WorksEngineering
AC Assessment of DC Power Flow Screening for Data Center Hosting Capacity and Siting
Abstract
A screening procedure based on direct current (DC) power flow estimates the data center hosting capacity of a candidate bus, termed the DC-screened capacity, and selects the generator commitment that supports it. This paper assesses that estimate under the alternating current (AC) network model. With the selected commitment fixed, an AC optimal power flow reoptimizes dispatch and voltages to find the largest data center load the bus can carry, termed the AC-assessed capacity. On a synthetic Texas system with 123 buses, the DC-screened capacity exceeds the AC-assessed capacity at 113 buses. Across all 123, the ratio of DC-screened to AC-assessed capacity has a median of 143% (a 43% overestimate) and a maximum of 1,070%. At these 113 buses, the DC-screened capacity violates a branch apparent power limit or a generator reactive power limit in the AC model, two limits the DC model omits. The overestimate changes decisions: a 1-GW project passes DC screening at every bus but exceeds the AC-assessed capacity at 37, and the bus ranked first by DC screening has 205.5 MW less AC-assessed capacity than the bus ranked second. Reporting the AC-assessed capacity beside each DC-screened capacity lets planners reject oversized projects or choose better sites.
Introduction
Data centers consumed 4% to 5% of United States electricity in 2024 and are projected to consume 9% to 17% by 2030 [1]. The large load interconnection queue of the Electric Reliability Council of Texas (ERCOT) held requests totaling approximately 410 GW as of March 2026, about 87% of them from data centers [2]. Screening a queue of that size means solving unit commitment and expansion planning models at scale, and those models rely on the direct current (DC) power flow approximation, which is linear in the voltage angles and scales to thousands of buses and hours [3]. The approximation gains that tractability by dropping losses, holding voltage magnitudes fixed, and omitting reactive power [4]; a dispatch that is optimal under it need not be feasible under the alternating current (AC) network model [5].
For data centers, these omitted quantities matter. Emerging large loads include complexes of several gigawatts connected at the transmission level; in data centers, the power electronic drives of the cooling systems can increase reactive power demand, and large loads in general can affect voltage response and stability [6]. A 2026 alert from the North American Electric Reliability Corporation (NERC) calls for the expected power factor of each computational load and for system studies that may include active power versus voltage (P-V) and active versus reactive power (P-Q) analysis [7]. These studies rest on reactive power and voltage, the two quantities the DC approximation excludes.
Many studies of data center integration at the transmission level are built on the DC model. Under it, spatial load flexibility has been valued for congestion relief and curtailment reduction [8], spatial and temporal flexibility have been coordinated within security-constrained unit commitment [9], and the sensitivity of hyperscale hosting capacity to security modeling fidelity has been examined [10]. Two related lines of work point at what such DC studies leave out: system-level analyses show that large loads interconnect faster when they can be curtailed [11], whereas a data center sited as a firm load cannot rely on that relief; and at the distribution level, an AC branch flow model finds edge data center hosting capacity bounded by branch current and voltage limits that vary strongly with location [12]. For a candidate bus, DC screening estimates the hosting capacity, the largest constant data center load the bus can accept, and selects the generator commitment that supports it.
Two established approaches narrow the gap between the DC and AC network models. A data-driven enhanced linearized AC model keeps voltage magnitude and reactive power in a linear formulation and reduces its error through regression [13]. Alternatively, the unit commitment problem can include AC optimal power flow (ACOPF) constraints directly [14]. Both approaches improve how a plan is chosen. This paper takes the plan as given and assesses it under the AC network model. Holding the generator unit commitment selected by DC screening fixed, the AC assessment determines how much firm, constant data center load a transmission bus can carry within generator reactive power, bus voltage magnitude, and branch apparent power limits.
For each candidate bus , let denote the DC-screened capacity and the AC-assessed capacity under the same commitment. Comparing the two shows how far the DC-screened capacity departs from what the AC network can carry, and whether the screened project size and site remain valid. The contributions are threefold.
An assessment procedure that fixes the generator unit commitment selected by DC screening and reoptimizes dispatch and voltages with ACOPF to find the largest
data center load each candidate bus can carry under the AC network model.
A diagnosis of the AC limits violated at the DC-screened capacity, obtained with penalized slack variables. At every candidate whose DC-screened capacity is not AC feasible, the diagnosis identifies a branch apparent power limit or a generator reactive power limit, neither of which is represented in the DC formulation.
