Covariates Are the Key to Accurate Probabilistic Building Energy Forecasting with Time Series Foundation Models

Probabilistic building energy demand forecasting is critical for advanced energy management but remains challenging due to building heterogeneity, data scarcity, and non-stationary behavior. Although Time Series Foundation Models (TSFMs) promise scalable generalization from large-scale pretraining and are intended for zero-shot forecasting without additional training, their default settings often fail to reach satisfactory accuracy. As a result, prior work frequently relies on fine-tuning to achieve competitive performance, thereby undermining the out-of-the-box applicability of the models by increasing computational cost, data requirements, and sensitivity to model configuration. Furthermore, successful covariate integration in TSFMs remains largely unexplored, limiting their effectiveness in settings dominated by external influences such as weather conditions or calendar effects. In this work, we aim to demonstrate that covariate integration in zero-shot settings constitutes a more robust and scalable alternative to fine-tuning for building energy parameter forecasting. We evaluate three TSFMs on two real-world building energy demand time series from an office building, with electrical load and cooling demand as target series that exhibit challenges like external influences, limited data and concept drift. Zero-shot and fine-tuned TSFMs are benchmarked against multiple baseline models. Our results show that covariate integration yields substantial performance gains for Chronos-2 and Chronos across both target series, outperforming all baseline models, while MOIRAI does not benefit from future covariates due to insufficient pretraining. Fine-tuning proves to be highly sensitive to the target series and model configuration, while comparable or superior performance can already be achieved in zero-shot settings by substantially extending the context length, which also increases robustness to concept drift. These findings highlight the potential of covariate-informed zero-shot forecasting with TSFMs as a practical and scalable alternative in real-world building energy applications.