Investigation of long short-term memory networks to temperature prediction for permanent magnet synchronous motors

Monitoring critical temperatures in permanent magnet synchronous motors (PMSMs) is crucial to ensure safe operation and maximum device utilization as well. In this work, the application of recurrent neural networks featuring memory blocks (LSTMs/GRUs) are investigated upon their suitability to accurate temperature time series prediction inside PMSMs or similar motor types, which is the first time in literature to the author's best knowledge. Considered motor components are stator yoke, teeth and winding as well as the rotor's permanent magnets of a highly-utilized PMSM for electric vehicle applications. Having benchmark data available, numerous neural networks are trained and optimized with the aid of the Chainer framework and particle swarm optimization is conducted for finding suitable model hyper-parameters (e.g. number of hidden neurons or layers) on a computing cluster. It is found, that the Euclidean norm performance (in the range of 1-3 K) is similar but the worst-case predication errors (in the range of 9-14 K) are significantly higher compared to established modeling techniques like lumped-parameter thermal networks (LPTNs). This initial investigation motivates future research to increase ANN-based estimation accuracy by taken other ANN topologies, training methods or hyper-parameter optimization approaches into account.