Indoor Wellness Optimization in Smart Buildings Based on Ambient Variable Dimensionality Reduction

Achieving thermal comfort in indoor environments is crucial for both occupant well-being and energy sustainability. Although recent machine learning (ML) and deep reinforcement learning (DRL) approaches have shown promise for real-time heating, ventilation, and air conditioning (HVAC) control, their reliance on simulated environments, high computational costs, and limited generalizability hinder practical adoption. In particular, these methods often require large datasets and offer limited interpretability of variable correlations. This paper proposes a lightweight and scalable alternative by integrating multivariate analysis (MVA) with Internet of Things (IoT) data streams for real-time comfort prediction and control on a flexible number of variables. The approach combines a nonlinear locally weighted regression (LWR) model with a sequential quadratic programming (SQP) optimization layer, enforcing physical and operational constraints to produce interpretable and feasible control actions. The optimization strategy is highly flexible, allowing independent boundary conditions on environmental variables. Dimensionality reduction techniques enable deployment on resource-constrained hardware without sacrificing accuracy. The proposed framework is validated on a real-world dataset collected from three indoor zones at the University of Bologna (Campus of Cesena), comprising 765 observations of environmental variables, including temperature (Temp), relative humidity (RH), and indoor air quality (IAQ), sampled every 20 minutes, along with hourly outdoor conditions and subjective satisfaction data from occupant surveys. The results show predicted wellness improvements of 14.4%, 77.8%, and 89.6% in three different rooms.