Thermal Neural Networks

This repository demonstrates the usage of thermal neural networks (TNNs) on an electric motor data set.

The TNN is introduced in:

@misc{kirchgässner2021_tnn_preprint,
  title={Thermal Neural Networks: Lumped-Parameter Thermal Modeling With State-Space Machine Learning}, 
  author={Wilhelm Kirchgässner and Oliver Wallscheid and Joachim Böcker},
  year={2021},
  eprint={2103.16323},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}

Further, this repository supports and demonstrates the findings around a TNN's generalization to Neural Ordinary DIfferential Equations as presented on IPEC2022. If you want to cite that work, please use

@INPROCEEDINGS{kirchgässner_node_ipec2022,
  author={Kirchgässner, Wilhelm and Wallscheid, Oliver and Böcker, Joachim},
  booktitle={2022 International Power Electronics Conference (IPEC-Himeji 2022- ECCE Asia)}, 
  title={Learning Thermal Properties and Temperature Models of Electric Motors with Neural Ordinary Differential Equations}, 
  year={2022},
  volume={},
  number={},
  pages={2746-2753},
  doi={10.23919/IPEC-Himeji2022-ECCE53331.2022.9807209}}

The data set is freely available at Kaggle and can be cited as

@misc{electric_motor_temp_kaggle,
  title={Electric Motor Temperature},
  url={https://www.kaggle.com/dsv/2161054},
  DOI={10.34740/KAGGLE/DSV/2161054},
  publisher={Kaggle}, 
  author={Wilhelm Kirchgässner and Oliver Wallscheid and Joachim Böcker}, 
  year={2021}}

Topology

Three function approximators (e.g., multi-layer perceptrons (MLPs)) model the thermal characteristics of an arbitrarily complex component arrangement forming a system of interest. One outputs thermal conductances, another the inverse thermal capacitances, and the last one the power losses generated within the components.

In contrast to other neural network architectures, a TNN needs at least to know which features are temperatures and which are not. The TNN's inner cell working is that of lumped-parameter thermal networks (LPTNs). A TNN can be interpreted as a hyper network that is parameterizing an LPTN, which in turn is iteratively solved for the current temperature prediction.

In a nutshell, a TNN solves the difficult-to-grasp nonlinearity and scheduling-vector-dependency in quasi-LPV systems, which an LPTN represents.

Code Structure

The streamlined usage can be seen in the jupyter notebooks, one with tensorflow2 (TNN) and two with pytorch (TNN and NODE). The tf2-version makes heavy use of auxiliary functions and classes declared in aux, whereas the pytorch notebooks are self-contained and do not import from 'aux'. In both frameworks, the TNN is defined as a cell class that is plugged into an outer RNN layer.