Huitong You

Indoor temperature forecasting in large public buildings is affected by heating, ventilation, and air-conditioning (HVAC) control, occupancy, airflow, and zone use, not only by recent temperature histories. This study asks whether operationally derived functional-zone proxy labels—room groupings derived heuristically from HVAC statistics rather than from field-verified building metadata—can provide a useful learnable context for temporal convolutional network (TCN)-based room-level forecasting when verified semantic metadata are unavailable. Using the public BEAR (Building Efficiency and Renewables) multizone building HVAC dataset, we construct a reproducible operational proxy scenario: eighty rooms are grouped from occupancy, airflow, damper, control-command, and temperature-variability statistics, and Niagara-inspired point-naming conventions are used only as structured metadata descriptors, not as tags exported from a deployed system. The proposed SemanticTCN augments a TCN backbone with learnable room, temperature-variability proxy embeddings. Experiments compare persistence, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), TCN, DLinear, PatchTST, a TCNOnly variant, SemanticTCN, and ablations for 1 h, 6 h, and 24 h forecasts, evaluated by mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and the coefficient of determination ($R^2$). Persistence is strongest at 1 h and 24 h, while DLinear is strongest at 6 h, indicating strong thermal inertia and daily regularity. Within the TCN family, SemanticTCN improves over TCNOnly in the locked run, shows a statistically significant within-distribution MAE improvement across multiple seeds (most consistently at 6 h), and reduces large-error rates; however, a leave-one-zone-out experiment shows that this advantage does not transfer to rooms unseen during training, and a five-seed operational-versus-random-proxy control shows that the specific operational grouping rule is not statistically distinguishable from arbitrary grouping. The bounded conclusion is that providing room/group context to the TCN backbone helps predict rooms whose identity was observed during training, especially at 6 h, but the gain reflects room-identity context rather than generalizable operational semantics: the proxy labels are heuristic operational constructs rather than verified building semantics; they do not outperform arbitrary grouping under multi-seed control, and they do not generalize to unseen rooms under this split.