THE DEVELOPMENT OF HVAC SYSTEM ENERGY MANAGEMENT BASED ON REINFORCEMENT LEARNING CONTROL USING ANALYSIS OF PREDICTED MEAN VOTE AND ENERGY EFFICIENCY UTILIZATION

• Main Author: Christian P. S, Hadi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/81476
• Institution: Institut Teknologi Bandung

Approximately 36% of the world's energy consumption and 38% of carbon emissions are attributed to the building sector. Indonesia has committed to achieving a 29% reduction in carbon emissions by 2030, focusing on energy conservation measures. Additionally, there is a specific target to achieve 23% utilization of renewable energy by 2025. The predominant energy consumption in building systems stems from HVAC, making it a key area of focus for researchers aiming to improve building energy efficiency while maintaining thermal comfort. Nevertheless, the pursuit of occupant thermal comfort with reduced energy usage poses a considerable challenge for researchers. Therefore, this research presents a new intelligent approach that has the ability to accurately control building HVAC systems based on thermal comfort using Predicted Mean Vote analysis and Energy Efficiency Utilization in buildings to reduce HVAC energy consumption. Currently, a lot of research has been carried out using Reinforcement Learning controllers, especially in HVAC research objects. The focus of this research is the development of a controller based on reinforcement learning, to reduce HVAC consumption in university buildings by minimizing energy use based on a temperature setpoint regulation system so that energy efficiency utilization is minimum without sacrificing thermal comfort. This research takes an approach based on modeling and simulation. The case study object used is the Labtek XIX SBM ITB Building, the Freeport Building, where the building was modeled using SketchUp, then redesigned in OpenStudio to define various parameters according to real building conditions. The modeled building is limited to 2 floors with 4 thermal zones for design simplicity. The HVAC system temperature setpoint in the model is set at a constant 24°C as the initial value. The building model that has been created is then validated by analyzing the power consumption patterns of the HVAC system from simulation results with measurement results from the Electrical Energy and Water System (SiElisA). After the building model is available, the Reinforcement Learning controller is designed in Python software. The controller algorithm chosen in this research is Proximal Policy Optimization (PPO) because it is stable for many building system cases, simple, policy-based and does not require large computing resources. However, in this research hyperparameter tuning has been carried out in the form of learning rate. The learning rate is set to the values 0.1, 0.01, 0.001, 0.0001, 0.00001 and 0.000001. From the second tuning, a controller with a learning rate of 0.001 was chosen because the reward at the end of the tuning was higher compared to other reward tunings. The controller that had been trained was then tested for a total of 4 weeks during which there were dry season, rainy season, equinox I and equinox II. The parameters evaluated for the 4 thermal zones include changes in temperature setpoint, changes in thermal comfort and cost savings. In this research, it was found that the controller was able to reduce building efficiency by 7.4% in the first equinox (March), 11.8% in the second equinox (September), 10.9% in the dry season (May) and 14% in the rainy season. (December). The evaluation results showed that the reduction in energy consumption for the case study object, Floor 2 of the Labtek XIX SBM ITB Building was 11.02% and was able to increase thermal comfort by 25% compared to without a controller.