TY - GEN

T1 - A hybrid-model time-series forecasting approach for reducing the building energy performance gap

AU - Chen, Xia

AU - Guo, Tong

AU - Geyer, Philipp Florian

N1 - Funding Information: We gratefully acknowledge the German Research Foundation (DFG) support for funding the project under grant GE 1652/3-2 in the Researcher Unit FOR 2363. We would like to thank Prof. Peng Xu and his research group at Tongji University, Shanghai, China, for data resources support.

PY - 2021

Y1 - 2021

N2 - The performance gap between predicted and actual energy consumption in the building industry remains an unsolved problem in practice. This paper aims to minimize this gap by proposing a hybrid-model using building simulation and machine learning (ML) models inspired by the concept of time-series decomposition: 1. Using first-principles methods in different levels of information to convert the building discrete features and predictable patterns in time-series format. 2. Import the physical model's output into the ML model as input. 3. Training the ML model to align the performance and calibrate the result. The approach is tested in the measured energy load from an office building in Shanghai. Hybrid-model shows higher accuracy in prediction with a better interpretation for gap magnitude investigation in building energy. In summary, the method demonstrates how domain knowledge via building simulation incorporated with data-driven methods, especially ML leads to improved predictions.

AB - The performance gap between predicted and actual energy consumption in the building industry remains an unsolved problem in practice. This paper aims to minimize this gap by proposing a hybrid-model using building simulation and machine learning (ML) models inspired by the concept of time-series decomposition: 1. Using first-principles methods in different levels of information to convert the building discrete features and predictable patterns in time-series format. 2. Import the physical model's output into the ML model as input. 3. Training the ML model to align the performance and calibrate the result. The approach is tested in the measured energy load from an office building in Shanghai. Hybrid-model shows higher accuracy in prediction with a better interpretation for gap magnitude investigation in building energy. In summary, the method demonstrates how domain knowledge via building simulation incorporated with data-driven methods, especially ML leads to improved predictions.

UR - http://www.scopus.com/inward/record.url?scp=85134228601&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:85134228601

T3 - EG-ICE 2021 Workshop on Intelligent Computing in Engineering, Proceedings

SP - 44

EP - 53

BT - EG-ICE 2021 Workshop on Intelligent Computing in Engineering, Proceedings

A2 - Abualdenien, Jimmy

A2 - Borrmann, Andre

A2 - Ungureanu, Lucian-Constantin

A2 - Hartmann, Timo

CY - Berlin

T2 - 28th International Workshop on Intelligent Computing in Engineering of the European Group for Intelligent Computing in Engineering, EG-ICE 2021

Y2 - 30 June 2021 through 2 July 2021

ER -