When can we trust a temperature model?
How hot a motor runs decides how much copper or cooling we can take out of it. A temperature model could guide that choice if we knew when its predictions were good enough. We tested three against published measurements.
In this notebook
The three models carry different amounts of physics. The physical network treats the rotor, winding, tooth and yoke as four masses exchanging heat. The causal regression is statistical and predicts temperature from the operating inputs and their recent history. The hybrid adds a learned correction to the network. All were scored on measurements from one Paderborn University prototype, a different motor from entry 001.
Estimate model parameters.
Choose settings and bands.
Score frozen predictions.
We supplied the measured initial temperatures, then predicted each session without further temperature feedback.
Across the held-out sessions, the hybrid had the lowest rotor error. In session 36 it did worse than the physical network it was built to correct. The figure opens there, and all 14 sessions are available.
Compare the prediction with the measurement.
Session 36, Hybrid band and error
We start with session 36, where the hybrid increased rotor error from 3.94 to 6.38 °C.
Time in minutes. Positive error means the model predicts too cool. The dashed 10 °C line is a research criterion, not a motor safety limit.
Hybrid band at this sample, 15.70 to 27.10 °C
Hybrid, selected session
All three predictions stay visible. The selected band comes from development errors; it does not guarantee coverage or describe sensor uncertainty. Displayed samples preserve extrema; metrics use all 22,608 scored samples, excluding the measured initial state. Session 36 data (CSV).
All 14 held-out sessions
Rotor root-mean-square error (RMSE), °C. Select a session to inspect it above. † The hybrid did worse than the physical network here.
| Session | Physical | Causal | Hybrid |
|---|---|---|---|
| 7.07 | 2.69 | 1.96 | |
| 9.73 | 4.32 | 4.46 | |
| 2.99 | 0.56 | 0.94 | |
| 5.38 | 1.23 | 1.49 | |
| 8.82 | 4.12 | 3.70 | |
| 11.69 | 1.15 | 1.98 | |
| 10.58 | 1.24 | 2.20 | |
| 11.67 | 7.60 | 3.11 | |
| 9.72 | 6.69 | 4.05 | |
| 3.94 | 3.72 | 6.38 | |
| 3.89 | 1.78 | 1.37 | |
| 5.53 | 2.37 | 1.92 | |
| 9.85 | 4.18 | 2.44 | |
| 6.22 | 3.52 | 2.13 |
Lower overall error didn’t settle the choice.
Our comparison rule required at least 10% lower overall error than the physical network, with no held-out session made worse. Causal regression met it. The hybrid failed it.1
Rotor RMSE across all held-out sessions
All 7 fitted candidates, including these three, failed the full thermal screen. Underprediction matters most, because the motor is then hotter than the model says. Each candidate underpredicted by more than 10 °C somewhere; the hybrid’s worst miss was 13.91 °C and causal regression’s was 21.11 °C. That screen was a research criterion, not a motor safety limit.
We’re keeping causal regression for further prediction work within this motor’s measured operating range. These results do not validate a changed magnet, geometry, cooling system or motor.
Too much heat, or too little heat transfer?
A rotor can run hot because the motor generates more heat or because it sheds heat poorly. The fitted network cannot separate the two. It identifies normalized parameters, not absolute losses and thermal conductances. That distinction matters when changing a design. At the same current, removing copper raises winding loss; reducing cooling can make it harder to shed that heat.
We’re continuing to investigate which measurements can distinguish heat generation from heat transfer well enough to guide those changes. Could a thermal transient, with a measured initial state and an informative operating cycle, separate the two? If you have relevant motor data or a different explanation, share it with us.
How we fitted and checked the models.
Three models and their development bands
Physical network
Four temperature states represent the rotor, winding, tooth and yoke. Positive reciprocal conductances connect them to each other and to the coolant and ambient boundaries. Fitted heat-source proxies use current, speed and voltage. The dynamics identify normalized parameters, not separate absolute losses, capacities and conductances.
Causal regression
The present operating inputs and their filtered memories at 30, 120, 600 and 1,800 seconds predict temperature. An initial-state correction uses the measured starting temperatures. Future temperature targets never enter the prediction.
Hybrid
A learned residual correction is added to the physical network and starts at zero. We selected it on the 11 selection sessions before seeing held-out results.
Development bands
For each model and target, the half-width is the largest of the selection sessions’ 95th-percentile absolute residuals. Width stays fixed through a held-out trace. The fraction of observations inside a band is measured afterward; it is not guaranteed coverage, sensor uncertainty or manufacturing reliability.
All seven candidates and the frozen criteria
The full screen requires rotor RMSE ≤ 5 °C, absolute bias ≤ 2 °C, and no session with underprediction above 10 °C. The comparison rule requires at least 10% lower RMSE than the physical network without making any session worse. These are separate tests.
Scroll horizontally to see both test results.
| Candidate | RMSE °C | Bias °C | Worst underprediction °C | Comparative utility | Full screen |
|---|---|---|---|---|---|
| Initial-temperature persistence | 38.162 | −32.473 | 87.40 | Failed | Failed |
| Coolant plus initial offset | 33.579 | −23.867 | 86.88 | Failed | Failed |
| Static regression | 12.916 | −0.581 | 61.33 | Failed | Failed |
| Causal regression | 3.805 | −1.177 | 21.11 | Met | Failed |
| Physical network v1 | 8.211 | −3.572 | 23.99 | Failed | Failed |
| Physical network | 8.180 | −3.621 | 24.11 | Reference | Failed |
| Hybrid | 3.072 | −0.208 | 13.91 | Failed | Failed |
RMSE uses equal session weights; bias is the mean of session biases. Bias is predicted minus measured; negative means too cool, the opposite sign to the error plot. Underprediction is the largest measured-minus-predicted error in the scored sessions. Seven candidates (CSV)
What was withheld, and what these results can establish
Public marginal distributions were known before the split. Whole held-out trajectories were withheld during fitting and selection. Their outcomes are now exposed and cannot serve as fresh confirmation for a revised model.
The seven candidates and criteria were frozen before scoring. The package verification matched all 115 recorded model, code and prediction hashes and reproduced the session scores from original samples. Every model began with four measured temperatures. No result establishes cross-unit, batch, changed-design or unknown-initial-state performance.
The figure connects selected original samples while retaining extrema. Scores use all 324,526 original samples per model after initialization, at 2 Hz. The source has no calibrated measurement-uncertainty record for these scores. Data schema and display method. Complete frozen summary.
Measurements by LEA, Paderborn University, published by Kirgsn (wkirgsn) as Electric Motor Temperature, version 3. Frozen predictions and adapted data figures by Dynamical, shared under CC BY-SA 4.0. Attribution and reuse terms.
The original authors’ research provides context. The seven fitted candidates here are our separate investigation.