Design of an adaptive identification method for faulty operating states of power system relay protection devices

• Proposes adaptive ResNet50-based method for relay protection fault identification in complex grids. • Integrates residual connections with probabilistic learning for enhanced fault feature extraction. • Leverages same-source data comparison to improve state recognition accuracy across system complexities. • Validates zero-sequence overcurrent adaptation for multi-scale fault pattern characterization. • Achieves 92.7% recognition accuracy with 15% faster convergence than conventional CNN architectures. To address the high complexity and diversity of faults in relay protection devices, as well as the challenges in fault feature extraction that affect fault identification accuracy, an adaptive recognition method is proposed for identifying the faulty operation states of power system relay protection devices. The zero-sequence inverse-time overcurrent protection method is utilized to enable adaptive adjustments within these devices. A same-source data comparison approach is adopted to collect faulty operation state data from power system relay protection devices, which serves as the input for the ResNet50 network. This network, consisting of 50 convolutional layers, processes complex fault features from the input data and captures extensive fault feature information through its powerful feature representation capability. The network layers are interconnected via residual connections, and adaptive probabilistic learning is applied during network training. The fault features extracted by the relay protection device are mapped to a label space through a fully connected layer, and the adaptive recognition results for the faulty operation states are generated using a softmax function. The experimental results demonstrate that the proposed method accurately identifies faulty operation states in relay protection devices and exhibits adaptability to power systems of varying complexities.