Photovoltaic Scenario Generation via Multimodal Fusion and Local Window Attention
IEEE Access · 2026 · 14 · 82428-82438
DOI: 10.1109/access.2026.3698126 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Zhipeng Kang und 3 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Shanghai Oriental Talents Program.
Abstract (Originalsprache)
The large-scale integration of photovoltaic (PV) power intensifies the uncertainty of electrical grids. Abrupt changes in PV output during diurnal transitions present challenges for high-resolution scenario generation due to deficiencies in multimodal fusion, constraints in long-sequence modeling, and inefficiencies in feature decoupling. This paper proposes a multimodal local window generative adversarial network (M-LGAN) to address these bottlenecks. M-LGAN utilizes PV-specific local window attention to capture critical-period fluctuations while reducing computational complexity. A dynamic feature balancing module integrates spatio-temporal-noise modalities to enhance spatial heterogeneity modeling. Furthermore, adversarial training with multi-scale constraints and composite loss suppresses mode collapse, preserving site-specificity. The performance of M-LGAN was evaluated on 77 PV sites at 15-minute resolution, with the results indicating a reduction in Maximum Mean Discrepancy of 9.91%-13.04% and an enhancement in Continuous Ranked Probability Score of 3.78%-7.29% in comparison to established benchmarks. Furthermore, it attains a Spatial Consistency Score of 0.0261, successfully mitigates rare-event reproduction errors by 38.5%, and enhances training efficiency by a factor of 3.2. Generated scenarios have been shown to accurately replicate gradient dynamics, thereby providing robust support for day-ahead scheduling and risk assessment in high-renewable systems.
Förderung
- National Natural Science Foundation of China
- Shanghai Oriental Talents Program
Zitieren
APA
Kang, Z., Shang, J., Zhu, R., & Zhu, X. (2026). Photovoltaic Scenario Generation via Multimodal Fusion and Local Window Attention. IEEE Access, 14, 82428-82438. https://doi.org/10.1109/access.2026.3698126
Vancouver
Kang Z, Shang J, Zhu R, Zhu X. Photovoltaic Scenario Generation via Multimodal Fusion and Local Window Attention. IEEE Access. 2026;14:82428-82438. doi:10.1109/access.2026.3698126
BibTeX
@article{Kang2026,
author = {Kang, Zhipeng and Shang, Jianbo and Zhu, Rui and Zhu, Xiaojing},
title = {Photovoltaic Scenario Generation via Multimodal Fusion and Local Window Attention},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {82428--82438},
doi = {10.1109/access.2026.3698126}
}
EUVIMED