Page 8 - InnoRenew CoE International Conference 2022
P. 8
imizing neural networks for building energy demand estimation
Tamás Storcz1*, István Kistelegdi2, Zsolt Ercsey3
1 Department of System and Software Technology, University of Pécs, Boszorkány u. 2, 7624 Pécs, Hungary, storcz.tamas@mik.pte.hu
2 János Szentágothai Research Centre, Energy Design Research Group, University of Pécs, Ifjúság u. 20, 7624 Pécs, Hungary,
kistelegdisoma@mik.pte.hu
3 Department of System and Software Technology, University of Pécs, Boszorkány u. 2, 7624 Pécs, Hungary, ercsey.zsolt@mik.pte.hu
* Corresponding author
For energy optimization, it is necessary to know the energy (e.g. annual heating energy) demands of buildings.
Basically, building energy demand simulations are done by complex software systems, where environmental,
building, and engineering parameters must be set and then an accurate simulation is computed. It takes a lot of
time to configure the system, calculate and process the simulation result. Moreover, this process requires time
of architect and IT experts.
Using regression models to replace these resource demanding simulations would simplify and accelerate
optimization processes. Among available regression models, a neural network-based regression model could be
suitable for energy demand estimations. Structural flexibility of neural networks makes them a good choice and
popular these days.
In this study, the process and results of hyper parameter selections for a neural network regression model is
presented. Effects of input selection, network structure, and initialization are measured to support and validate
the selection of hyper parameters of the network and model creation. Simplification of network structure and
reducing model complexity, has multiple effects. It reduces learning time and could increase generic regression
accuracy by less chance of overfitting.
The overall result is a dense neural network-based regression model with maximized network performance and
minimized network structure and length of learning process. This model can be integrated into building energy
demand optimization processes with ease or can serve as base of experiments enhancing regression models by
changing input structure.
Keywords: energy, building, demand estimation, neural network
Acknowledgment: This work was partially supported by the [2019-2.1.11-TÉT Bilateral Scientific and
Technological Cooperation].
8 17–18 NOVEMBER 2022 I IZOLA, SLOVENIA
Tamás Storcz1*, István Kistelegdi2, Zsolt Ercsey3
1 Department of System and Software Technology, University of Pécs, Boszorkány u. 2, 7624 Pécs, Hungary, storcz.tamas@mik.pte.hu
2 János Szentágothai Research Centre, Energy Design Research Group, University of Pécs, Ifjúság u. 20, 7624 Pécs, Hungary,
kistelegdisoma@mik.pte.hu
3 Department of System and Software Technology, University of Pécs, Boszorkány u. 2, 7624 Pécs, Hungary, ercsey.zsolt@mik.pte.hu
* Corresponding author
For energy optimization, it is necessary to know the energy (e.g. annual heating energy) demands of buildings.
Basically, building energy demand simulations are done by complex software systems, where environmental,
building, and engineering parameters must be set and then an accurate simulation is computed. It takes a lot of
time to configure the system, calculate and process the simulation result. Moreover, this process requires time
of architect and IT experts.
Using regression models to replace these resource demanding simulations would simplify and accelerate
optimization processes. Among available regression models, a neural network-based regression model could be
suitable for energy demand estimations. Structural flexibility of neural networks makes them a good choice and
popular these days.
In this study, the process and results of hyper parameter selections for a neural network regression model is
presented. Effects of input selection, network structure, and initialization are measured to support and validate
the selection of hyper parameters of the network and model creation. Simplification of network structure and
reducing model complexity, has multiple effects. It reduces learning time and could increase generic regression
accuracy by less chance of overfitting.
The overall result is a dense neural network-based regression model with maximized network performance and
minimized network structure and length of learning process. This model can be integrated into building energy
demand optimization processes with ease or can serve as base of experiments enhancing regression models by
changing input structure.
Keywords: energy, building, demand estimation, neural network
Acknowledgment: This work was partially supported by the [2019-2.1.11-TÉT Bilateral Scientific and
Technological Cooperation].
8 17–18 NOVEMBER 2022 I IZOLA, SLOVENIA
