Smart Building Temperature Forecasting with Probabilistic Temporal Fusion Transformers

Accurate long-horizon forecasting of thermal dynamics in smart buildings is essential for energy optimization, occupant comfort, and predictive maintenance. In this work, we propose a probabilistic forecasting framework based on a Temporal Fusion Transformer (TFT) architecture, trained on zone-level temperature observations and multimodal exogenous factors. Unlike deterministic models, our approach estimates both predictive means and uncertainties via Gaussian likelihood modeling. We evaluate the model on a large-scale dataset of building temperatures, measuring accuracy via mean absolute error (MAE) and distributional quality via Kullback–Leibler (KL) divergence. Our framework achieves stable 6-month autoregressive forecasts with interpretable uncertainty quantification, demonstrating the feasibility of reliable long-term predictive control in smart building environments.