An evaluation of the DC-screened plans at all 123 buses of the synthetic Texas backbone transmission system (TX-123BT) [15], over the 24 hours of its peak day.
Assessing a DC-Screened Plan
Plan and Load Definition
Let be the set of buses, the set of branches, and the set of hourly periods in one day. All active and reactive power variables and limits in the formulations below are expressed in per unit on a common system base; capacity values are converted to MW for reporting.
Let denote the added active demand of the data center. If the data center is located at bus , its constant active and reactive demands at bus and hour are
where the superscript c denotes the added data center load, equals one for and zero otherwise, and pf is the lagging power factor of the data center.
Hosting capacity is the largest feasible . DC screening maximizes and returns the generator unit commitment that supports it; the AC assessment fixes that commitment and maximizes again, now subject to generator reactive power, bus voltage magnitude, and branch apparent power limits. Figure 1 shows the complete procedure at one candidate bus.

Figure 1. Assessment procedure for a DC-screened plan, repeated at each of the 123 candidate buses.
DC screening maximizes in a unit commitment model. Its variables include the active output of each conventional or renewable generator , the commitment status and reserve of each conventional generator, the bus voltage angle , and the active branch flow . Renewable output is bounded by its hourly availability. Commitment transitions, minimum up and down times, ramping, and generation limits are enforced using standard mixed integer linear constraints [16], [17]. In each hour, the reserve held by the other units must cover the output of every online unit, implementing the largest unit reserve rule. For every bus , branch , and hour , the DC network constraints are
Here, is the set of conventional and renewable generators at bus , is the native active demand, is the bus branch incidence matrix with outgoing orientation, is the branch reactance, and is the hourly active flow limit. The factor increases native and added active demand by two percent to approximate losses omitted by the lossless DC model.
The optimal value of is the DC-screened capacity . Because multiple commitments can attain the same , a second solve fixes and selects the commitment with the lowest daily operating cost, including fuel, no-load, and startup costs of the conventional units. The resulting commitment is denoted by , with elements . The pair is the DC-screened plan for bus .
AC-Assessed Capacity Under the Selected Commitment
The AC assessment fixes the generator commitment selected by DC screening and solves an ACOPF problem that reoptimizes every continuous variable. Let be the complex per unit voltage at bus , the entries of the per unit bus admittance matrix, the reactive output of generator , the native reactive demand, and the imaginary unit; an overbar denotes the complex conjugate. For every bus and hour , the AC nodal balance is
Conventional generators remain subject to the selected commitment and their active and reactive power limits, including , where and are the minimum and maximum reactive output of generator . Renewable generation follows its hourly availability and reactive power limits. Voltage magnitudes, branch angle differences, and apparent power at both ends of each branch are bounded, while the ramping and reserve requirements of DC screening are retained. The AC assessment maximizes the added data center load subject to these constraints. The optimal value of is the AC-assessed capacity , the largest load the selected commitment supports under the AC network model.
Where the DC-Screened Capacity Fails
An infeasible AC solve at does not reveal which limit is violated. To find it, a second ACOPF problem fixes under the same commitment and relaxes three limits with nonnegative slacks: voltage magnitude, generator reactive power, and branch apparent power. Each bus also receives an active power shortfall. All other constraints remain exact, including the reactive nodal balances, active generation limits, ramping limits, and reserve requirements.
The problem minimizes the weighted sum of these slacks. If every slack is zero, is AC feasible. Otherwise, the largest slack of each type gives the violation of that limit and shows where the plan fails.
How the Gap Changes a Decision
The capacity overestimate of DC screening at bus is
the percentage by which the DC-screened capacity exceeds the AC-assessed capacity. For example, means that the DC-screened capacity is 150% of the AC-assessed capacity.
The two capacities also determine whether a project size is screened correctly. A project of size at bus is a screening error if
that is, if it passes DC screening but exceeds the AC-assessed capacity.
The AC-assessed capacities also compare sites. Let be the DC-preferred bus and any other candidate. Both and are assessed under their own DC-selected commitments, so a positive means that the alternative can carry more load than the bus that DC screening prefers.
