Journal: IEEE Transactions on Industry Applications
Authors: Ji-Hyeon Lee, Hyun-Su Kim, Soo-Hwan Park, Jin-Cheol Park, Myung-Seop Lim
DOI: 10.1109/TIA.2026.3721841
In recent years, induction motors (IMs), which do not require rare-earth materials, have attracted increasing attention. The efficiency of IMs is a critical design parameter because it directly affects the overall system efficiency, and iron loss accounts for a major portion of the total loss. Therefore, accurate prediction of iron loss during the design stage is essential. Transient analysis can evaluate iron loss with high accuracy; however, it requires a high computational cost owing to the induced rotor bar current and the frequency difference between the stator current and slip frequency. Conversely, the virtual blocked-rotor (VBR)-based frequency analysis significantly reduces the computational burden by eliminating time-stepping integration, but its accuracy deteriorates because slot-harmonic components caused by rotor motion are not reflected. To address this limitation, this paper proposes a transfer learning-based iron loss prediction method that improves the accuracy of VBR-based frequency analysis while preserving its computational advantage over transient analysis. A deep neural network is first trained using a large VBR-based dataset that excludes slot harmonic components and is then fine-tuned using a small amount of transient analysis data that includes slot harmonics. As a result, iron loss including slot-harmonic effects can be predicted accurately with only a limited amount of computationally expensive transient analysis data, thereby improving both accuracy and efficiency in IMs design.