TopoBrick Enables Zero-Shot IoT Forecasting via Topology Sampling
A new training-free framework uses building knowledge graphs to select the most relevant exogenous variables for sensor forecasting.
TopoBrick solves the problem of treating building sensors as isolated data streams by using an agentic topology sampler to identify the most relevant external variables for any given target. This training-free framework leverages building knowledge graphs to create a structural skeleton, allowing it to forecast IoT data without needing building-specific training sets.
https://arxiv.org/abs/2607.06349
The system organizes variables by their availability at deployment, distinguishing between past sensor states and future-known data like meteorological conditions and schedules. This approach allows the model to outperform strong zero-shot foundation-model baselines and remain competitive with models that are fully trained on a specific building's data.
Topology-aware sampling proves more reliable than random or fixed-hop selection. The gains are most evident in physically coupled systems, such as HVAC and weather-driven sensing variables, where spatial hierarchy and operational context dictate the data's behavior.
Moving toward zero-shot capabilities reduces the overhead of deploying customized models for every new facility. The ability to maintain high accuracy without site-specific training streamlines the scaling of automated building energy management.