A research team has built a forecasting system for building sensors that requires no site-specific training and still matches purpose-built models.
TopoBrick treats building sensor networks as what they physically are: nodes embedded in spatial hierarchies, not isolated data streams. The framework constructs a structural skeleton from building knowledge graphs, then uses an agentic sampler to select the external variables most relevant to each target sensor. Those variables are split by when they become available: past sensor readings on one side, future-known data like weather forecasts and occupancy schedules on the other. Tested across three real buildings, TopoBrick outperformed zero-shot foundation-model baselines and stayed competitive with models trained specifically on each building.
The interesting engineering choice here is the agentic sampler. Rather than pulling in every available covariate or walking a fixed number of hops through the graph, the system selects variables based on physical coupling - notably HVAC systems and weather-driven sensors, where generic approaches tend to fail. That specificity is why topology-aware selection beat random, ontology-only, and fixed-hop alternatives in ablation tests. For building operators who manage dozens of sites with inconsistent data histories, a model that works without retraining on each one is practically useful, not just academically tidy.
Zero-shot performance in time-series forecasting has improved fast - foundation models like TimesFM and Moirai have pushed the baseline up - which makes it notable that a structurally-informed, training-free approach can still close the gap with supervised methods.