A building does not respond instantly to its heating system. Turn the thermostat up and the air temperature rises over minutes, but the walls, floor, and furniture absorb heat over hours. The building has thermal inertia — a memory of past thermal states encoded in the mass of its structure. Predicting indoor temperature 150 minutes ahead requires modeling this inertia, not just the current HVAC output.
A deep Kalman filter handles the prediction by treating the building as a dynamical system with hidden states (arXiv:2507.02351). The observable is the sensor reading — noisy, sparse, sometimes biased by direct sunlight or proximity to a vent. The hidden states are the thermal masses of the building's structural elements, which absorb and release heat on timescales much longer than the sensor sampling rate.
The Bayesian framework provides uncertainty quantification alongside the point prediction. When the building's behavior deviates from the model — a door left open, an unusual occupancy pattern, equipment generating unexpected heat — the uncertainty estimate widens rather than the prediction silently degrading. The model knows what it doesn't know.
The denoising capability is as important as the prediction. Building sensors are cheap, imprecise instruments in harsh environments. Dust, age, calibration drift, and placement errors introduce systematic and random noise. The Kalman framework separates sensor noise from thermal signal, recovering the building's true thermal trajectory from corrupted measurements.
The structural insight: a building is a thermal system with memory, not a thermostat with a room attached. The walls remember yesterday's temperature. The concrete floor remembers last week's heatwave. Predicting comfort requires modeling the building's thermal memory, which means treating the structure as a participant in the thermal system, not just a container for air. The building breathes through its vents and remembers through its mass. Controlling it requires acknowledging both.