Title: | Identification of Thermal Model Parameters Using Deep Learning Techniques |
Authors: | Ševčík, Jakub Šmídl, Václav Votava, Martin |
Citation: | ŠEVČÍK, J. ŠMÍDL, V. VOTAVA, M. Identification of Thermal Model Parameters Using Deep Learning Techniques. In 2022 IEEE 31st International Symposium on Industrial Electronics (ISIE) : /proceedings/. Piscataway: IEEE, 2022. s. 978-981. ISBN: 978-1-66548-240-0 , ISSN: 2163-5145 |
Issue Date: | 2022 |
Publisher: | IEEE |
Document type: | konferenční příspěvek ConferenceObject |
URI: | 2-s2.0-85135786103 http://hdl.handle.net/11025/51295 |
ISBN: | 978-1-66548-240-0 |
ISSN: | 2163-5145 |
Keywords in different language: | deep learning;junction temperature;multistep prediction;neural network;ordinary least squares;thermal model |
Abstract in different language: | Identification of thermal model parameters using multi-step prediction is proposed. Even in the case of a linear model, the multi-step prediction is a non-linear complex function, hence we use techniques of deep learning for its identification. Specifically, we use stochastic gradient descent optimization with importance sampling of mini-batches. The importance function is designed to match the character of thermal experiments in which the step change is less frequent than steady-state operation. The proposed method is demonstrated on the identification of an IGBT module SK 20 DGDL 065 ET. The maximum error of the model identified by the multi-step approach is almost two times smaller than that of the model identified by the least squares. |
Rights: | Plný text je přístupný v rámci univerzity přihlášeným uživatelům. © IEEE |
Appears in Collections: | Konferenční příspěvky / Conference papers (RICE) Konferenční příspěvky / Conference Papers (KEV) OBD |
Files in This Item:
File | Size | Format | |
---|---|---|---|
Sevcik_Identification.pdf | 1,9 MB | Adobe PDF | View/Open Request a copy |
Please use this identifier to cite or link to this item:
http://hdl.handle.net/11025/51295
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.