Learning goals¶
Separate binary labels, margin estimates, and trajectory surrogates as different products
Separate trajectory error from boundary / CCT decision error
Place a learned surrogate inside an operational CCT search carefully
Require classical simulation baselines on the same scenarios
Reading¶
What the model is asked to produce¶
| Approach | Typical output | Operational use |
|---|---|---|
| Binary classifier | Stable / unstable | Fast screen only — no clearing margin |
| Margin regressor | CCT-like scalar | Needs careful labels and honesty about extrapolation |
| Trajectory surrogate | (and related states), then search | Reuse one model for nominal TSA and CCT search if the same criterion is used |
Binary stable/unstable labels are not enough for risk-informed decisions that need how close the system is to the boundary. Margin regression and trajectory-based search aim at that gap; both still require a stated criterion and a reference simulator.
One trajectory model, two assessment tasks¶
A useful teaching pattern (planning-oriented screening):
Time-domain reference (e.g. Lab 1 SMIB / later ANDES)
→ train trajectory surrogate (supervised net or forward PINN)
→ TSA screen at nominal clearing
→ CCT search by bisection on clearing timeThe same trained checkpoint can serve both tasks only if the stability rule (horizon, threshold, persistence window, …) matches the rule used on the reference trajectories. Mixing criteria between surrogate and reference invalidates the comparison.
Forward surrogates for TSA¶
A forward surrogate predicts trajectories quickly for many clearing times or operating points. CCT search may call the surrogate instead of the full simulator — but then CCT error and false-safe rate must be reported against the reference simulator.
| Safe practice | Unsafe practice |
|---|---|
| Same fault, clearing action, horizon, criterion | Mixing criteria between surrogate and reference |
| Boundary-focused test set | Only easy stable cases |
| Signed CCT error + false-safe | Accuracy-only leaderboard |
| Report that better trajectories need not yield protection-grade CCT | Treating low trajectory RMSE as CCT certification |
Improved trajectory fidelity helps screening and comparative studies; it does not automatically deliver protection-grade CCT estimation.
Worked example¶
On SMIB Lab 1 scenarios, compare CCT from dense reference simulation vs CCT from a surrogate with a coarse time grid. Where do rankings disagree? Keep the stability rule identical.
Lab¶
Further reading¶
Optional published SMIB trajectory-surrogate example (Further reading only — not a course lab): Poulose (2026)
Common misconceptions¶
“Fast surrogate implies trustworthy CCT”
“Binary accuracy is enough for operational margins”
Retraining on the CCT test bracket without disclosure
Skipping classical energy-function or TDS baselines when available
- Poulose, A. (n.d.). Unified physics-informed neural surrogate SMIB TSA/CCT benchmark. GitHub repository. Retrieved October 14, 2026, from https://github.com/albert8943/unified-physics-informed-neural-surrogate-smib-tsa-cct-benchmark