Battery Energy Storage in Data Centres: A Review of Sizing Methods, Control Strategies and Grid Services

Data centres built for artificial intelligence (AI) are becoming large and fast-changing electrical loads. Battery storage is already installed in these facilities for backup, and it is now also expected to smooth load swings and to serve the grid. This review examines battery storage at data centres, from the server board to the point of interconnection. It covers load characteristics, the storage layers, sizing methods, control strategies, and grid services and standards. The evidence base comprises 322 studies, selected through a structured screening process guided by PRISMA 2020. Reported load behaviour was converted into a common storage duty. Smoothing the 0.5–1.5 Hz training cycle needs only about 0.1–0.5 s of energy per unit swing, but tens of thousands of shallow cycles per day. Storage at the board, rack, facility uninterruptible power supply (UPS) and grid connection serves complementary rather than interchangeable functions. When normalised to facility rating, reported storage sizes differ by a factor of about 24 in energy and 12 in power, and this spread follows the assumed requirement more closely than the sizing method. Sizing studies often conflate local smoothing with shaping of the grid-side profile. The facility UPS battery appears to be the most promising under-used network resource, but its grid role rests on feasibility studies, simulation and one reported pilot. Five research needs are identified, led by experimental validation of grid-forming facility storage with a critical information technology (IT) load connected.