Data-Driven Generator Transient Prediction for Digital Twin Decision Support

Overview

The study asks how a digital twin can quickly and reliably predict, before a planned load command is applied to a generator, whether voltage and frequency deviations will stay within acceptable limits. The proposed model is an event-conditioned Hankel dynamic mode decomposition with control (Hankel-DMDc), combining delay-coordinate lifting, command-event memory features and an event-weighted Hankel basis, with conformal prediction bands for uncertainty.

The model was trained and tested on hardware data from a 180 kW/225 kVA, 480 V, 60 Hz generator (Kohler KG180) loaded by a programmable four-quadrant amplifier: random active-power commands of 0-49 kW and reactive commands of 0-18 kVAr (apparent power limited to 50 kVA, power factor 0.8 lagging to unity). Even at steady load, the low-inertia, governor-controlled generator showed low-frequency oscillations of about 3 V and 0.5 Hz.

Results: event-window RMS forecast errors were 1.058 V and 0.155 Hz; for a nominal 90% target the joint voltage-frequency bands achieved 90.17% coverage, with mean widths of 3.55 V and 0.566 Hz. A 50 s open-loop rollout ran in 181 ms on one CPU core, about 280 times faster than real time, with accuracy evaluated up to 5 s horizons. The model uses a 100-dimensional reduced state, 2.6 s of measured history and about 15 MB of memory; offline basis construction takes 1,057 s.

Limits: evaluation used data from the same physical generator (the test set came from a separate experimental run), so transfer to a new operating profile is shown but not generalisation to other generators, controllers, sensor chains or load magnitudes. Open-loop rollout can accumulate error with horizon, so the calibrated band should be used rather than the point forecast alone. The work is not DC-specific, but it addresses the islanded-generator step-loading problem that governs data-centre load-transfer and retransfer sequencing.

Key points

1. The testbed used a Kohler KG180 generator rated 180 kW/225 kVA, 480 V L-L, 60 Hz, with random active-power commands of 0-49 kW and reactive commands of 0-18 kVAr (apparent power limited to 50 kVA, PF 0.8 lagging to unity).
2. Even at steady load, the low-inertia, governor-controlled generator showed low-frequency oscillations of about 3 V and 0.5 Hz.
3. Event-window RMS forecast errors were 1.058 V and 0.155 Hz; for a nominal 90% target the joint bands achieved 90.17% coverage with mean widths of 3.55 V and 0.566 Hz.
4. A 50 s open-loop rollout ran in 181 ms on one CPU core, about 280 times faster than real time; accuracy was evaluated up to 5 s horizons.
5. The model uses a 100-dimensional reduced state and about 15 MB of memory; a 1 s forecast takes 5.28 ms and offline basis construction takes 1,057 s.
6. The authors recommend flagging a planned load command for caution or higher-fidelity validation if the predicted band crosses system voltage or frequency limits.

Using it in design

• In data-centre generator load-transfer and retransfer sequencing, size load steps against generator voltage and frequency limits, and split steps whose prediction band crosses a limit into smaller steps.
• In digital-twin decision support, rely on the uncertainty band rather than the point forecast, and send commands whose band leaves the limits to additional validation.
• Retrain such surrogates for the specific generator, governor and sensor chain; do not assume generalisation to a different generator.

Who should read it

Power-system engineers and researchers developing generator load-transfer sequencing, digital twins and decision-support tools.

Document contents

1. I. Introduction
2. II. Background and Related Work (A. Digital Twins and Generator Surrogate Modeling, B. Koopman-Based Dynamic Surrogates)
3. III. Event-Conditioned Surrogate Model (A. Command-Conditioned Forecasting Map, B. Delay-Coordinate Observable, C. Event-Aware Command Features, D. Event-Weighted Hankel Basis, E. Controlled Reduced-Order Evolution, F. Transient-Aware Model Selection)
4. IV. Hardware Data and Evaluation Protocol (A. Hardware Setup and Measured Generator Profile, B. Normalized Coordinates and Preprocessing, C. Data and Metrics)
5. V. Deterministic Prediction Results
6. VI. Conformal Prediction Bands (A. Calibration Principle, B. Event-Conditioned Joint Bands, C. Held-Out Coverage Results)
7. VII. Discussion and Deployment Scope
8. VIII. Conclusion

Related standards

• ISO 8528-5

Read the document

Publication details

Authors

Emad Sadeghi; Emerson Miller; Kerry Sado; Adel Nasiri

Publisher

arXiv (University of South Carolina)

Year

2026

Document number

arXiv:2609.05356

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.48550/arXiv.2609.05356

Licence

CC BY 4.0 — share with attribution

Licence evidence

arxiv.org

Citation