Test System and Evaluation Design
The test system is the modified TX-123BT case [15]: 123 buses, 255 branches, 138 conventional and 154 renewable generators, with the wind, solar, and hourly line rating profiles of its time series and an ERCOT weather zone label on every bus. The eight zones are Coast (COAST), East (EAST), Far West (FWEST), North (NORTH), North Central (NCENT), South (SOUTH), South Central (SCENT), and West (WEST). All 123 buses are evaluated over the 24 hours of the case’s peak day.
tbl:I lists the conditions shared by all assessments. Hourly line ratings apply as MW limits in DC and as numerically equal MVA limits at both ends of each branch in AC.
Computations ran on the Carya cluster at the University of Houston, using Intel Xeon Gold 5320 CPUs, with candidates solved in parallel; the study took 59.5 h of aggregate compute time, 7.2 h for DC screening of the 123 candidates and 52.3 h for the AC assessments. Gurobi 13.0.2 solves both the capacity maximization and cost minimization stages of DC screening using a relative mixed integer optimality gap tolerance of , with optimal status reported at every candidate bus. PowerModels.jl on JuMP formulates the ACOPF problems, which Ipopt solves. In the diagnosis problem, all slacks are expressed in per unit on the 100 MVA system base, and the weighted objective assigns a weight of 100 to the voltage magnitude slacks and active power shortfalls and a weight of 1 to the generator reactive power and branch apparent power slacks.
| Quantity | Setting |
| Data center load | Constant , at the candidate bus |
| Data center power factor | 0.98 lagging |
| Voltage magnitude bounds | 0.94 to 1.06 per unit |
| DC loss allowance | 2% of native and added load |
| AC commitment | Candidate’s DC-selected commitment |
Table I. Common Conditions for All Candidate Assessments
Results: Capacity, Project Size, and Location
The Gap Between DC-Screened and AC-Assessed Capacity
The DC-screened capacity exceeds the AC-assessed capacity at 113 of the 123 candidates. Over all 123, the capacity overestimate has a median of 43.2%, so at the median site the DC-screened capacity is 143% of the AC-assessed capacity. Figure 2 shows the full distribution, and it has two parts. Across 110 candidates, rises without a break from 97% to 275%; the remaining 13 sit apart, between 410% and 1,070%, with nothing in between. Brownsville 2 (bus 72) closes the sequence at 1,070%, the largest value in the study.

Figure 2. DC-screened capacity as a percentage of the AC-assessed capacity at each of the 123 candidates, in ascending order. Values above 100%, found at 113 candidates, mean that DC screening overestimates the capacity. The solid line marks 100% and the dashed line the median, 143%; the vertical axis is logarithmic.
At 10 candidates, is below 100% and is negative: the AC-assessed capacity exceeds the DC-screened capacity. DC screening inflates every load by a fixed 2% to cover losses, and at these ten sites the actual AC losses are smaller than that, so the AC model has slightly more room than DC screening allowed.
What Breaks at the DC-Screened Capacity
The AC model, unlike the DC model, includes the reactive demand of the data center. By (1) with a 0.98 lagging power factor, a 1-GW data center adds 203.06 Mvar at its bus. The slack diagnosis fixes at in the AC model, with the commitment unchanged. Ten of the 123 plans pass this check, with every slack at zero.
The check fails at the other 113, and every one of the returned diagnostic solutions violates a branch apparent power limit (111 plans), a generator reactive production limit (100), or both; 29 also violate a reactive absorption limit and 33 a voltage limit (11 high only, 11 low only, 11 both). The violations are large. Among the 111 plans with a branch apparent power violation, the largest violation of any branch in any hour is 126 MVA at the median plan and 1,003 MVA at the most; among the 100 with a reactive production violation, the largest excess of any generator in any hour is 528 Mvar at the median plan and 2,396 Mvar at the most, and at Plano 2 (bus 114) it is 923 Mvar. The active power shortfall, by contrast, stays below 0.001 MW at all 113 plans. Active generation therefore covers the DC-screened capacity in full; the returned solution exceeds limits because of the reactive power the load brings with it and the apparent power headroom it consumes on the branches that carry it. Neither limit exists in the DC model: it bounds active flow alone and cannot see the reactive component that pushes apparent power over the same rating, and it carries no generator reactive capability at all.
The Discrepancy Depends Strongly on Location
Figure 3 groups the 123 candidates by ERCOT weather zone. The gap is heterogeneous in two ways. Between zones, the median of moves from 116% in COAST to 216% in SOUTH; within zones the spread is just as wide, WEST alone spanning 98% to 449%. The 13 candidates with above 400% all lie in three zones: eight in SOUTH, four in SCENT, and one in WEST. Four of the ten candidates whose AC-assessed capacity exceeds the DC-screened capacity are in COAST. A single correction factor, global or zonal, therefore cannot replace an AC assessment at each bus.

