Causality-Informed Time-Series Foundation Models for Building Performance Prediction

Abstract: Time-Series Foundation Models, pretrained on large amounts of time-series data, offer strong zero-shot forecasting capabilities. Recent TSFM versions provide improved support for covariate forecasting. This renders them a promising alternative to address current limitations of data-driven building performance prediction approaches: they are highly dependent on data availability, are mostly case-specific, and often favor correlation over causation. In this paper, we compare the predictive performance of two TSFMs with graph neural network approaches and a statistical baseline. We also present an approach for configuring causal covariates for TSFM predictions from graph representations. All models are evaluated using two real-world building datasets.