Dynamic Energy Storage Control for Reducing Electricity Cost in Data Centers

Abstract EN

As the scale of the data centers increases, electricitycost is becoming the fastest-growing element in their operationcosts.

In this paper, we investigate the electricity cost reductionopportunities utilizing energy storage facilities in data centersused as uninterrupted power supply units (UPS).

Its basic ideais to combine the temporal diversity of electricity price and theenergy storage to conceive a strategy for reducing the electricitycost.

The electricity cost minimization is formulated in theframework of finite state-action discounted cost Markov decisionprocess (MDP).

We apply Q-Learning algorithm to solve theMDP optimization problem and derive a dynamic energy storagecontrol strategy, which does not require any priori informationon the Markov process.

In order to address the slow-convergenceproblem of the Q-Learning based algorithm, we introduce aSpeedy Q-Learning algorithm.

We further discuss the offlineoptimization problem and obtain the optimal offline solution asthe lower bound on the performance of the online and learningtheoretic problem.

Finally, we evaluate the performance of theproposed scheme by using real workload traces and electricityprice data sets.

The experimental results show the effectivenessof the proposed scheme.