Parag Mitra
The increased reliance on digital services, growth of big data analytics, and popularity of cloud-based services, among other things, have led to the recent boom in data centers. North America leads the world in the total number of installed data centers, and it is estimated that the total installed peak megawatt (MW) capacity is around 4,458 MW, with another 1,144 MW under construction according to the latest assessment by CBRE, a commercial real estate services and investment firm based in North America. Notably, the Silicon Valley (California), Northern Virginia, Atlanta (Georgia), and Hillsboro (Oregon) regions in the United States are seeing significant new construction of data centers.
Data centers are unique loads in the sense that these can be significantly large single-point loads that are mostly composed of electronic components, unlike other industrial loads, which are predominantly motor-type loads. Data centers are highly energy-dense loads, which means that these installations have a very high energy consumption as compared to the square footage of the facility. Since data centers can be significantly large, especially the hyperscale data centers (in the hundreds of MW), planning for these loads is becoming a challenge that many transmission planners are faced with. Understanding how these loads behave is critical for transmission planners to ensure that the grid can support these loads without resulting in any reliability issues.
According to a 2009 report by the Lawrence Berkeley National Laboratory, depending upon their primary usage, data centers can be categorized as flat-use data centers or mixed-use data centers. Flat-use data centers typically house only the servers that may be running particular services and have a flat and uniform daily as well as seasonal load consumption pattern. Mixed-use data centers, on the other hand, have an on-site facility for staff, and hence, their load shape is similar to that of a flat-use data center with a load shape of a typical office building superimposed on it. Understanding the type of data center being built is therefore important for transmission planners as that influences the MW load that planners have to consider in their transmission planning studies.
Another aspect that transmission planners need to be mindful of is the stability implications of having such large concentrated electronic loads on the system. Traditionally, much of the electronic load in a power grid was composed of consumer electronics that were fairly distributed in the system. As such, only a smaller number of such devices contributed to the grid dynamics in the event of a disturbance, and modeling these as simple constant power loads was considered sufficient. However, data centers are challenging that way of thinking. Data centers typically have two major types of loads: switch mode power supplies connecting to the server racks and air conditioning systems required to control the temperature within the facility. Depending on the size of the data center, the power consumed by the servers can comprise approximately 50% (for the kW scale) to 80% (for the MW scale) of the total load inside the facility, according to a 2012 white paper by the Green Grid. Additionally, these facilities are also equipped with power conditioning equipment like online uninterrupted power supplies, local backup generators, and in some cases, on-site photovoltaic generation, which can result in complex (and sometimes undesirable) dynamic responses at the point of interconnection during grid disturbances.
Notably, the ride-through behavior of data center loads has become quite a concern for many utility planners and operators. One of the main challenges that utility planners are concerned with is the disconnection of large data centers from the grid during momentary shallow voltage sags (caused by remote faults) due to the action of the facility protection systems or over-sensitive power conditioning equipment. While this may be desirable for data centers close to the fault as it helps with voltage recovery, the disconnection of large loads farther away from the affected portion of the grid followed by delayed reconnection can cause frequency excursions, especially during periods of low overall system load. This particular issue is still being investigated, and utilities are conducting a variety of simulation studies to gauge the risk that this poses to system stability. The power system community is trying to identify if certain “grid-friendly” ride-through behavior can be made a part of a large load connection request that can address this concern. As an example, see ENTSO-E’s demand connection code, which has placed some requirements on loads connecting to the transmission system to induce grid-friendly behavior.
Current and voltage oscillation in the subsynchronous range from data centers is another active issue that a lot of power system engineers are trying to understand. In 2017, researchers from Meta (formerly Facebook) and Rensselaer Polytechnic Institute presented a paper documenting oscillation caused by a Meta data center in the 10–20-Hz range. Oscillation in this frequency range can be detrimental to the power grid. This has led to some concerns regarding the operation of large data centers in close proximity to large inverter-based resources as some of these oscillatory issues can result in subsynchronous control interactions between the plant and the loads. These issues, however, can be addressed by the proper tuning of controls within either or both facilities if the actual cause of such interactions can be pinpointed.
Data centers are a new breed of loads that will be of a sizable MW amount in many footprints. As such, understanding the quasi-static (time-series) as well as the dynamic nature of the load will be extremely important to ensure that such loads can be served reliably without exposing the system to reliability challenges. Modeling the load properly in planning studies plays a key role and further highlights the importance of load modeling as well. However, to ensure that all of this can be achieved, a platform for having a dialog between transmission planners and data center owners and operators needs to be created. Many transmission planners are struggling to understand the very nature of these loads to model them appropriately for screening assessments. Open communication allows transmission planners to understand and model these loads better and take advantage of the flexibility that these loads might offer. A collaborative approach is the best way forward to plan for a secure and reliable grid.
“North America data center trends H2 2022,” CBRE, Dallas, TX, USA, 2023. [Online] . Available: https://www.cbre.com/insights/reports/north-america-data-center-trends-h2-2022
G. Ghatikar et al., “Demand response and open automated demand response opportunities for data centers,” Lawrence Berkeley National Lab., Berkeley, CA, USA, Tech. Rep. LBNL-3047E, 2009. [Online] . Available: https://www.osti.gov/servlets/purl/981725
“PUE™: A comprehensive examination of the metric,” The Green Grid, Washington, DC, USA, White Paper #49, 2012. [Online] . Available: https://datacenters.lbl.gov/sites/default/files/WP49-PUE%20A%20Comprehensive%20Examination%20of%20the%20Metric_v6.pdf
“Demand connection code,” ENTSO-E, Brussels, Belgium, 2016. [Online] . Available: https://www.entsoe.eu/network_codes/dcc/
J. Sun et al., “Modeling and analysis of data center power system stability by impedance methods,” in Proc. IEEE Energy Convers. Congr. Expo. (ECCE) Conf., 2019, pp. 107–116, doi: 10.1109/ECCE.2019.8913185.
Parag Mitra (pmitra@epri.com) is with EPRI, Palo Alto, CA 94304 USA.
Digital Object Identifier 10.1109/MELE.2023.3291509
2325-5897/23©2023IEEE