Data mining for modeling chiller systems in data centers

Abstract. We present a data mining approach to model the cooling infrastructure in data centers, particularly the chiller ensemble. These infrastructures are poorly understood due to the lack of “first principles ” models of chiller systems. At the same time, they abound in data due to instrumentation by modern sensor networks. We present a multi-level framework to transduce sensor streams into an actionable dynamic Bayesian network model of the system. This network is then used to explain observed system transitions and aid in diagnostics and prediction. We showcase experimental results using a HP data center in Bangalore, India.