Examining the generalizability of inverse surrogate models for different geometries and locations

• Investigates generalizability of inverse surrogate models (ISMs). • Compares CNN, RNN, and transformer architectures in building energy modeling. • Uses internal temperature and energy data as time-series inputs. • Demonstrates ISM applicability across diverse building shapes and locations. • Highlights transformers’ strengths in sequence learning for predictive modeling. While building surrogate modelling has been shown to accurately replicate the outputs of computationally intensive building energy modelling, successfully adopting surrogate modelling in practice still has challenges. As surrogate models are machine learning models, they require an extensive quantity of training data in order to train effectively. The process of acquiring training data often requires numerous simulation runs of a building energy model. To offset this issue, surrogate models that demonstrate a suitable level of generalizability can be applied successfully to multiple projects without the need for the further generation of data. This study examines the generalizability of multiple inverse surrogate models. Inverse surrogate modelling is a more difficult task than traditional surrogate modelling as it tries to extract building energy model inputs from output data. As the output data required to do this is often comprehensive, deep learning models are preferred. For the inverse surrogate models, a basic deep artificial neural network, convolutional neural network, recurrent neural network and transformer were examined. Output data in this study consisted primarily of temperature and energy time series data with input data being building energy model parameters reflective of thermally important building characteristics. Generalizability is assessed by first training the inverse surrogate models on data from 3 separate building energy models. Each of the building energy models contain geometry that is randomly scaled. Additionally we examine training the inverse surrogate models on building energy model data produced with multiple locations as well as on data from all building energy models at once. Parameters relating to the building envelope demonstrated the highest prediction performance among the models, whereas the prediction performance for less influential parameters was more varied depending on the inverse surrogate model. Overall, the convolutional neural network typically outperformed the other models with the recurrent neural network and transformer producing slightly worse performance. The artificial neural network was unable to accurately predict parameters outside of a select few that were highly influential to the time-series data. In the cases of training with data from multiple locations or all buildings at once, prediction performance decreased, however several parameters remained predictable.