Collective learning of lithium-ion aging model parameters for battery health-conscious demand response in datacenters

This paper examines the degree to which a large-scale datacenter employing lithium-ion (Li-ion) batteries for demand response can learn the physics-based aging and degradation dynamics of the underlying batteries by measuring their input/output current/voltage data. Battery degradation dynamics are chemistry dependent and change significantly for newer chemistries. Moreover, characterizing these degradation dynamics requires time-consuming and expensive laboratory testing. Together, these facts motivate the following question: is it possible to use battery current and voltage measurements to learn battery degradation behavior in a datacenter where numerous distributed batteries are being used for demand response? If so, what are the benefits and challenges associated with such learning? The goal of this paper is to provide preliminary answers to these questions building on earlier work by authors on health-conscious stochastic battery control. Specifically, we show that when datacenters exploit its demand management flexibility at the rack-level to control different batteries in accordance with different management policies, the resulting data is sufficiently rich to the point where (1) the learning of degradation behavior is possible within the span of approximately 1 year, (2) with a reasonable number of cells, (3) even when the batteries are used simultaneously for degradation learning and demand response.