EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Dual-Layer Physics-Residual-Constraint-Driven SAE-Mini-Transformer: DC-Side Fault Diagnosis of Photovoltaic Arrays for Edge Terminals

Ying Su, Shaorong Jia, Jiajie Li, Honglu Zhu

Energies · 14.11.2026 · 19 (22) · 5499

DOI: 10.3390/en19225499 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Energies, online erschienen am 14.11.2026. Verfasst von Ying Su und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: China Three Gorges Corporation.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Direct current-side (DC-side) faults, including string open circuits, string short circuits, string underperformance, sensor faults, and partial shading, threaten the safe operation of photovoltaic (PV) plants. Existing data-driven methods lack physical priors, struggle to extract coupled features from multivariate long time series, or require excessive computational resources for edge deployment. Overlapping fault features further limit fault discrimination and cross-site robustness. To address the aforementioned issues, this study aims to develop a DC-side fault diagnosis method that integrates physical consistency with PV operation, multi-class fault identification under complex operating conditions, and computational efficiency for edge deployment. The objective is to improve fault diagnosis performance and cross-site generalization while meeting the real-time diagnostic requirements of edge devices in PV power plants. This study proposes a lightweight stacked autoencoder (SAE)-Mini-Transformer fault diagnosis method driven by dual-layer physics-residual constraints. Three physical residuals, the electrical-consistency residual, power-ramp residual, and performance ratio (PR) out-of-range residual, embed physical constraints through sample weighting and joint-loss penalty terms. These constraints apply only during training, adding no computational overhead during inference. The SAE denoises and reduces the dimensionality of multi-source time-series data, while a Mini-Transformer with a streamlined encoder captures long-range temporal coupling between electrical and meteorological variables. A feature discrimination head outputs posterior probabilities for the normal state and five typical faults. Validation using measured operating data from three PV plants in China shows that the macro-averaged F1-score (Macro-F1) exceeds 93% at each plant, indicating good cross-site robustness. The model outperforms one-dimensional convolutional neural network (1D-CNN) and long short-term memory (LSTM) models and significantly reduces model complexity while maintaining recognition performance close to that of a standard Transformer. Single-sample inference takes approximately 0.8–1.0 ms, meeting real-time diagnosis requirements. Integrating PV operating physics with lightweight temporal modeling, the method provides theoretical support and an engineering reference for intelligent DC-side operation and maintenance.

Förderung

  • China Three Gorges Corporation

Zitieren

APA

Su, Y., Jia, S., Li, J., & Zhu, H. (2026). Dual-Layer Physics-Residual-Constraint-Driven SAE-Mini-Transformer: DC-Side Fault Diagnosis of Photovoltaic Arrays for Edge Terminals. Energies, 19(22), 5499. https://doi.org/10.3390/en19225499

Vancouver

Su Y, Jia S, Li J, Zhu H. Dual-Layer Physics-Residual-Constraint-Driven SAE-Mini-Transformer: DC-Side Fault Diagnosis of Photovoltaic Arrays for Edge Terminals. Energies. 2026;19(22):5499. doi:10.3390/en19225499

BibTeX
@article{Su2026,
  author = {Su, Ying and Jia, Shaorong and Li, Jiajie and Zhu, Honglu},
  title = {Dual-Layer Physics-Residual-Constraint-Driven SAE-Mini-Transformer: DC-Side Fault Diagnosis of Photovoltaic Arrays for Edge Terminals},
  journal = {Energies},
  year = {2026},
  volume = {19},
  number = {22},
  pages = {5499},
  doi = {10.3390/en19225499}
}
Kostenlos starten →