Details
| Originalsprache | Englisch |
|---|---|
| Titel des Sammelwerks | Proceedings of the 31st International Workshop on Intelligent Computing in Engineering |
| Seiten | 385-392 |
| Seitenumfang | 8 |
| Publikationsstatus | Veröffentlicht - 3 Juli 2024 |
| Veranstaltung | 31st International Workshop on Intelligent Computing in Engineering, EG-ICE 2024 - Vigo, Spanien Dauer: 3 Juli 2024 → 5 Juli 2024 |
Abstract
Building energy modeling plays a crucial role in decreasing global energy demand. Data-driven models has become highly popular in building energy modeling due to high robustness and accuracy. These methods mainly rely on measured data from real buildings. Hence, the limitations in measured data also limits the data-driven approaches’ performance and reliability. To address this challenge, transfer learning approaches have become known in learning from a source dataset and transfer the knowledge to another similar dataset. In this study, we have trained a Random Forest Regressor (RFR) and a Long Short Term Memory (LSTM) model on simulation data as the source dataset. Then, we transferred the trained models and train them on a subset of the real building data as target data. Finally, we tested the models on a subset of the target dataset to evaluate the performance of the developed models in transferring the knowledge from simulation dataset to the measured dataset from buildings. Moreover, as a baseline, we tested the developed models on the target dataset without training on the target dataset to evaluate and compare the performance of transfer learning and non-transfer learning models on the real dataset. The results shows that the LSTM approach has a high capability in learning from source dataset and transfer the knowledge to the target dataset while keep the knowledge from the source dataset. However, the RFR model forgets the collected knowledge from the source data during the training on the target data. This ends up in significantly lower performance even in comparison with non-transfer learning RFR model. The transfer learning has the highest performance in forecasting building energy with R-Squared, MAPE, and RMSE of 0.87, 15%, and 1.91 GJ, respectively.
ASJC Scopus Sachgebiete
- Informatik (insg.)
- Angewandte Informatik
- Ingenieurwesen (insg.)
- Allgemeiner Maschinenbau
Zitieren
- Standard
- Harvard
- Apa
- Vancouver
- BibTex
- RIS
Proceedings of the 31st International Workshop on Intelligent Computing in Engineering. 2024. S. 385-392.
Publikation: Beitrag in Buch/Bericht/Sammelwerk/Konferenzband › Aufsatz in Konferenzband › Forschung › Peer-Review
}
TY - GEN
T1 - Transfer Learning Modeling for Building Energy Performance Prediction Using Measured Data and Simulation Information
AU - Nabavi, Seyed Azad
AU - Geyer, Philipp
N1 - Publisher Copyright: © 2024 Proceedings of the 31st International Workshop on Intelligent Computing in Engineering, EG-ICE 2024. All rights reserved.
PY - 2024/7/3
Y1 - 2024/7/3
N2 - Building energy modeling plays a crucial role in decreasing global energy demand. Data-driven models has become highly popular in building energy modeling due to high robustness and accuracy. These methods mainly rely on measured data from real buildings. Hence, the limitations in measured data also limits the data-driven approaches’ performance and reliability. To address this challenge, transfer learning approaches have become known in learning from a source dataset and transfer the knowledge to another similar dataset. In this study, we have trained a Random Forest Regressor (RFR) and a Long Short Term Memory (LSTM) model on simulation data as the source dataset. Then, we transferred the trained models and train them on a subset of the real building data as target data. Finally, we tested the models on a subset of the target dataset to evaluate the performance of the developed models in transferring the knowledge from simulation dataset to the measured dataset from buildings. Moreover, as a baseline, we tested the developed models on the target dataset without training on the target dataset to evaluate and compare the performance of transfer learning and non-transfer learning models on the real dataset. The results shows that the LSTM approach has a high capability in learning from source dataset and transfer the knowledge to the target dataset while keep the knowledge from the source dataset. However, the RFR model forgets the collected knowledge from the source data during the training on the target data. This ends up in significantly lower performance even in comparison with non-transfer learning RFR model. The transfer learning has the highest performance in forecasting building energy with R-Squared, MAPE, and RMSE of 0.87, 15%, and 1.91 GJ, respectively.
AB - Building energy modeling plays a crucial role in decreasing global energy demand. Data-driven models has become highly popular in building energy modeling due to high robustness and accuracy. These methods mainly rely on measured data from real buildings. Hence, the limitations in measured data also limits the data-driven approaches’ performance and reliability. To address this challenge, transfer learning approaches have become known in learning from a source dataset and transfer the knowledge to another similar dataset. In this study, we have trained a Random Forest Regressor (RFR) and a Long Short Term Memory (LSTM) model on simulation data as the source dataset. Then, we transferred the trained models and train them on a subset of the real building data as target data. Finally, we tested the models on a subset of the target dataset to evaluate the performance of the developed models in transferring the knowledge from simulation dataset to the measured dataset from buildings. Moreover, as a baseline, we tested the developed models on the target dataset without training on the target dataset to evaluate and compare the performance of transfer learning and non-transfer learning models on the real dataset. The results shows that the LSTM approach has a high capability in learning from source dataset and transfer the knowledge to the target dataset while keep the knowledge from the source dataset. However, the RFR model forgets the collected knowledge from the source data during the training on the target data. This ends up in significantly lower performance even in comparison with non-transfer learning RFR model. The transfer learning has the highest performance in forecasting building energy with R-Squared, MAPE, and RMSE of 0.87, 15%, and 1.91 GJ, respectively.
UR - http://www.scopus.com/inward/record.url?scp=85207205276&partnerID=8YFLogxK
UR - https://3dgeoinfoeg-ice.webs.uvigo.es/proceedings
M3 - Conference contribution
AN - SCOPUS:85207205276
SP - 385
EP - 392
BT - Proceedings of the 31st International Workshop on Intelligent Computing in Engineering
T2 - 31st International Workshop on Intelligent Computing in Engineering, EG-ICE 2024
Y2 - 3 July 2024 through 5 July 2024
ER -