基于时空门限融合网络的电力变压器励磁涌流识别OA

The inrush current of power transformer shares similar characteristics with internal fault currents,which tends to cause the misoperation of the main differential protection.The existing methods of inrush current identification suffer from the problems such as insufficient information utilization and low accuracy under short time windows.To address the problems,a gated temporal-spatial fusion network is proposed in this paper.The three-phase differential current of transformers is taken as the monitoring target,and the one-dimensional time-series waveforms is converted into two-dimensional images via the Gramian angular field.A dual-branch structure is adopted to extract the temporal-image multimodal features.The temporal branch utilizes a Transformer encoder to mine the global temporal features and the spatial branch parses the image texture features based on a residual network.Meanwhile,a gated attention fusion mechanism is integrated to achieve the adaptive feature fusion with dynamic gating.The short-time window-sampled simulation data of inrush current and part of the on-site recorded data are utilized to construct the training set for the network training,while another part of the short-time window-sampled on-site recorded data is employed as the test set to verify the network performance.The results show that an inrush current identification accuracy of 95.20%and an F1-score of 95.65 are achieved.It verifies the efficient identification capability.The reliable technical support is provided for the real-time decision-making of differential protection in power transformer.