TopoBrick Improves Zero-Shot Building IoT Forecasting.

Existing forecasting methods for building IoT sensors often treat them as isolated data streams or rely on fixed sets of covariates, ignoring their inherent physical, spatial, and operational contexts. This limitation hinders accurate predictions, especially in zero-shot scenarios where no prior training data is available for a specific building. TopoBrick, a novel training-free framework, addresses this by integrating building knowledge graphs to construct a structural skeleton. It then employs an agentic topology sampler to intelligently select target-specific exogenous variables. These variables are organized based on their availability at deployment time, distinguishing between past-known sensor states and future-known calendar, schedule, and meteorological data. Evaluations across three real-world buildings demonstrate that TopoBrick significantly outperforms strong zero-shot foundation model baselines. It even remains competitive with fully trained, building-specific models. Ablation studies confirm that its topology-aware sampling method is more reliable than random or fixed-hop selection, particularly for critical HVAC and weather-driven sensing variables.