Data-driven approach to predicting the energy performance of residential buildings using minimal input data

Abstract

To achieve carbon neutrality, the South Korean government has been retrofitting existing buildings to reduce their energy consumption. However, existing buildings often lack sufficient information for building energy modeling. In this study, a model was developed for predicting heating energy demand using only information obtained from a preliminary survey. Three different models were considered: multiple linear regression (MLR), artificial neural network (ANN), and support vector regression (SVR). They were then trained with data on old houses of low-income households in South Korea and were used to predict the heating energy demand of individual household units. Different input variables were applied to the initial models to identify target variables and tune the hyperparameters. In tests, ANN was slightly more accurate than SVR. SVR required a shorter total running time (training and prediction), but ANN was 10 times faster than SVR when only prediction was considered. Therefore, ANN was selected. The selected model method takes 0.215 s for 10,000 cases. On the other hand, the previous method takes approximately an hour for one case except time for moving to a field. This shows the suggested method is much faster than the previous one. The proposed model was applied to a case study, and the predicted and true values had a relative error of only 1.40%. The proposed model can be used to predict the heating energy demand of old houses while requiring only the heating area and construction year as inputs.