The active distribution network with its strong cyber-physical coupling

is vulnerable to unconventional anomalies like uncertain internal resource inputs/outputs and external network attacks

challenging situational awareness and safe operation. A new framework based on embedded physical knowledge is proposed for detecting and explaining anomalies in system situational signals. First

four typical anomalies are selected

and their physical models are defined from a system operation perspective. A spatio-temporal graph convolutional neural network is then employed to detect these anomalies

capture complex dependencies between nodes

and efficiently aggregate features by embedding physical information. Next

a three-layer fully connected neural network classifies and identifies multiple signals using a multi-label classification mechanism. Finally

to analyze the mechanisms behind the anomalies

causal relationships between nodes are established to quantify their impact on the anomaly signals. Numerical simulations and comparisons demonstrate the method’s accuracy and robustness

which can provide a valuable reference for future situational warnings and defense strategies.