计量时序驱动的台区变压器异常监测方法OA

The health states of distribution transformers are critical to the safe operation and regulation efficiency of dis-tribution networks.To address the challenge that the traditional methods struggle to map the transformers'health states due to data quality issues and limited characterization dimensions,a metering time-series driven anomaly monitoring method for distribution transformers is proposed in this paper.First,a multi-dimensional evaluation index system is con-structed based on the metering data,and the health levels are classified accordingly.Second,an adaptive sliding window is introduced to achieve the optimal sample segmentation,and the feature is extracted in depth with the combination of Bayesian optimization algorithm and temporal convolutional network-bidirectional long short-term memory network.In addition,a health baseline is also constructed.Finally,a dual-threshold strategy is employed to implement joint deter-mination of indicator residuals between the measured data and the health baseline,significantly improving the accuracy of anomaly detection.The validation results based on actual distribution substations demonstrate that the proposed meth-od can effectively identify the trend anomalies caused by health degradation in metering data streams,achieving an im-provement of 8.9%in the detection accuracy compared with the traditional threshold criterion method.This method en-ables rapid assessment of transformer conditions and provides a technically feasible solution with high engineering adaptability for anomaly monitoring of equipment in smart distribution networks.