Prediction of energy use intensity of urban buildings using the semi-supervised deep learning model

Abstract

Prediction of building energy performance is a critical strategy for building energy management. Extant studies established city-scale prediction models only based on buildings with energy data. However, building energy data in most cities is limited, which may impair model performance. A large number of unlabeled buildings (without energy data) may reveal important energy use knowledge, but few studies have explored their capability to improve building energy prediction. Therefore, a novel semi-supervised deep learning method, namely dynamically updated multi-fold semi-supervised learning method based on deep neural networks (DUMSL-DNN) is proposed to predict building energy use intensity (EUI) by utilizing unlabeled samples. Manhattan is selected as a case study, which contains 4854 labeled samples and 34,456 unlabeled samples. Compared with the optimal DNN model, DUMSL-DNN can improve building EUI prediction with root-mean-square error (RMSE) reduced by 9.36% and mean absolute error (MAE) reduced by 9.43%. The DUMSL method is superior to typical semi-supervised learning methods with the lowest RMSE of 0.5207 and the lowest MAE of 0.3325. By the implementation of DUMSL-DNN, more areas with high EUI are identified in Manhattan. Specifically, commercial buildings and residential buildings built before 1965 have higher EUI. Measures are accordingly proposed to improve building energy efficiency.