PV Arc Fault Circuit Interrupter with Knowledge Distillation-Based Lightweight Convolutional Neural Network and SSCB Integration

DC arc faults in photovoltaic (PV) systems pose serious safety risks, including fire hazards and power disruption, making accurate and timely detection essential. This paper presents PArcNet, a lightweight convolutional neural network (CNN) architecture optimized through knowledge distillation for efficient arc fault detection in resource-constrained environments. Trained using a high-capacity teacher model, PArcNet achieves over 98% detection accuracy with only 5.04 K parameters and a memory footprint of 90 kB-making it suitable for real-time deployment on low-cost microcontrollers. The model outperforms traditional machine learning (ML) classifiers such as Support Vector Machines (SVM) and LightGBM in both accuracy and model efficiency, and its robustness has been confirmed through 5-fold cross-validation. For real-time validation, PArcNet was implemented on an STM32H743ZI2 microcontroller and integrated with a solid-state circuit breaker (SSCB), achieving arc detection and interruption in as little as 19 ms. To further reduce false positives during transient disturbances, a 5-cycle confirmation strategy was adopted, resulting in a total response time of 95 ms. Extensive testing under varied PV operating scenarios-including inverter startup, shadow occlusion, and load fluctuations-demonstrates the model's resilience and reliability. PArcNet offers a scalable, accurate, and fast-response solution for embedded PV arc fault protection, contributing to safer and more robust renewable energy systems.