A Novel Machine Learning-Based Load-Adaptive Power Supply System for Improved Energy Efficiency in Datacenters

Power Supplies are a key part of the modern Internet and Communications Technologies (ICT) industry. Modern Uninterruptible Power Supply (UPS) systems are modular and as such, consist of several Power Supply Units (PSUs). Even though various PSU designs are used to optimize operation efficiency at specific loading conditions they engender inefficient operation at other loading conditions. In order to optimize the energy efficiency in various loading conditions, this paper proposes a novel power supply multiplexing system engaging different combinations of PSUs which are controlled through machine learning techniques to maximize efficiency depending on the loading conditions. Each PSU combination is given a state number. Due to the vast number of combinations (states) that can occur in such systems and redundancy requirements, machine learning techniques are proposed. It is shown that by using the proposed novel system, an efficiency improvement of over 78% can be achieved in low loading conditions and an average 5.23% in all loading conditions.