Short-Term Prediction Method of Transient Temperature Field Variation for PMSM in Electric Drive Gearbox Using Spatial-Temporal Relational Graph Convolutional Thermal Neural Network

Accurate short-term prediction of transient temperature field (TTF) variation is crucial for the effective thermal management and safe operation of permanent magnet synchronous machine (PMSM) in electric drive gearbox. In this work, a novel short-term prediction method of TTF is introduced, which unifies both, thermal network topology (TNT) graph construction, and modified relational graph convolutional thermal neural network (RGCN) with supervised machine learning. First, the thermal network model of PMSM with composite cooling type is established based on lumped parameter thermal network (LPTN). Furthermore, the thermal network model is simplified as
$n$
-node TNT through loss analysis. Second, ordinary least square (OLS) is used to estimate the main components' temperatures. Considering the TNT, and integrating modified RGCN, the TTF variation prediction model of the spatial-temporal relational graph convolutional thermal neural network based on OLS (OLS-RGCN) is constructed. Finally, the accuracy of OLS-RGCN is verified by using a water-cooled PMSM dataset and a self-measured oil-cooled PMSM dataset. Experimental results show that OLS-RGCN exhibits better comprehensive prediction performance compared with the other two methods during the two datasets. Meanwhile, the global temperature errors are controlled within 5.88 
$^\circ \rm {C}$
and 2.6
$\,^\circ \rm {C}$
, respectively, when prediction time is 10 s, indicating the successful applicability of OLS-RGCN for different PMSM.