Sensor-Only Occupant-Centric Ventilation Control with LLM-Augmented Decision Support for Human–Building Interactions

A sensor-only occupant-centric ventilation control framework augmented with large language model based decision support is proposed to address human-building interactions in smart building operations. The system uses low-cost environmental sensors including carbon dioxide concentration, temperature, humidity, passive infrared motion detection, door state, and smart meter data to preserve privacy while capturing occupant presence and activity patterns. Occupancy is inferred using a carbon dioxide mass-balance inversion model with Bayesian filtering, incorporating motion and door state signals for accuracy. Short-term forecasts of indoor air quality and thermal conditions are generated using time-series models such as long short-term memory networks or gradient boosting. These predictions are applied by a model predictive controller to optimize ventilation and cooling setpoints under air quality, comfort, and equipment constraints while minimizing energy consumption. The large language model module acts as a supervisory layer that reviews control proposals, detects potential violations or inefficiencies, incorporates contextual factors such as scheduled events, and recommends adjustments. The module also generates human-readable, context-aware explanations linking each control decision to predicted occupancy, air quality trajectories, and energy trade-offs, enhancing operator trust and enabling human-in-the-loop interactions. The framework is evaluated using real-building sensor datasets and EnergyPlus or CONTAM simulations, with performance compared against fixed-schedule control, threshold-based demand-controlled ventilation, and predictive control without decision support. Results show reductions in air quality exceedance hours and ventilation or cooling energy use, while maintaining stable actuation and providing clear decision rationales. This integration of predictive control and large language model based decision augmentation offers a practical, adaptive, and explainable approach for occupant-centric and privacy-preserving ventilation management, advancing research in human-building interaction and sustainable smart building practices.