Figure 3. The zonal median of ranges from 116% in COAST to 216% in SOUTH. The horizontal axis is logarithmic. Boxes span the interquartile range (IQR) with a median line, whiskers extend to the most extreme candidates within 1.5 IQR of the box, dots mark candidates, and parentheses give sample sizes.
Project Sizes That Pass DC Screening but Exceed the AC-Assessed Capacity
Suppose a site is to be chosen for a 1-GW data center. All 123 candidate buses have DC-screened capacities above 1 GW, so DC screening passes every one of them. At 37 of them, however, the AC-assessed capacity lies below 1 GW, so (5) holds and each is a screening error. The 37 are unevenly distributed: NCENT accounts for 20 and SOUTH for 8, while the 40 candidates of COAST, EAST, and WEST together have none.

Figure 4. Four of the 37 sites at which a 1-GW data center passes DC screening but exceeds the AC-assessed capacity. Bars give the DC-screened capacity and the AC-assessed capacity , labeled to the nearest MW. Each site is a separate case under its own DC-selected commitment.
The margins differ widely (Fig. 4). The AC-assessed capacity at Brownsville 2 is 146 MW, 854 MW below 1 GW; at San Antonio 25 (bus 47) it is 485 MW, 515 MW below; at Plano 2 it is 559 MW, 441 MW below; and at the closest case, Odessa 8 (bus 59), it is 973 MW, 27 MW below. There MW against MW, so : an overestimate of four percent is enough to reverse a project decision.
The DC-Preferred Location Is Not the Best Plan
DC screening prefers Pasadena 4 (bus 100) with MW, the largest DC-screened capacity in the study. Under that plan the AC-assessed capacity is MW. Calliham (bus 65), which DC screening ranks second with MW, has the largest AC-assessed capacity of all 123 candidates, MW (Figure 5), so

Figure 5. DC screening prefers Pasadena 4 on , while Calliham has 205.5 MW more AC-assessed capacity. Circles give and crosses at the two sites, and the dashed line carries across both. The axis is zoomed on the range that contains all four values.
Choosing the DC-preferred site therefore gives up 205.5 MW of AC-assessed capacity relative to an alternative that DC screening ranks lower.
Conclusion
This paper assessed DC-screened data center hosting capacity under the AC network model, holding fixed the generator commitment selected by DC screening. On the peak day of the TX-123BT system, DC screening overstates hosting capacity at 113 of the 123 candidate buses, and the median overestimate across all 123 is 43%. At each of these 113 buses, active generation covers the DC-screened capacity, and the diagnostic solution violates a branch apparent power limit or a generator reactive power limit, neither of which the DC model represents. The overestimate varies widely by location and changes decisions. A 1-GW project passes DC screening at every bus but exceeds the AC-assessed capacity at 37 of them; 28 of these are in NCENT and SOUTH, and none is among the 40 candidates of COAST, EAST, and WEST. The bus that DC screening ranks first has 205.5 MW less AC-assessed capacity than the bus it ranks second. Reporting the AC-assessed capacity beside each DC-screened capacity gives planners a basis to reject a project size or revise a site.
Acknowledgment
This work was supported in part by the University of Houston Energy Transition Institute Seed Grant Program and by computational resources from the Research Computing Data Core at the University of Houston.
References
- [1]T. F. Wilson, G. J. Blanford, J. Bistline, and N. Johnson, “Powering intelligence 2026: Updated scenarios of U.S. data center electricity use and power strategies,” Electric Power Research Institute (EPRI), Palo Alto, CA, USA, Tech. Rep. 3002034696, Feb. 2026. [Online]. Available: https://esca.epri.com/products/3002034696/
- [2]Electric Reliability Council of Texas, “ERCOT update,” Presentation to the Texas Senate Committee on Business and Commerce, Apr. 2026. [Online]. Available: https://www.ercot.com/files/docs/2026/04/01/ERCOT_LargeLoad_Update_April2026_B-C_-Hearing.pdf
- [3]B. Stott, J. Jardim, and O. Alsac, “DC power flow revisited,” IEEE Trans. Power Syst., vol. 24, no. 3, pp. 1290–1300, Aug. 2009.DOI
- [4]K. Purechala, L. Meeus, D. Van Dommelen, and R. Bermans, “Usefulness of dc power flow for active power flow analysis,” in IEEE Power Engineering Society General Meeting, 2005, 2005, pp. 454–459 Vol. 1.
