A Lightweight, Transferable, and Self-Adaptive Framework for Intelligent DC Arc-Fault Detection in Photovoltaic Systems
Summary: arXiv:2603.25749v1 Announce Type: cross
Introduction
Arc-fault circuit interrupters (AFCIs) play a crucial role in enhancing safety by mitigating fire hazards in residential photovoltaic (PV) systems. However, achieving reliable DC arc-fault detection under real-world operating conditions has proven to be a significant challenge. Conventional AFCI solutions often struggle due to various factors such as spectral interference from inverter switching, hardware heterogeneity, operating-condition drift, and environmental noise.
Proposed Solution: The LD-Framework
This paper introduces a novel framework known as the lightweight, transferable, and self-adaptive learning-driven framework (LD-framework) for intelligent DC arc-fault detection. The LD-framework is designed to address the limitations of traditional AFCIs by incorporating three innovative components:
- LD-Spec: This component learns compact spectral representations that allow for efficient on-device inference and achieve near-perfect arc discrimination.
- LD-Align: This mechanism performs cross-hardware representation alignment, ensuring robust detection across heterogeneous inverter platforms despite hardware-induced distribution shifts.
- LD-Adapt: This component introduces a cloud-edge collaborative self-adaptive updating mechanism that detects previously unseen operating regimes and facilitates controlled model evolution.
Experimental Results
To validate the effectiveness of the LD-framework, extensive experiments were conducted involving over 53,000 labeled samples. The results showcased remarkable performance metrics:
- Achieved an accuracy of 0.9999.
- Obtained an F1-score of 0.9996.
- Maintained a 0% false-trip rate across diverse nuisance-trip-prone conditions, such as inverter start-up, grid transitions, load switching, and harmonic disturbances.
Cross-Hardware Transfer and Field Adaptation
The framework also demonstrated reliable adaptation across different hardware platforms, requiring only 0.5%-1% of labeled target data while preserving performance from the source. Furthermore, field adaptation experiments revealed a remarkable recovery of detection precision from 21% to 95% under previously unseen conditions.
Conclusion
The findings from this study indicate that the LD-framework provides a scalable, deployment-oriented AFCI solution that maintains highly reliable detection across heterogeneous devices and supports long-term operation. This innovative approach not only enhances safety in residential PV systems but also paves the way for future advancements in intelligent fault detection technologies.
As the demand for renewable energy sources continues to grow, the implementation of such intelligent systems will be crucial in ensuring both safety and efficiency in photovoltaic applications.