Robust Oscillation Source Tracing in Auxiliary Power Systems Based on Physics-Constrained Spatiotemporal Graph Learning

Low-frequency oscillation tracing in auxiliary power systems is affected by load changes, measurement noise, and missing sensor data. To improve localization under these conditions, an Adaptive ST-GCN model with electrical-topology constraints is developed. Steady-state voltage, active-power, and reactive-power signals are transformed into the complex-frequency domain, where both magnitude and phase are retained. The admittance relations calculated from line parameters are used as the basic adjacency matrix. A compensation matrix bounded by the Tanh function allows the edge weights to vary within a limited range, while random masking of node features during training is used to simulate missing measurements. Tests are carried out on a three-node Simulink system. Under the combined condition of 70% light load, 5 dB noise, and loss of the Node 2 measurement, Adaptive ST-GCN achieves 100.00% tracing accuracy on the test samples, compared with 66.67% for the MLP baseline. Ablation and single-disturbance tests show that electrical topology, phase features, and node masking all affect the tracing result. Within the operating conditions considered in this study, the model maintains relatively stable oscillation-source identification performance.