Optimizing Energy Efficiency in Data Centers through Learning-Based Air Conditioner-Server Rack Mapping

Traditional approaches, reliant on static rules and manual oversight, are inadequate in managing complex modern data centers. This research introduces a novel framework leveraging unsupervised and semi-supervised learning to optimize cooling systems and improve energy efficiency. The framework employs constrained k-means++ clustering and ensemble learning techniques to analyze historical temperature data, automatically revealing device interdependencies and mapping air conditioners to server racks without layout diagrams. It utilizes a data-driven approach to efficiently model and detect airflow organization variations within data centers that impact the air conditioning-racks relationship. It provides insights that go beyond what can be deduced merely from layout diagrams. This reduces human involvement and establishes a robust foundation for cooling control. Semi-supervised learning within the framework allows for accurate temperature correlations and iterative mapping refinement. Essentially, the algorithm also optimizes server room alarms and decreases alarm rates simultaneously via its proactive correction mechanism, thus ensuring safety from a different angle. The framework demonstrates adaptability to diverse operational conditions. Experiments conducted in two separate data centers confirm that this adaptive strategy achieves significant enhancements in energy efficiency, responsiveness, and overall cooling management. Consequently, it lays down the groundwork for advanced air conditioning control strategies with considerable potential to boost energy efficiency in data centers.