TY - JOUR
T1 - Short-term building energy model recommendation system
T2 - A meta-learning approach
AU - Cui, Can
AU - Wu, Teresa
AU - Hu, Mengqi
AU - Weir, Jeffery D.
AU - Li, Xiwang
N1 - Funding Information:
This research was partially supported by funds from the National Science Foundation award under grant number CNS-1239257 and from the United States Transportation Command (USTRANSCOM) in concert with the Air Force Institute of Technology (AFIT) under an ongoing Memorandum of Agreement. The U.S. Government is authorized to reproduce and distribute for governmental purposes notwithstanding any copyright annotation of the work by the author(s). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of USTRANSCOM, AFIT, the Department of Defense, or the U.S. Government.
Publisher Copyright:
© 2016 Elsevier Ltd.
PY - 2016/6/15
Y1 - 2016/6/15
N2 - High-fidelity and computationally efficient energy forecasting models for building systems are needed to ensure optimal automatic operation, reduce energy consumption, and improve the building's resilience capability to power disturbances. Various models have been developed to forecast building energy consumption. However, given buildings have different characteristics and operating conditions, model performance varies. Existing research has mainly taken a trial-and-error approach by developing multiple models and identifying the best performer for a specific building, or presumed one universal model form which is applied on different building cases. To the best of our knowledge, there does not exist a generalized system framework which can recommend appropriate models to forecast the building energy profiles based on building characteristics. To bridge this research gap, we propose a meta-learning based framework, termed Building Energy Model Recommendation System (BEMR). Based on the building's physical features as well as statistical and time series meta-features extracted from the operational data and energy consumption data, BEMR is able to identify the most appropriate load forecasting model for each unique building. Three sets of experiments on 48 test buildings and one real building were conducted. The first experiment was to test the accuracy of BEMR when the training data and testing data cover the same condition. BEMR correctly identified the best model on 90% of the buildings. The second experiment was to test the robustness of the BEMR when the testing data is only partially covered by the training data. BEMR correctly identified the best model on 83% of the buildings. The third experiment uses a real building case to validate the proposed framework and the result shows promising applicability and extensibility. The experimental results show that BEMR is capable of adapting to a wide variety of building types ranging from a restaurant to a large office, and gives excellent performance in terms of both modeling accuracy and computational efficiency.
AB - High-fidelity and computationally efficient energy forecasting models for building systems are needed to ensure optimal automatic operation, reduce energy consumption, and improve the building's resilience capability to power disturbances. Various models have been developed to forecast building energy consumption. However, given buildings have different characteristics and operating conditions, model performance varies. Existing research has mainly taken a trial-and-error approach by developing multiple models and identifying the best performer for a specific building, or presumed one universal model form which is applied on different building cases. To the best of our knowledge, there does not exist a generalized system framework which can recommend appropriate models to forecast the building energy profiles based on building characteristics. To bridge this research gap, we propose a meta-learning based framework, termed Building Energy Model Recommendation System (BEMR). Based on the building's physical features as well as statistical and time series meta-features extracted from the operational data and energy consumption data, BEMR is able to identify the most appropriate load forecasting model for each unique building. Three sets of experiments on 48 test buildings and one real building were conducted. The first experiment was to test the accuracy of BEMR when the training data and testing data cover the same condition. BEMR correctly identified the best model on 90% of the buildings. The second experiment was to test the robustness of the BEMR when the testing data is only partially covered by the training data. BEMR correctly identified the best model on 83% of the buildings. The third experiment uses a real building case to validate the proposed framework and the result shows promising applicability and extensibility. The experimental results show that BEMR is capable of adapting to a wide variety of building types ranging from a restaurant to a large office, and gives excellent performance in terms of both modeling accuracy and computational efficiency.
KW - Building energy consumption
KW - Feature reduction
KW - Machine learning
KW - Meta-learning
KW - Recommendation system
KW - Time series forecasting
UR - http://www.scopus.com/inward/record.url?scp=84962162042&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84962162042&partnerID=8YFLogxK
U2 - 10.1016/j.apenergy.2016.03.112
DO - 10.1016/j.apenergy.2016.03.112
M3 - Article
AN - SCOPUS:84962162042
SN - 0306-2619
VL - 172
SP - 251
EP - 263
JO - Applied Energy
JF - Applied Energy
ER -