Efficient and Lightweight Convolutional Neural Network-based Series DC Arc Fault Protection for PV Systems

Series DC arc faults in photovoltaic (PV) systems pose substantial fire risks and compromise electrical safety and reliability. Traditional AI-based arc fault detection methods, while highly accurate, are computationally burdensome to integrate into resource-constrained, low-cost microcontrollers for real-time applications. This research introduces an innovative AI algorithm that detects arc faults with high accuracy and employs a MOSFET-based solid-state circuit breaker (SSCB) for rapid circuit interruption. The model utilizes an efficient and lightweight convolutional neural network (CNN) architecture to analyze the frequency spectrum of the load current, featuring a compact design with just 14.6 k trainable parameters. Following the UL1699B standard, a comprehensive arc fault database is developed to enhance the algorithm's robustness under diverse loading conditions. The proposed algorithm detects arc faults with over 98.15% accuracy. TensorFlow-Lite was employed to reduce the model's binary size, facilitating efficient implementation on microcontrollers. The algorithm interrupts arc faults within 25 ms when implemented on a commercial STM32H743ZI2 microcontroller. This timeframe includes 12.8 ms for data acquisition, preprocessing with Fast Fourier Transformation (FFT), and inference time during real-time operations. To the best of our knowledge, this is the fastest AI-based arc fault circuit interrupter (AFCI). This method ensures rapid detection and interruption of arc faults, significantly enhancing the safety and reliability of PV systems.