Automated machine learning-based building energy load prediction method

Abstract

The application of data-driven building energy load prediction technologies remains a time-consuming effort, since it highly relies on human expertise to train data-driven building energy load prediction models. To address this issue, this study proposes an automated machine learning-based method which can develop accurate data-driven building energy load prediction models with little human assistance. The potential of six representative automated machine learning frameworks (AutoWeka, H2O, TPOT, AutoGluon, FLAML and AutoKeras) are investigated through utilizing them to forecast 1-h ahead heating, cooling and electrical loads of three real-world buildings. The results show that the proposed method outperforms manual modeling. The accuracy of the proposed method increases by 1.10%–18.66% compared to manual modeling. It is also demonstrated that AutoGluon and FLAML can obtain high prediction accuracy using short training times. Moreover, AutoWeka, H2O and TPOT usually need longer training times than AutoGluon and FLAML to develop high-accuracy prediction models. As for AutoKeras, it doesn't show high accuracy on building energy load prediction. This study provides guidelines for making data-driven building energy load prediction technologies easier to use in practical applications.