EUVIMEDEUVIMED entdecken
Crossref

Lightweight Edge-Deployable Voltage Prediction Method for Distribution Networks With High Photovoltaic Penetration

Zhijin Lyu, Wei Liu, Wei Du, Tian Gao, Xibin Shan, Kaiyuan Hu

IEEE Transactions on Industrial Informatics · 2026-07 · 22 (7) · 6022-6033

DOI: 10.1109/tii.2026.3674492 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Industrial Informatics, 2026-07. Verfasst von Zhijin Lyu und 5 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Gefördert von: State Grid Corporation of China Research Project.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Under the continued growth of high penetration distributed photovoltaics (PV), PV interface voltages exhibit pronounced randomness, rapid fluctuations, and even overvoltage, making short-term voltage prediction increasingly critical yet more challenging under dynamic operating conditions. This growing reliance on accurate and timely prediction directly exposes the shortcomings of existing cloud-based forecasting schemes, which are constrained by communication delays, bandwidth limitations, and unstable connectivity, ultimately restricting their ability to provide the fast response required by a prediction-control loop. To address these issues, this study proposes a lightweight edge-deployable voltage prediction (LEDVP) method that consist of an ensemble algorithm and a distributed framework. The ensemble algorithm combines variational mode decomposition for voltage-signal decomposition and extreme gradient boosting for extracting the most informative features, which are then fed into a Mamba state-space sequence model to perform efficient linear-time temporal inference, forming a compact and lightweight prediction process suitable for embedded edge platforms. Moreover, by quantizing model parameters, pruning redundant structures, reusing intermediate caches, and optimizing the model specifically, the LEDVP can be deployed in the distributed framework. Experimental results show that the LEDVP method achieves near-lossless accuracy with a 1–3 s latency on NVIDIA Orin and yields approximately 40–50% lower root mean square error compared with representative decomposition-based neural network methods, while the LEDVP framework maintains stable resource usage. These results demonstrate the engineering feasibility and practical value of LEDVP for real-time voltage prediction, PV interface voltage regulation, and edge-side autonomy in PV-rich distribution networks.

Förderung

  • State Grid Corporation of China Research Project

Zitieren

APA

Lyu, Z., Liu, W., Du, W., Gao, T., Shan, X., & Hu, K. (2026). Lightweight Edge-Deployable Voltage Prediction Method for Distribution Networks With High Photovoltaic Penetration. IEEE Transactions on Industrial Informatics, 22(7), 6022-6033. https://doi.org/10.1109/tii.2026.3674492

Vancouver

Lyu Z, Liu W, Du W, Gao T, Shan X, Hu K. Lightweight Edge-Deployable Voltage Prediction Method for Distribution Networks With High Photovoltaic Penetration. IEEE Transactions on Industrial Informatics. 2026;22(7):6022-6033. doi:10.1109/tii.2026.3674492

BibTeX
@article{Lyu2026,
  author = {Lyu, Zhijin and Liu, Wei and Du, Wei and Gao, Tian and Shan, Xibin and Hu, Kaiyuan},
  title = {Lightweight Edge-Deployable Voltage Prediction Method for Distribution Networks With High Photovoltaic Penetration},
  journal = {IEEE Transactions on Industrial Informatics},
  year = {2026},
  volume = {22},
  number = {7},
  pages = {6022--6033},
  doi = {10.1109/tii.2026.3674492}
}
Kostenlos starten →