Engineering question: How will the system evolve under specified inputs, parameters, and disturbances?
Planned depth (stubs later)¶
Surrogate rollouts and event interfaces
Hybrid physics–ML dynamic models
Multimachine trajectory tasks (after SMIB mastery)
Labs¶
Lab 1 (reference data), Lab 2 (surrogate vs PINN), Lab 4 (learned dynamics)
Further practice (optional)¶
After Labs 1–2: ANDES PINN surrogate gallery (SMIB + symbolic DAE path) — does not replace course labs
Renewables / frequency projects: low-inertia IEEE 14 gallery
Network reference overview: ANDES as network reference
Model reduction and long-horizon rollouts¶
A learned dynamics model is usually trained on short windows but used for long rollouts. Two ideas help:
Reduced state. Learn the dynamics in a small latent state (an encoder compresses the measured state, a learned map advances it, a decoder returns physical variables). For SMIB the full state is already small; the idea matters for multimachine and converter-rich systems.
Rollout error growth. One-step accuracy does not guarantee multi-step accuracy, because errors feed back into the next prediction. Lab 4 should report error versus horizon, not one number.
Method background: Thuerey et al. (2021), model reduction and time series (fast-forward overview with pointers to external code; check each linked implementation’s own license).
Discussion: what Dynamics methods can and cannot claim¶
| Can claim (with evidence) | Cannot claim |
|---|---|
| Trajectory error on stated test disturbances, horizons, and operating points | Accuracy on disturbances, horizons, or operating points outside the test set |
| Speed-up over the reference simulator for the same task | That a fast surrogate replaces the simulator for new topologies |
| Physics residual size, if a residual is reported | That a low residual means the physics is exactly satisfied |
Rule: every Dynamics result names its reference simulator (Lab 1), its test disturbances, and its horizon.
Status¶
Overview stub only — grow independently after foundations.
- Thuerey, N., Holzschuh, B., Holl, P., Kohl, G., Lino, M., Liu, Q., Schnell, P., & Trost, F. (2021). Physics-based Deep Learning. arXiv. 10.48550/arXiv.2109.05237