Options
Anwendung künstlicher neuronaler Netze zur Bestimmung von U-Werten
Citation Link: https://doi.org/10.15480/882.14481
Publikationstyp
Conference Paper
Date Issued
2024-03
Sprache
German
Author(s)
TORE-DOI
Citation
Bauphysiktage in Weimar 2024
Contribution to Conference
Publisher DOI
The in-situ estimation of U-values is usually associated with high inaccuracies and long measuring campaigns. Literature sources mention time intervalls of 72 hours up to more than two weeks as well as derivations of up to 40 % (Evangelisti et al., 2016). Main reasons cited for these circumstances are high thermal capacities, low temperature gradients, transient boundary conditions and alternating directions of the heat flow.
Artificial neural networks (ANNs) offer the possibility of quantifying the correlations listed above. By choosing a regression-based approach implemented with ANNs, transient environmental influences (e.g., heat flow, air temperature and solar radiation) can be correlated with variables that are by definition stationary, such as the U-value.
In this paper, the authors present the application of ANNs for estimating U-values using simulation data. The training data is generated exclusively by transient simulations of building elements based on finite elements, eliminating the need for extensive measurements to generate the training data. In this work, three-layer neural networks with 2 to 20 neurons within the hidden layer are used for the regression of U-values. The regression results obtained are compared with U-values of the mean value method for referencing the proposed methodology.
Artificial neural networks (ANNs) offer the possibility of quantifying the correlations listed above. By choosing a regression-based approach implemented with ANNs, transient environmental influences (e.g., heat flow, air temperature and solar radiation) can be correlated with variables that are by definition stationary, such as the U-value.
In this paper, the authors present the application of ANNs for estimating U-values using simulation data. The training data is generated exclusively by transient simulations of building elements based on finite elements, eliminating the need for extensive measurements to generate the training data. In this work, three-layer neural networks with 2 to 20 neurons within the hidden layer are used for the regression of U-values. The regression results obtained are compared with U-values of the mean value method for referencing the proposed methodology.
Subjects
MLE@TUHH
DDC Class
006: Special computer methods
Publication version
publishedVersion
Loading...
Name
06C2 ANWENDUNG KÜNSTLICHER NEURONALER NETZE ZUR BESTIMMUNG VON U-WERTEN.pdf
Size
812.26 KB
Format
Adobe PDF