A Multimodal Transformer for Joint Prediction of Comfort and Energy Consumption in Smart Buildings

This paper presents a multimodal transformer-based framework for the joint prediction of indoor thermal comfort and energy efficiency using real-world building management system (BMS) datasets. Unlike traditional comfort models that rely on fixed physical assumptions and subjective surveys, the proposed approach adopts physics-guided, data-driven learning to capture nonlinear and time-dependent interactions among environmental conditions, HVAC operation, and occupancy-related variables. Thermal comfort labels are computed using the PMV–PPD formulation defined by ASHRAE Standard 55, assuming standard metabolic rate and clothing insulation due to the lack of direct measurements in routine BMS data. A temperature-driven baseline HVAC energy proxy is derived using change-point regression. The proposed transformer architecture fuses multivariate temporal sequences to jointly predict both comfort and energy baseline targets through a dual-head regression formulation. The model is validated on two complementary datasets representing steady-state and dynamically perturbed thermal conditions. The proposed approach consistently outperforms linear regression, random forest, and LSTM baselines, achieving mean absolute errors below 0.03 and R2 values exceeding 0.98 with corresponding RMSE values below 0.035 for both targets. Residual and calibration analyses confirm stable, unbiased prediction behavior across wide temperature ranges. The results highlight the strong potential of attention-based multimodal learning for future comfort-aware building energy optimization and digital twin integration.