Does ChatGPT Know Building Physics? Exploiting Foundation Models for Building Performance Prediction with GNNs

Abstract: Graph Neural Networks have shown promising results to make predictions for time-series data collected by IoT sensors in buildings. While the process of collecting and structuring data is mostly automated, correctly capturing the physical causalities in the models still requires domain knowledge and manual labor. In this paper, we evaluate the capabilities and challenges of Foundation Models like ChatGPT to configure the GNNs. We conduct experiments prompting two different Foundation Models to construct graphs on small and large scale and compare the resulting graphs and the performance of GNNs based on these graphs for a simulated small scale data set.