Abstract
Transient stability assessment must identify instability and its likely sources despite incomplete and delayed phasor measurement unit (PMU) data. This study develops a physics-guided graph-temporal framework that jointly predicts stability margins, instability types, risk sources, and line fault positions. A physical risk prior derived from differential-algebraic equations aggregates adverse scenarios through conditional value at risk and is fused with measurement-based source probabilities. A graph attention encoder and causal temporal convolution capture spatial and temporal dependencies, with PMU availability and elapsed-time indicators representing measurement degradation. Auxiliary algebraic power balance consistency guides training, while probability calibration and conformal prediction support confidence-based decisions. On the grouped 39-bus test set, rotor-angle and voltage balanced accuracies reached 99.22% and 98.86%, respectively, and Top-1 source localization accuracy reached 96.32%, 1.54 percentage points above the graph-temporal reference. The mean expected calibration error was 0.91%, and nominal 95% source sets achieved 95.40% empirical coverage with 1.42 candidates on average; coverage under 30% PMU loss was 94.62%. Under 30% PMU-channel loss and a six-sample delay, balanced accuracy was 97.31%, compared with 95.08% without measurement quality indicators. After retraining on an independent 68-bus system, rotor-angle balanced accuracy and Top-1 localization accuracy reached 98.37% and 92.68%, respectively. Model-only 95th-percentile latency was 21.8 ms for a single model and 58.9 ms for the five-model ensemble. The evaluated simulations quantify joint prediction accuracy, empirical coverage, and computational cost under topology and measurement variation.