Thermal neural networks: Lumped-parameter thermal modeling with state-space machine learning

With electric power systems becoming more compact and increasingly powerful, the relevance of thermal stress especially during overload operation is expected to increase ceaselessly. Whenever critical temperatures cannot be measured economically on a sensor base, model lends itself estimate those unknown quantities. Thermal models for are usually required both, real-time capable high estimation accuracy. Moreover, ease implementation time production play an important role. In this work, neural network (TNN) introduced, which unifies consolidated knowledge in form heat-transfer-based lumped-parameter models, data-driven nonlinear function approximation with supervised machine learning. A quasi-linear parameter-varying system identified solely from empirical data, where relationships between scheduling variables matrices inferred statistically automatically. At same time, TNN has physically interpretable states through its state-space representation, end-to-end trainable -- similar deep learning automatic differentiation, requires no material, geometry, nor expert design. Experiments motor data set show that achieves higher temperature accuracies than previous white-/grey- or black-box mean squared error $3.18~\text{K}^2$ worst-case $5.84~\text{K}$ at 64 parameters.