TopoBrick: Zero-shot building IoT forecasting with topology

TopoBrick, a training-free framework for zero-shot building IoT forecasting, was submitted to arXiv on 7 Jul 2026 by Xiachong Lin, Du Yin, Arian Prabowo, Hao Xue, Wen Hu, Imran Razzak, Matthew Amos, Sam Behrens and Flora D. Salim. The system builds a compact structural skeleton from building knowledge graphs and uses an agentic topology sampler to select target-specific exogenous variables, separating past-known sensor states from future-known calendar, schedule and meteorological variables.

What is TopoBrick and how does it work?

TopoBrick is a training-free pipeline that constructs a compact structural skeleton from building knowledge graphs and then applies an agentic topology sampler to select exogenous variables for zero-shot forecasting. The framework organizes the selected variables by deployment-time availability, explicitly separating past-known sensor states from future-known calendar, schedule and meteorological exogenous variables, and feeds that organized input into a zero-shot forecaster.

The paper describes two core steps. First, building knowledge graphs encode physical topology, spatial hierarchy and operational context and are reduced into a structural skeleton that highlights relevant couplings. Second, an agentic topology sampler examines that skeleton to pick target-specific exogenous variables rather than using fixed covariate sets. That selection is intended to reflect the actual physical and operational links a target sensor has across the building.

How does TopoBrick perform against baselines?

Across three real-world buildings, TopoBrick outperforms strong zero-shot foundation-model baselines and remains competitive with fully trained building-specific models. The authors report experiments on three buildings and include ablations that compare topology-aware sampling against random selection, ontology-only selection and fixed-hop selection.

Those ablations show topology-aware sampling is more reliable than the alternatives, particularly for sensing variables that are physically coupled to HVAC systems and for sensors driven by weather. The paper contains 12 pages, 4 figures and 3 tables documenting the experiments and comparisons.

Why it matters

TopoBrick shifts the input-selection problem from fixed or heuristic covariate sets to a structure-aware, agentic process that uses explicit building knowledge. That matters because building sensors are embedded in physical topology, spatial hierarchy and operational context; treating them as isolated time series misses those couplings. If topology-aware sampling reliably picks the right exogenous inputs, practitioners can deploy forecasting models in new buildings without model retraining or building-specific data collection.

Practically, TopoBrick promises lower setup friction: the framework is training-free and designed for zero-shot deployment, so it is aimed at scenarios where collecting labeled, building-specific data or training bespoke models is costly or slow.

What to watch

Watch for replication across a wider set of building types and sensor modalities beyond the three buildings in this paper, and for published code or datasets that implement the agentic topology sampler. The arXiv submission lists a DOI via DataCite pending registration and includes links to the paper PDF and TeX source.

Acknowledgements and provenance: the manuscript appears on arXiv as arXiv:2607.06349 and was submitted on 7 Jul 2026 by the nine authors listed above. The document notes 12 pages, 4 figures and 3 tables of results.