A controller needs a timestamped thermal state, which may be partly inferred. Keep modelled winding temperature distinct from a measured oil temperature; Jain’s EV study uses a particular transformer thermal model, not universal asset parameters. [1]
Keep three records: inputs, parameters and state
Inputs include current history, ambient temperature and known cooling status. Parameters describe heat generation, thermal storage and heat removal for the particular transformer. State describes the estimated condition at a specific time, such as oil temperature and winding temperature rise. Changing an input should advance the state; it should not silently redefine the transformer’s parameters.
A conceptual observer first predicts the next state from the previous estimate and current inputs, then compares a corresponding predicted measurement with the actual measurement. That residual may correct the state estimate. The correction must respect what the sensor can observe: a top-oil reading does not directly observe every winding hot spot.
| Residual pattern | Possible explanation | Useful discriminating check |
|---|---|---|
| Persistent offset at several loads | Sensor bias or an unsuitable parameter | Compare an independent temperature reference and asset data |
| Mismatch concentrated during load changes | Timing, dynamics or input alignment | Check timestamp consistency and the load-step history |
| Mismatch after cooling changes | Wrong cooling-state input | Compare actual fan/pump status with the recorded command |
| A single abrupt departure | Data fault or real event | Inspect quality flags and independent electrical observations |
A small residual is not full validation
A model can match the measured oil temperature while estimating the winding rise poorly. Evaluate the state quantities relevant to the control limit and the conditions used for that evaluation. If a required state lacks direct validation, retain an uncertainty allowance and identify the evidence that would reduce it.
A small residual does not establish a correct unobserved state. NIST’s model-validation guidance calls for examining residual structure and evidence beyond a fitted numerical measure [2]. Reserve independent histories and cooling modes for evaluation.
Hand the controller a range with a timestamp
The control interface should contain the estimate, its time, the quality of the inputs and an upper bound or other defined uncertainty representation. A central value without those fields leaves the scheduler unable to distinguish a well-observed cool transformer from an old estimate that happens to be low.
For an independent one-state observer, assume time constant 120 min, ambient 30 °C, initial rise 50 K and new equilibrium rise 70 K. The lumped energy model [3] predicts 70+(50−70)e⁻⁰·²⁵ = 54.424 K rise after 30 min, or 84.424 °C. With assumed matching-state sensor reading 86.0 °C and correction gain 0.40, residual is 1.576 K and corrected state 85.054 °C. A memoryless equilibrium value would be 100 °C. Advance the state rather than resetting it, then validate dynamics and uncertainty on independent data. This single-state correction is not a validated hot-spot observer from top-oil alone.
Evaluate the intended decision with the uncertain state, not only the best-fit trajectory. Record whether the next charging allowance changes materially across plausible estimates. If it does, the limitation is an observation problem as much as a capacity problem. That result helps prioritize a better sensor, a corrected time base or a more suitable thermal model.
References
- A. Jain and M. Karimi-Ghartemani, Mitigating Adverse Impacts of Increased Electric Vehicle Charging on Distribution Transformers, Energies 15, 9023, 2022. DOI: 10.3390/en15239023.
- NIST/SEMATECH, e-Handbook of Statistical Methods, Model Validation, section 4.4.4, and Estimation, section 4.1.3.1.
- U.S. Department of Energy, DOE-HDBK-1012/2-92, Thermodynamics, Heat Transfer, and Fluid Flow, Volume 2, 1992.