Minimization of Electricity Consumption in a Centralized Ventilation System

In times of growing concern over the environmental impact of high-energy physics organizations, this thesis, conducted within CERN's Cooling and Ventilation group, explores energy-saving strategies for heating, ventilation, and air conditioning (HVAC) systems. HVAC systems, essential for maintaining temperature, humidity, and air quality in residential and industrial settings, account for up to 40% of residential and 70% of industrial energy consumption. At CERN, these systems ensure proper air conditions in the accelerator complex. Together with general water-cooling systems, these systems account for up to 15% of total electricity consumption of CERN's flagship accelerator. Despite their energy intensity, these systems are typically managed by classical controllers, which are reliable but not optimal in terms of energy efficiency. This study aims to quantify the potential energy savings in HVAC systems using model predictive control (MPC), an advanced control strategy that incorporates system behaviour prediction and external data, such as weather forecasts. The methodology involves coding both classical and advanced controllers in a virtual environment, developing a digital twin model for a selected plant, and running simulations to confirm the improved thermal performance and electricity reduction with the MPC approach. To ensure a reproducible solution that can be easily adapted to different HVAC plants, the digital twin is built using neural networks, following recent advances in academic research. Optimization is performed using a genetic algorithm to minimize energy consumption while maintaining operational constraints, such as temperature limits. The comparative study shows significant savings of 77% achieved by MPC, far exceeding the 20% savings reported in the literature. While the results highlight the potential of advanced control strategies, the calculated savings are overly optimistic. Additionally, unexpected model behaviors in specific configurations provide valuable insights for future improvements, such as refining neural network architecture, enlarging the training data set and enhancing digital classical controller implementation.