Optimal shifting of peak load in smart buildings using multiagent deep clustering reinforcement learning in multi-tank chilled water systems

In an effort to achieve energy efficiency, cost savings, and reduce carbon footprint, researchers are exploring ways to strategically redistribute energy use in cooling systems, particularly by flattening peak demand. This study proposes a novel approach using multiagent deep clustering reinforcement learning (MADCRL) to optimize load-shifting within multi-tank chilled water (MTCW) systems. Unlike traditional single-tank configurations, MTCW utilizes natural temperature stratification to store and distribute cool water, eliminating the need for physical barriers. The core of this approach lies in the MADCRL policy, which intelligently sequences chiller operations by taking advantage of cooler nighttime temperatures. By running chillers at full capacity while simultaneously charging the MTCW tanks, the system achieves substantial energy efficiency improvements. This dynamic control strategy ensures sufficient cooling capacity during peak periods while minimizing energy consumption during low-demand times. The effectiveness of this approach is validated through a complex MTCW plant simulation, highlighting the optimization of the value function by MADCRL, leading to improved decision-making and overall system efficiency. Implementing this reinforcement learning approach leads to significant energy savings and reduced power consumption. Compared to conventional methods using PID controllers, which struggle to shift loads efficiently, MADCRL achieves a superior coefficient of performance (COP), potentially exceeding energy savings by up to 25.47 %. This research demonstrates the promising potential of MADCRL in optimizing energy consumption within HVAC systems and beyond, offering substantial benefits in terms of energy efficiency and cost savings.