Smart HVAC optimization using machine learning and self-adaptive NSGA-II for energy-efficient thermal comfort

A Building Management System (BMS) aims to maintain optimal thermal comfort within an air-conditioned space, even as external conditions fluctuate. Integrating an optimization model with the Air Handling Unit (AHU) enhances the unit’s overall performance while minimizing power consumption. This work focuses on designing and constructing an Air Handling Unit that evaluates performance parameters under varying climatic conditions. Air passes through a helical-coil dehumidifier, an ultrasonic humidifier, and a damper into the room to regulate the required conditions. This study proposes a machine learning-based optimization framework to regulate thermal comfort while minimizing energy consumption in Air Handling Units (AHUs). A predictive model was developed using Random Forest Regressor, XGBoost Regressor, and Artificial Neural Network (ANN), trained on experimental data to estimate PMV, PPD, and energy consumption based on input air conditions. A self-adaptive Non-dominated Sorting Genetic Algorithm II (NSGA-II) was employed to predict optimal inlet air parameters—including temperature, velocity, and relative humidity—within defined thermal comfort constraints. The optimization results were validated experimentally using a test configured with input conditions derived from adaptive NSGA-II predictions, and the resulting thermal comfort indices and energy usage were measured. The prediction errors were minimal—0.8% for energy consumption, 1.2% for Predicted Mean Vote (PMV) and 2.7% for Predicted Percentage of Dissatisfaction (PPD)—demonstrating the accuracy and robustness of the approach. Experimental validation under optimized inlet conditions confirmed the model's reliability, with minimal prediction errors of 0.8% in energy consumption, 1.2% in PMV, and 2.7% in PPD relative to measured values. This work confirms the viability of using ML-based multi-objective optimization for clean, energy-efficient, and comfort-focused HVAC control in smart building environments. • Fabricated an AHU test rig to study comfort indices and energy consumption. • Investigated effects of air velocity, temperature, and humidity on PMV and PPD. • Integrated machine learning with optimization to enhance AHU performance. • XGBoost with self-adaptive NSGA-II predicted optimal inlet air conditions. • Experimental results closely matched optimized PMV, PPD, and energy values. • Low prediction errors confirmed high accuracy of the ML-based framework.