Optimal- heating-start strategies can reduce unnecessary heating, ventilation, and air conditioning (HVAC) operation while maintaining acceptable indoor temperatures at occupancy. However, forecasting models are often evaluated using aggregate prediction errors that may not reflect operational performance. This study compares four data-driven approaches: gated recurrent unit (GRU) and long short-term memory (LSTM), recurrent networks with different memory mechanisms; temporal convolutional network–GRU (TCN-GRU), combining temporal convolutions and recurrence; and extreme gradient boosting (XGBoost), using gradient-boosted decision trees. Using measurements from an occupied office building, these models were compared with four conventional optimal-start strategies using 24 h of historical inputs and a 6 h prediction horizon. Performance was assessed through temperature prediction errors, recovery-time errors, temperature-deficit exposure, and unnecessary heating runtime. Uncertainty was evaluated using paired day-level bootstrap resampling of 22 recovery events. XGBoost achieved the lowest temperature mean absolute error (MAE) (0.182

°

C), while LSTM achieved the lowest recovery-time MAE (25 min) and unnecessary heating runtime (5.7 min/day). TCN-GRU achieved the lowest temperature-deficit exposure (4.9 K

\cdot

min/day), despite a higher aggregate forecasting error. With considerable uncertainty in some comparisons, no model was consistently superior across objectives. These findings show that optimal-start model selection should consider operational trade-offs, not forecasting accuracy alone.

Publications with shared reference:

0 publications found