The development of cloud computing and AI workloads as well as digital services of large scale has advanced extremely fast. As a result, the energy consumption of new data centers has escalated dramatically. Cooling systems represent a very big fraction of the total cost of running a data center. Hence, conventional rule, based cooling approaches, that mainly cannot adapt to workload changes effectively, besides causing energy wastage, also temperature control problems as their side effects. This work proposes a Reinforcement Learning (RL) based cooling optimisation framework, which implements the Proximal Policy Optimisation (PPO) algorithm to generate smart and adaptive cooling management. The RL agent is trained in a simulated data center environment where the server temperature and energy consumption are mainly used as feedback. A very careful reward function tries to maintain the temperature close to the ideal operating range (24, 26 Degree Celsius) as much as possible while, simultaneously, the reward function also discourages energy overuse. So, a balance between thermal comfort and energy saving is attained. The test confirmed that the method described above not only helps to stabilize the temperature very precisely, but at the same time, it is more energy, efficient than the traditional cooling methods. As well as maintaining the system's performance under continuously fluctuating workloads, the framework also stabilizes the cooling actions even when the thermal load rises very quickly. The modular design of the system architecture offers the possibility of integrating live sensor data and building management systems in the future. The PPO algorithm ensures that the policy updates are done in a stable manner, and it prevents the cooling changes from being too sudden, which is very important for facilities that need to be safe, such as data centers. Long term reward convergence plots indicate a stable learning process throughout extended training duration. The study emphasizes the necessity for a good trade, off between exploration and exploitation to obtain the best cooling strategies. The framework, by being flexible to changes in cooling capacities and thermal dynamics, thus allows the scalable installation of heterogeneous data center infrastructures. Adding IoT, live monitoring, and predictive workload, aware cooling control can further enhance the system. This research demonstrates that using artificial intelligence for thermal optimization is feasible and it helps in making the next generation data center (DC) infrastructure autonomous, sustainable, and energy, efficient.
S et al. (Tue,) studied this question.