Abstract: In response to the strong concealment of abnormal patterns in power metering data and the poor accuracy of existing recognition methods, a research on anomaly recognition model based on Boosting Ensemble Learning is carried out. By constructing an adaptive weighted feature selection mechanism and a progressive residual learning framework, the discriminative boundaries of multiple weak classifiers are deeply integrated to sensitively capture and accurately separate weak abnormal signals in non-stationary and high noise metric data. This model adopts a serialization training strategy, focusing on optimizing the iterative attention to minority class abnormal samples, effectively overcoming classification bias caused by data imbalance. Comparative experiments show that the proposed method is significantly superior to existing methods in terms of recognition accuracy indicators, and has strong recognition ability and generalization stability for various hidden abnormal forms such as distortion, drift, pulse, and constant value. It provides a high-order application value solution for the intelligent operation and data quality governance of power system metering devices.