Real-time temperature prediction of electric machines using machine learning with physically informed features

Abstract

Accurate estimation of the internal temperatures of electric machines is critical to increasing their power density and reliability since key temperatures, such as magnet temperature, are often difficult to measure. This work presents a new machine learning based modelling approach, incorporating novel physically informed feature engineering, which achieves best-in-class accuracy and reduced training time. The different features introduced are proportional to sources of machine losses and require no prior knowledge of the machine, hence the models are completely data driven. Evaluation using a standard experimental dataset shows that modelling errors can be reduced by up to 82.5%, resulting in the lowest mean squared error recorded in the literature of 2.40 K2. Additionally, models can be trained with less training data and have lower sensitivity to data quality. Specifically, it was possible to train a loss enhanced multilayer perceptron model to a mean squared error <5 K2 with 90 h of training data, and an enhanced ordinary least squares model with just 60 h to the same criteria. The inference time of the model can be 1–2 orders of magnitude faster than competing models and requires no time to optimise hyperparameters, compared to weeks or months for other state-of-the-art prediction methods. These results are highly important for enabling low-cost real-time temperature monitoring of electric machines to improve operational efficiency, safety, reliability, and power density.