TY - JOUR
T1 - Meta-regression framework for energy consumption prediction in a smart city
T2 - A case study of Songdo in South Korea
AU - Carrera, Berny
AU - Peyrard, Suzanne
AU - Kim, Kwanho
N1 - Publisher Copyright:
© 2021 Elsevier Ltd
PY - 2021/9
Y1 - 2021/9
N2 - Nowadays the concept of smart city has gained in popularity in urban studies. A smart city collects diverse information to monitor and analyze urban systems, such as energy management. It is crucial for smart cities to monitor energy efficiency to be sustainable. In this study, we search to expose the possibilities offered by the energy data of Songdo, a South Korean smart city. First, we have highlighted the ability of Songdo to generate energy data. Second, we used those data to predict its evolution. As a result, we develop a short-term stacking ensemble model for energy consumption in Songdo, focusing on a three-months-ahead prediction problem. To obtain this result, first we design a baseline regressors for the prediction, second, we apply a three-combination of each best model of the base regressors, and finally, a weighted meta-regression model was applied using meta-XGBoost. We call the resulting model stacking ensemble model. The proposed stacking ensemble model combines the best ensemble networks to improve performance prediction, yielding an R2 value of 97.89 %. The results support the effectiveness of the ensemble networks, which use Artificial Neural Networks (ANN), CatBoost and Gradient Boosting. This study also shows that the weighted meta model outperforms several machine learning models in terms of R2, MAE and RSME.
AB - Nowadays the concept of smart city has gained in popularity in urban studies. A smart city collects diverse information to monitor and analyze urban systems, such as energy management. It is crucial for smart cities to monitor energy efficiency to be sustainable. In this study, we search to expose the possibilities offered by the energy data of Songdo, a South Korean smart city. First, we have highlighted the ability of Songdo to generate energy data. Second, we used those data to predict its evolution. As a result, we develop a short-term stacking ensemble model for energy consumption in Songdo, focusing on a three-months-ahead prediction problem. To obtain this result, first we design a baseline regressors for the prediction, second, we apply a three-combination of each best model of the base regressors, and finally, a weighted meta-regression model was applied using meta-XGBoost. We call the resulting model stacking ensemble model. The proposed stacking ensemble model combines the best ensemble networks to improve performance prediction, yielding an R2 value of 97.89 %. The results support the effectiveness of the ensemble networks, which use Artificial Neural Networks (ANN), CatBoost and Gradient Boosting. This study also shows that the weighted meta model outperforms several machine learning models in terms of R2, MAE and RSME.
KW - Deep learning
KW - Energy consumption
KW - Ensemble meta regressor
KW - Machine learning
KW - Smart city
KW - Topical information
UR - http://www.scopus.com/inward/record.url?scp=85106522009&partnerID=8YFLogxK
U2 - 10.1016/j.scs.2021.103025
DO - 10.1016/j.scs.2021.103025
M3 - Article
AN - SCOPUS:85106522009
SN - 2210-6707
VL - 72
JO - Sustainable Cities and Society
JF - Sustainable Cities and Society
M1 - 103025
ER -