Preemptive HVAC setpoint scheduling enabled by large language model-based occupancy prediction: A lecture hall case study

In large-volume spaces such as lecture rooms, the indoor air temperature responds slowly to the cooling operation because of high thermal inertia. Additionally, feedback control based on the current temperature can introduce an additional delay in the equipment response. Together, these time lags may prevent the room from reaching the intended setpoint during occupied hours, thereby degrading thermal comfort. This challenge is exacerbated in lecture rooms, where attendance varies widely across lectures and causes substantial variations in internal heat gain and cooling load. Therefore, preemptive setpoint scheduling based on occupancy prediction is required. In this study, we evaluated a preemptive temperature setpoint scheduling strategy using a simulation framework that explicitly incorporates lecture-specific occupancy prediction. We constructed a control-oriented HVAC simulation for a university lecture room and coupled it with a simplified 5R2C room model. Occupancy was predicted from syllabus information using a large language model (LLM); in prior work, the Direct-LLM approach achieved a mean absolute error of 12.06 and root mean squared error of 15.41 in lecture attendance prediction. Candidate setpoint schedules were exhaustively simulated and screened by using a comfort criterion (indoor air temperature within 23-25 °C during class hours). Among the feasible schedules, the schedule with the lowest total electricity consumption was selected as the optimal schedule. Case studies compared schedules optimized using measured occupancy with those optimized using LLM-predicted occupancy. The results indicated that schedules derived from the predicted occupancy lie near the measured optimal schedule in the setpoint-parameter space and achieve similarly low electricity consumption while satisfying the comfort criterion. These findings support the practical applicability of LLM-based semantic occupancy prediction for preemptive setpoint scheduling in lecture rooms with uncertain and highly variable attendance.