Electricity Price-Aware Scheduling of Data Center Cooling

TL;DR

This paper introduces a Markov Decision Process-based model for electricity price-aware cooling scheduling in data centers, aiming to reduce cooling costs by adapting to price fluctuations and seasonal patterns.

Contribution

The paper presents a novel MDP framework combined with Quantile Fourier Regression for dynamic cooling scheduling based on electricity prices and temperature data.

Findings

Model achieves lower cooling costs compared to heuristics

Simulated over 14 years of historical data

Effective for demand response and investment decisions

Abstract

Data centers are becoming a major consumer of electricity on the grid, with cooling accounting for about 40\% of that energy. As electricity prices vary throughout the day and year, there is a need for cooling strategies that adapt to these fluctuations to reduce data center cooling costs. In this paper, we present a model for electricity price-aware cooling scheduling using a Markov Decision Process(MDP) framework to reliably estimate the cooling system operational costs and facilitate investment-phase decision-making. We utilize Quantile Fourier Regression (QFR) fits to classify electricity prices into different regimes while capturing both daily and seasonal patterns. We simulate 14 years of operation using historical electricity price and outdoor temperature data, and compare our model against heuristic baselines. The results demonstrate that our approach consistently achieves lower cooling costs. This model is useful for grid operators interested in demand response programs and data center investors looking to make investment decisions.