A Transformer With Variable-Wise Attention for Multi-Zone Indoor Air Temperature Prediction in a Retail Store

초록

Accurate long-horizon indoor air temperature (IAT) prediction is essential for control-aware heating, ventilation, and air conditioning (HVAC) operation in large commercial buildings, yet the task is complicated by partial observability and the coupled influence of multiple HVAC units on multiple zones. This study presents a Transformer-based predictive model for multi-zone IAT prediction in a large-scale retail store using routinely available building operation data. The decoder augments temporal attention with a variable-wise attention pathway that captures cross-variable interactions among HVAC actuation, external drivers, and zone-temperature feedback, and is trained with an autoregressive rollout loss over day-aligned operating-hour sequences. On the real test set under day-level closed-loop evaluation, the proposed model achieves MSE = 0.140 and R-2 = 0.916, outperforming Transformer, Informer, LSTM-MIMO, and NNARX baselines, and a controlled five-seed ablation confirms that the variable-wise attention pathway provides a statistically significant gain (zone-pooled paired t-test, p = 0.0015). A matched-condition residual analysis further indicates that the model has approached the information ceiling imposed by the observable input set. These results demonstrate that the proposed approach is a practical data-driven solution for long-horizon multi-zone IAT prediction in existing retail buildings with limited sensing.

키워드