Abstract:

DC series arc faults (SAFs) in photovoltaic (PV) systems are a significant concern due to potential causes such as loose connectors and damaged cable insulation.These faults pose serious fire risks and can lead to increased operational costs.A lightweight machine learning (ML) framework for real-time SAF detection is proposed,with feature optimization adopted to balance detection accuracy and computational efficiency.A hybrid set of 44 time-frequency features is extracted via a 6-level Daubechies-4 wavelet transform combined with time-domain analysis.Four different optimization strategies are evaluated,with XGBoost feature importance (XGB_FI) showing superior performance by reducing the features to 12 key indicators.The Optuna-optimized XGBoost model achieved 99.94% accuracy during 10-fold cross-validation.When deployed on a Raspberry Pi 4B,it achieved 99.90% accuracy on 4 000 unseen samples,with an inference time of just 3.07 ms per sample,which meets the latency requirments specified in UL 1699B—2018 standard for safety for photovoltaic (PV) DC arc-fault circuit protection latency requirements.Effective identification of SAF can be realized by the proposed method.

Key words: photovoltaic system, DC series arc fault detection, feature optimization, lightweight machine learning, embedded real-time system