Dynamic Control of HVAC Attributes. Improving Energy Efficiency of HVAC System Using Machine Learning and Computational Fluid Dynamics

Abstract

Heating, Ventilation and Air-Conditioning (HVAC) accounts for a major share of building energy use. This thesis develops a data-driven HVAC control framework that couples high-fidelity Computational Fluid Dynamics (CFD) with machine learning. Three-dimensional CFD simulation is performed to model the flow field in the room using Star-CCM+. 973 steady state CFD simulations representing realistic boundary conditions were performed to form a comprehensive dataset of temperature and velocity fields. Fourier Neural Operator (FNO) was trained as a surrogate model to CFD. This model reproduces the temperature and velocity flow fields with adequate resolution, with over 90% of predicted temperature and velocity fields across unseen samples deviating within a 5% error, making it suitable for closed-loop use. The error distribution analysis shows that the median temperature error is at 0.0169 °C and median velocity error is 0.0034 m/s, indicating that the surrogate model is reliable to predict temperature and velocity fields. The surrogate model coupled with a Soft Actor-Critic (SAC) controller, which was designed to regulate inlet air temperature, inlet mass-flow rate, and radiator surface temperature to maximize the reward, that provides an optimal control solution to reduce the energy cost. The controller was evaluated for a year of weather data with a 20-minute control step and benchmarked against a PID controller. Results show that the SAC consumed 4,836 kWh compared to 7,460 kWh for PID, which corresponds to approximately 35% in energy cost reduction. SAC occasionally produces larger deviations than PID, leading to a higher median temperature error (0.487 °C vs 0.273 °C), but fluctuations beyond ±2.5 °C occurred only 2.3% of the time, indicating that comfort violations remained rare. Seasonal analysis shows SAC controller’s energy savings persist across the year and strengthen in the late-year window (36% vs 34% earlier), reflecting adaptive use of outdoor conditions and smoother control. Overall, the work demonstrates that combining CFD data trained surrogate model with entropy-regularized reinforcement learning can deliver substantial energy savings with acceptable comfort tracking, and provides a practical route to incorporate detailed physics (e.g. radiative gains, occupancy, humidity/CO2) and more advanced control designs in future studies.