A voltage sag detection framework based on constraint-driven neural networks

Voltage sags are critical power quality (PQ) disturbances that threaten the reliability of modern power systems. The increasing penetration of inverter-based resources and nonlinear loads gives rise to complex sag events superimposed with harmonics and transients, which pose a challenge to the effectiveness of traditional signal analysis methods for standard compliance monitoring. Although advanced signal processing methods are capable of addressing the dynamic and complex nature of voltage sags, these methods have shortcomings in automatically learning sag patterns and ensuring the physical consistency of detection results. To address this limitation, a novel voltage sag detection framework that integrates data-driven techniques with signal analysis is proposed, which is named the Voltage Sag-Constrained Neural Network (VoltSagConNN). To achieve accurate and physically consistent detection of voltage sags, the proposed VoltSagConNN incorporates two key processes: voltage sag estimation and the incorporation of physical constraints. The voltage sag estimation combines discrete wavelet transform (DWT) with attention-based neural networks (ABNN). In the physical constraint process, a voltage sag physical constraint loss function is proposed to ensure that the physical laws governing voltage sag disturbance are considered, thereby enhancing the reliability and physical consistency of the model. To reduce the training burden while efficiently extracting voltage sag characteristics, a voltage sag feature vector is constructed. The effectiveness of the VoltSagConNN is validated through extensive case studies using both synthetic and real-world voltage sag disturbance signals.