Regression-based Model for Predicting Simulated vs Actual Building Performance Discrepancies

Ivan Izonin, Roman Tkachenko, Rosana Caro, Antonio Latorre de la Fuente, Kyrylo Yemets, Stergios Aristoteles Mitoulis

Abstract: Accurately predicting discrepancies between simulated and actual building performance is becoming increasingly crucial in building management and optimization of e.g., Digital Twins and energy efficiency assessments. This challenge is amplified by the growing dependence on simulation models which predict building behavior, energy consumption, and operational efficiency. Despite advancements in simulation technology, aligning these models with real-world data remains a persistent challenge. This study addresses this challenge by developing a regression-based model designed to predict discrepancies between simulated and actual operational characteristics of buildings. The model identifies differences between synchronized time series data to generate a new series that highlights these discrepancies. By employing a sliding window technique, the model processes actual operational data to predict discrepancies. They implemented this approach using AdaBoost and Random Forest, evaluating performance across eight datasets of Indoor Air Temperature from two cities. The results demonstrate that the Random Forest algorithm significantly outperforms AdaBoost, with improvements in R² scores of up to 23 percent and reductions in RMSE by up to 4.5 times.