Leveraging surrogate modeling using synthetic data for energy efficient decision-making in residential buildings

Thermal building simulations play an important role in achieving energy efficiency, resulting in reduced carbon emissions and operational costs. However, such simulations are input-intensive and may need to be performed by modeling experts. This becomes a bottleneck in the building design process, especially for small residential projects, where there are limitations in terms of time and access to thermal modeling experts and tools. To address this issue, we explored the viability of replacing the simulation modeling approach with Machine learning (ML) models. We accessed and evaluated the suitability, precision and performance of ML models trained on data generated by building thermal simulations. Our study focused on a 2-storey residential building model simulated using the modeling software EnergyPlus™, representing a typical small residential project. We trained and tested multiple classical ML models and Deep learning models to predict the Energy performance index (EPI) of the building. Sensitivity analysis was also performed on the data to assess the effect of various parameters on the EPI and to identify which factors contributed the most to the accuracy of the model. After extensive experimentation and evaluation, we found that the Artificial Neural networks (ANN) was the most accurate. Our model was able to make predictions faster than traditional simulation models without significant loss of accuracy, highlighting the potential for using such surrogate models in place of input- intensive and computation-heavy simulations during the early design stages. This may help in monitoring and exploring multiple design scenarios, streamlining the assessment of building efficiency, non-intrusive load monitoring, and facilitating informed decision-making for policymakers.