Abnormal data detection is of great significance in ensuring the safe and stable operation of power systems. To address the limited state-coupling modeling capabilities and the insufficient dynamic adaptability of topological structures in existing methods

a power system abnormal data detection method based on high-order relations and topology is proposed. Firstly

an adaptive state-aware graph convolution module is constructed to improve robustness to abnormal disturbances by jointly regulating the direction of state propagation and the response intensity. Subsequently

a cross-graph embedding alignment module is designed to associate the topological structure information of the learned state graph with the node semantic space

thereby achieving consistent topological semantic alignment across multiple state graphs. Finally

a residual decomposition scoring module is proposed to deconstruct the aligned embedded residual into two orthogonal subspaces of structure and state

and anomaly detection is realized based on the weighted residual scoring function. Experiments conducted in typical power system scenarios show that the proposed model outperforms existing methods in both anomaly detection accuracy and robustness.