OCSRL: An Model-Based Reinforcement Learning Approach to Optimize Energy Consumption of Cooling Systems

Cooling systems consume more than 40% of overall buildings' energy consumption. These cooling systems provide comfortable and productive environments in building environments. The increase in global warming leads to the ever-increasing demand for cooling systems. There are two types of cooling systems — ducted-centralized and ductless-split. Optimizing energy consumption by ducted-centralized cooling systems has been the focus of researchers in the literature. However, very few have considered optimizing the energy consumed by the ductless-split. The solutions for ducted-centralized cooling systems control dampers inside ducts. In ductless-split cooling systems, there are no ducts. Hence, it is not possible to use a solution proposed for ducted-centralized to ductless-split. The solutions available for ductless-split cooling systems perform scheduling based on mathematical/machine learning models. In either case, since the solutions are static, they cannot adapt to the complex dynamics of the environment. In this paper, we propose OCSRL, a model-based reinforcement learning approach to optimize the energy consumption of ductless-split cooling systems. We evaluate the proposed OCSRL in both real-world experimentation and simulation. We compare OCSRL with the greedy and mathematical model-based approach Cooling-System Linear-Scheduling Algorithm (CLA) and All ACs ON (AA). In terms of the learning rate of the algorithm, energy consumption, temperature maintained, and execution time. OCSRL cuts the learning time by 97% compared to CLA in the simulation. Our proposed OCSRL cuts the learning time by 99% compared to CLA in real-world experiments. After learning, it immediately starts optimizing energy consumption. The proposed OCSRL saves up to 15% of energy compared to CLA in simulation and real-world experimentation while maintaining the desired temperature. OCSRL saves up to 50% of energy compared to AA. The proposed OCSRL reduces the execution time of ductless-split cooling systems in both simulation and experimentation by up to 9% and 65% compared to CLA and AA, respectively. This leads to a reduction in refrigerant consumption.