- [5]K. Baker, “Solutions of DC OPF are never AC feasible,” in Proc. 12th ACM Int. Conf. Future Energy Syst. (e-Energy ’21), Virtual Event, Italy, Jun. 2021, pp. 264–268.arxiv.org/abs/1912.00319
- [6]NERC Large Loads Task Force, “Characteristics and risks of emerging large loads,” North American Electric Reliability Corporation, White Paper, Jul. 2025. [Online]. Available: https://www.nerc.com/globalassets/who-we-are/standing-committees/rstc/whitepaper-characteristics-and-risks-of-emerging-large-loads.pdf
- [7]North American Electric Reliability Corporation, “Essential action to industry: Computational load modeling, studies, instrumentation, commissioning, operations, protection, and control,” Level 3 NERC Alert, May 2026. [Online]. Available: https://www.nerc.com/globalassets/programs/bpsa/alerts/level-3-computational-load-alert.pdf
- [8]H. Wan, L. Fang, and X. Li, “Grid operational benefit analysis of data center spatial flexibility: Congestion relief, renewable energy curtailment reduction, and cost saving,” in Proc. IEEE Power & Energy Soc. Gen. Meeting (PESGM), Montreal, QC, Canada, Jul. 2026.
- [9]H. Wan and X. Li, “Data center spatio-temporal load flexibility in security-constrained unit commitment for enhanced grid efficiency and reliability,” in Proc. 2026 IEEE Industry Applications Society Annual Meeting (IAS), Vancouver, BC, Canada, Oct. 2026. [Online]. Available: https://arxiv.org/abs/2605.18517
- [10]H. Wan, L. Fang, J. Kwon, J. Silva-Rodriguez, J. Yang, J. Yellajosula, F. Jiang, and X. Li, “Impact of security modeling fidelity on hyperscale data center hosting capacity assessment,” in Proc. 2027 IEEE PES Grid Edge Technologies Conf. & Expo. (Grid Edge), Salt Lake City, UT, USA, Apr. 2027, accepted for publication. [Online]. Available: https://arxiv.org/abs/2610.18405
- [11]T. H. Norris, T. Profeta, D. Patiño-Echeverri, and A. Cowie-Haskell, “Rethinking load growth: Assessing the potential for integration of large flexible loads in US power systems,” Nicholas Institute for Energy, Environment & Sustainability, Duke University, Durham, NC, USA, Tech. Rep. NI R 25-01, Feb. 2025. [Online]. Available: https://nicholasinstitute.duke.edu/publications/rethinking-load-growth
- [12]L. Fang and X. Li, “Hosting capacity assessment and enhancement for edge data centers in active distribution networks,” in Proc. 2026 IEEE Industry Applications Society Annual Meeting (IAS), Vancouver, BC, Canada, Oct. 2026. [Online]. Available: https://arxiv.org/abs/2606.01407
- [13]X. Li, “Fast heuristic AC power flow analysis with data-driven enhanced linearized model,” Energies, vol. 13, no. 13, Jun. 2020, art. no. 3308.
- [14]A. Castillo, C. Laird, C. A. Silva-Monroy, J.-P. Watson, and R. P. O’Neill, “The unit commitment problem with AC optimal power flow constraints,” IEEE Trans. Power Syst., vol. 31, no. 6, pp. 4853–4866, Nov. 2016.
- [15]J. Lu, X. Li, H. Li, T. Chegini, C. Gamarra, Y. C. Yang, M. Cook, and G. Dillingham, “A synthetic Texas power system with time-series weather-dependent spatiotemporal profiles,” Sustain. Energy Grids Netw., vol. 43, Sep. 2025, art. no. 101774.
- [16]M. Carrión and J. M. Arroyo, “A computationally efficient mixed-integer linear formulation for the thermal unit commitment problem,” IEEE Trans. Power Syst., vol. 21, no. 3, pp. 1371–1378, Aug. 2006.
- [17]G. Morales-España, J. M. Latorre, and A. Ramos, “Tight and compact MILP formulation for the thermal unit commitment problem,” IEEE Trans. Power Syst., vol. 28, no. 4, pp. 4897–4908, Nov. 2013.