Deep feature fusion and physical constraints-based photovoltaic system resonant overvoltage detection

To address the challenges of complex characteristics, limited accuracy, and insufficient real-time performance in resonant overvoltage detection for grid-connected photovoltaic systems, this paper proposes a deep detection network that integrates multi-scale feature extraction with circuit physical constraints. By constructing a parallel dual-branch architecture to simultaneously capture time-domain transient characteristics and frequency-domain resonant components of voltage signals, and introducing an adaptive weighting mechanism to achieve multi-scale feature fusion, the sensitivity of high-frequency resonance mode identification is significantly improved. The circuit topology prior knowledge (Kirchhoff's laws, impedance characteristics) is innovatively transformed into physically regularized constraints embedded in the loss function, guiding the network to learn feature representations that conform to power system laws, enhancing model generalization and interpretability. Furthermore, a lightweight dynamic attention module is designed, which strengthens feature weights during critical resonance periods through sparse computation strategies, substantially reducing computational redundancy in non-fault intervals. Tests on real photovoltaic power plant overvoltage datasets demonstrate that this method outperforms mainstream models in both F1-score (96.8%) and false positive rate (1.2%).