Early predicting cooling loads for energy-efficient design in office buildings by machine learning

Abstract

Energy-efficient building design has become imperative for energy conservation, emissions reduction, and life quality enhancement of occupant. Physics-based whole building energy simulation is widely used to access building energy performance, which requires large amount of information to specify values of input parameters and includes underlying assumptions. This study proposed an alternative model based on machine learning (ML) to predict cooling loads of buildings with few common parameters in the design phase. The ML models were developed and evaluated using a dataset of 243 buildings. Predicted cooling loads from these models were compared to those from the physics-based whole building energy simulation. The proposed model exhibits good agreement with the physics-based whole building energy simulation. The analytical results present that the ML models obtained the correlation coefficient (R) of 0.98–0.99, the mean absolute percentage error (MAPE) of 6.17–12.93% which shows the high agreement between observed and predicted values of cooling loads in building. Notably, the ensemble bagging artificial neural networks yielded the highest R of 0.99, the lowest root-mean-square error (RMSE) of 158.77 kW, the lowest MAE of 112.07 kW, and the lowest MAPE of 6.17% among all models in this study. This research contributes to (i) the state of the knowledge by examining various ML models that can predict cooling loads in the early building design stage; and (ii) the state of practice by providing an alternative tool in the design process through better understanding of relationships between building cooling loads and building characteristics for enhancing energy efficiency in buildings.