Learning tenant behavior and evolutionary approach for demand response in colocation datacenters

This article addresses a relevant problem for super-computer and colocation datacenters that participate in energy-efficiency demand response programs. A combination of computational intelligence methods (artificial neural networks and evolutionary computation) is applied for computing accurate plannings to attend demand response events, while minimizing cost and quality of service degradation. Artificial neural networks are applied to predict the behavior of tenants within the demand response program and an evolutionary algorithm is applied to solve the underlying control/planning problem, considering the (fuzzy) neural network prediction as a surrogate of the unknown real function that models the behavior of tenants. Results computed for a case study modeling a realistic datacenter operation during a demand response event indicate that the proposed approach is able to compute accurate plannings, improving up to 71% over a Business as Usual method.

I. Introduction

The new generation of large computing infrastructures, including datacenters, supercomputing facilities, Distributed Content Networks, etc., requires smart management models that improve upon traditional approaches [1]. Due to the large dimension of the provided infrastructure, the large number of actors (tenants, clients, etc.) that participate in the paradigm, human experts are not able to handle the complexities of all the information to take advantage for proper decision making. This is also the case for datacenters and supercomputing facilities that participate in the electricity market by providing demand response and ancillary services [2], [3].