TL;DR
Data centers house massive battery arrays that usually sit idle, waiting for a power outage that rarely comes. This paper demonstrates that by jointly using these batteries for both "peak shaving" (lowering demand charges) and "frequency regulation" (selling flexibility back to the grid), data centers can reduce electricity bills by up to 30%—achieving a "superlinear gain" where the combined savings exceed the sum of the individual parts.
What Problem Does This Paper Address?
Data centers are power-hungry giants, with electricity bills often reaching millions of dollars annually. These bills typically consist of two parts: the total energy consumed and a "peak demand charge" based on the highest amount of power drawn at any single moment. To avoid these charges, data centers can use on-site batteries to "shave" the peaks. Additionally, emerging "regulation markets" pay large consumers to adjust their power usage in real-time to help the grid balance supply and demand.
The challenge is that data center batteries are designed for short-term backup; they have high power capacity but low energy capacity. This makes them poor at shaving long, "flat" peaks because they run out of juice before the peak ends. Previous research treated peak shaving and regulation as separate tasks. This paper asks: can we do both at once, and does the synergy between these tasks solve the energy-capacity limitation?
Methods
The authors develop a convex optimization framework that captures three competing factors:
- Electricity Costs: Minimizing both energy usage and the maximum power draw (peak).
- Regulation Revenue: Maximizing the "option fee" earned by bidding battery capacity into the PJM (Pennsylvania-New Jersey-Maryland) regulation market.
- Battery Health: Accounting for the degradation costs of lithium-ion batteries to ensure the strategy is actually profitable in the long run.
The core novelty lies in the joint optimization. Instead of dedicating a slice of the battery to each task, the framework allows the battery to switch fluidly between them. To understand why this works, the authors also introduce a Peak Abstraction Model, categorizing power spikes by height, width, and shape (triangular vs. rectangular).
Datasets
The study validates its framework using two primary real-world sources:
- Microsoft Data Center Traces: Six months of power demand data from geo-distributed Microsoft data centers, sampled at 20-second intervals.
- PJM Regulation Signals: Real-world "RegD" signals (fast-changing signals used for frequency regulation) from the PJM Interconnection.
- Synthetic Traces: Custom-built load profiles used to stress-test how different peak shapes (narrow vs. wide) affect the "superlinear" savings.
Key Insights and Ideas
- The "Superlinear" Aha Moment: The most striking finding is that $1 + 1 > 2$. The savings from joint optimization are greater than the sum of doing peak shaving and regulation separately.
- Regulation as a "Peak Breaker": Why does the synergy exist? Regulation signals are zero-mean and highly volatile. When a data center faces a long, flat peak that would normally drain a battery, the "randomness" of the regulation signal creates tiny valleys. These valleys allow the battery to briefly recharge or pause its discharge, effectively "breaking" one long, impossible peak into several smaller, manageable ones.
- Complementary Nature: Regulation requires fast, short bursts of power (high power density), while peak shaving requires sustained energy (high energy density). By combining them, the high-power capability of data center batteries is fully utilized to "fix" their low-energy limitation.
- Real-World Peaks are Ideal: The authors found that 90% of real data center peaks are "narrow" (lasting less than 2 minutes) and "triangular." This specific profile is perfectly suited for battery intervention.
Key Results
- Bill Reduction: Using real Microsoft traces, the joint strategy reduced total electricity costs by up to 30%.
- Frequency of Success: In a simulation of over 4,400 hours, the "superlinear gain" occurred in 93.4% of the hours.
- The "q" Factor: The authors defined a superlinear saving ratio ($q$). On average, joint optimization provided an extra 6.5% in savings beyond what could be achieved by simply adding the results of two separate optimization programs.
Limitations and Future Work
- Offline vs. Online: The current study uses "offline" optimization, meaning it assumes the data center knows exactly what the future power demand and regulation signals will be. In the real world, these must be predicted. The authors suggest that because regulation signals are zero-mean, a simple "look-ahead" controller should perform nearly as well.
- Battery Degradation Complexity: While the paper uses a linear model for battery wear, real-world lithium-ion degradation is non-linear and affected by temperature and depth of discharge. More sophisticated aging models could refine the cost-benefit analysis.
- Market Volatility: The study assumes a stable regulation market. In reality, the prices for regulation capacity ($\lambda_c$) fluctuate, which could change the optimal bidding strategy.
Summary updated