Hybrid modeling based on integrating simulation and operational data to improve indoor air temperature predictions, a controlled variable in digital twin models

To achieve net zero emissions in the building and construction sector, there is a growing interest in how buildings can be digitalized to improve energy efficiency through optimal operational strategies to reduce energy consumption during the operational phase. The validity of the savings scenarios is highly dependent on the accuracy of the digitalized building model. However, considering the accuracy of the developed model determines the validity of the energy-saving scenarios, creating accurate models is difficult owing to the limited amount of physical data collected from buildings. Hence, in this study, a hybrid modeling method is proposed to improve the prediction accuracy by integrating the physical model results and operational data to improve the prediction accuracy for actual operating buildings where only partial data collection is provided, mainly for air conditioners. The hybrid model predicts the next day's room temperature by learning the difference between the simulated room temperature based on the laws of physics and historical measurement data. The results showing that the coefficient of variance of root mean squared error (CVRMSE) was 1.5% for the training period, a significant improvement compared to the existing RC model; moreover, the R2 was 0.93 for the hybrid model, indicating high explanatory power. In addition, an average CVRMSE of 3.8% in the period outside the training area was obtained, resulting in a model with improved prediction accuracy compared with the existing RC model. Similar results were obtained for design models without calibration.