Work

A thermal model of a 52 kW traction permanent magnet synchronous motor from documentation and test-bench electrical data: Modelica model, virtual sensors, temperature forecasting and permissible load on the open Electric Motor Temperature data set

Year: 2026
Type: article
Abstract: The open Electric Motor Temperature data set of Paderborn University contains 185 h of test-bench recordings of a 52 kW traction interior permanent magnet synchronous motor (IPMSM): 69 load profiles sampled every 0.5 s. We show how a physics-based model, built from documentation and refined with electrical data of the test bench itself, solves practical tasks: virtual temperature sensors, forecasting under a known load profile, time to threshold, replacement of a failed sensor and permissible load. Model. Our model synthesis method built a Modelica model from the datasheet: a quasi-static dq electromagnetic part, losses by kind and a 7-node thermal network of standard-library primitives. Missing data were taken from bench measurements. The permanent-magnet flux linkage (59.5 mVs,−0.121 %/K), the inductances and the copper resistance were identified from controller voltages and currents. Thermal parameters were refined by a portfolio of methods of the ADAO library: global search and local refinement of every solution found, with selection of several distinct optima. Correcting the model with data. A node-by-node error analysis revealed underestimated winding heat capacity and a wrong coolant boundary. After the correction the cross-validated error fell from 9.71 to 6.73 K², for the magnets from 24.7 to 13.7 K². The parameters converged to the values predicted by the analysis. Tasks (held-out set, generalisation profiles G): open loop — 5.8 K² mean error over four temperatures; virtual magnet sensor (model and voltage-based flux observer in a Kalman filter) — 2.8 K RMSE against 4.7–5.5 K for the same model before the correction; winding forecast at 5–60 min — 2.3–2.6 K, magnets at 15–60 min — 3.9–4.3 K against 11–34 K for persistence; time to threshold — median error 11 s for the winding and 1.7 min for the magnets; a failed stator thermocouple is reconstructed with 1.2–2.1 K error; a stuck sensor is detected within 6–8 min without false alarms. The model is exported as an FMU that matches OpenModelica to 0.0004 K. The initial model was tested blind under a pre-registered protocol. The corrected model and the tasks on it were computed on the held-out set after the analysis and are labelled as not blind.
Language: English
Open Access status: green
OpenAlex ID: W7236847963
FWCI:
Cites:
Cited by:
Related to: 10