enabling more precise and efficient cloud change predictions. To enhance PV power probabilistic forecasting, this paper proposes a novel diffusion-based framework that leverages cross-attention on multi-resolution sky image patches. Central to this framework is the Cross Branch Visual Informer (CB-ViInf), the first model to integrate multi-resolution patches into the Vision Informer architecture, improving feature extraction and temporal modeling for solar forecasting. Furthermore, we propose an innovative Temporal Encoder with a dual-block architecture to refine temporal dependencies. The first block leverages Spatiotemporal Ridgelet Transform (STRT) and Multi-Frame Adaptive Singular Value Decomposition (MF-ASVD) as a dynamic feature refinement process, reducing noise and computational complexity while preserving critical features. The second block, a Spatio-Temporal Attention mechanism, simultaneously incorporates local and global attentions to capture fine-grained short-term variations and broader trends, enhancing adaptability to rapid cloud movements and irradiance fluctuations. Additionally, we propose a novel loss function based on Variational Inference and the Evidence Lower Bound (ELBO) to improve uncertainty quantification and prediction accuracy. Extensive evaluations on a real-world dataset against 64 baseline models using four evaluation metrics demonstrate the superiority of the proposed framework, establishing it as a state-of-the-art approach in ultra-short-term PV power forecasting. The proposed model offers an accurate, robust, and uncertainty-aware solar power forecasting methodology for improved power system operation and management.</p><h2 dir="ltr">Other Information</h2><p dir="ltr">Published in: IEEE Access<br>License: <a href="https://creativecommons.org/licenses/by/4.0/deed.en" target="_blank">https://creativecommons.org/licenses/by/4.0/</a><br>See article on publisher's website: <a href="https://dx.doi.org/10.1109/access.2025.3600713" target="_blank">https://dx.doi.org/10.1109/access.2025.3600713</a></p> |
|---|