Flattening power curves in smart buildings: Deep reinforcement learning for enhancing chiller performance and binary phase change materials

Reducing energy consumption and offsetting peak loads have become crucial for designing smart buildings. The main aim of this study is to balance energy demand and lower the cost per kWh by combining machine learning with phase change materials. Additionally, this approach aims to improve thermal energy storage efficiency and increase the coefficient of performance (COP) for chillers, ultimately easing the load on overloaded power grids. To meet these goals, an innovative combination of phase change material (PCM) integration with advanced machine learning techniques was employed to address key challenges in energy demand and grid efficiency. By dividing a binary mixture of tetradecane and hexadecane into four tanks, the system enhances heat transfer and facilitates efficient thermal energy storage. Its real innovation lies in the use of cooperative multi-agent reinforcement learning with deep clustering (CMARLDC), which intelligently sequences tank operations and maximises chiller COP during off-peak periods. The application of Lagrangian trajectory curve clustering to handle nonlinear regression adds a unique technical dimension to the control strategy. The outcome results in saving over 28 % more energy compared to traditional structures and features a highly compact system. This reduces PCM tank size to just 16.8 % of traditional designs, while significantly lowering energy costs, easing grid stress, and supporting sustainable, resilient cooling infrastructure.