Authors: Xingyu Zhao, Yingyi Tang, and Junbo Zhao
Affiliation: Dartmouth College
PSCAD version: 5.0.2
Last updated: October 2, 2026
Contact: xingyu.zhao.th@dartmouth.edu

This project provides a modular electromagnetic transient (EMT) simulation model of data center loads supplied by a centralized uninterruptible power supply (UPS). It is intended for power system studies of control interactions among power electronic modules and dynamic interactions between the data center and the grid.

The PSCAD library, DC_EMT_lib.pslx, includes power electronic converters, IT load profiles, control blocks, grid components, and a mechanical load model. Two example cases demonstrate how to assemble a complete data center model and compare converter models at different fidelity levels.

The supplied project files use PSCAD 5.0.2, as shown by both the Product Version and File Version in the project settings.

Figure 1. Power electronic modules and supporting components in DC_EMT_lib.pslx.

• AVM: Represents average converter behavior while retaining circuit and control dynamics.
• SWM: Represents switching devices and PWM operation, including switching ripple.
• ROM: Simplifies fast dynamics while retaining the outer controls and energy-storage dynamics represented by each module.

Parameters retained in the ROM use the same values as in the corresponding SWM.

• IT load profiles: Ramp, sinusoidal, and square-wave demand variations.
• Control blocks: PI controller, UPS supervisory control, and voltage ride-through (VRT) control.
• Grid components: Grid equivalent and voltage-dip profile.
• Mechanical load: Shaft-load model for motor-driven cooling loads.

The following table summarizes the control objectives and structures. AVM and SWM retain the explicit current loops where used; ROM replaces these fast dynamics with ideal current tracking or an algebraic current response while retaining the represented outer controls.

The VFD's V/F profile aims to maintain approximately constant motor flux in its constant-ratio operating region. For background, see Texas Instruments' Scalar (V/f) Control of 3-Phase Induction Motors.

The PI-regulated loops use a second-order target to relate response speed and damping to controller gains. Instead of selecting gains independently, specify a tuning frequency $f_{\mathrm{bw}}$ in hertz and a damping ratio $\zeta$.

For a first-order plant approximation and a parallel PI controller,

$$ G(s)=\frac{K}{as+b},\qquad C_{\mathrm{PI}}(s)=K_p+\frac{K_i}{s}, $$

where $a>0$ and $K>0$, negative unity feedback gives the characteristic polynomial

$$ a s^2+(b+K K_p)s+K K_i. $$

Matching this to $a(s^2+2\zeta\omega_n s+\omega_n^2)$ gives

$$ \omega_n=2\pi f_{\mathrm{bw}},\qquad K_p=\frac{2\zeta\omega_n a-b}{K},\qquad K_i=\frac{a\omega_n^2}{K}. $$

A higher tuning frequency gives faster target dynamics; the damping ratio sets the damping of the target poles. The coefficients $a$, $b$, and $K$ must represent the particular loop, including its per-unit scaling and operating point.

For example, a capacitor voltage loop with ideal inner current tracking has $a=C_{\mathrm{pu}}/\omega_b$, $b=0$, and $K=1$, when the controller commands capacitor-side current and load current is treated as a disturbance. With physical time in seconds and $\omega_b=2\pi f_{\mathrm{base}}$,

$$ K_p=\frac{2\zeta\omega_n C_{\mathrm{pu}}}{\omega_b},\qquad K_i=\frac{\omega_n^2 C_{\mathrm{pu}}}{\omega_b}. $$

Tune inner current loops first, then choose slower outer voltage loops so that the fast-current-loop approximation is reasonable. Verify the resulting response with the connected converter chain, including filters, switching delays, and limits.

Interpretation of bandwidth: The parameter labeled “bandwidth” sets the target natural frequency through $\omega_n=2\pi f_{\mathrm{bw}}$. It is not generally equal to the measured closed-loop −3 dB bandwidth: the PI zero and additional dynamics also affect the response. ROM uses the same tuning values as SWM for every retained controller; tuning inputs for removed loops are omitted.

This case demonstrates how to use the modules in DC_EMT_lib.pslx to build a complete data center model for power system dynamic studies. It uses average converter models and includes disturbances on both the grid and workload sides.

Figure 2. Example data center model with UPS-supplied IT loads, VFD-driven cooling loads, and static lighting loads.

The simulation sequence is as follows:

1. Initialization (t = 0–5 s): The IT load is initialized to 0.6 pu. The cooling load, represented by an induction motor, accelerates toward steady operation under the commanded V/F profile.

2. Grid-side disturbance (t = 5 s): A voltage dip is applied at 5 s. In the result shown below, the voltage begins to recover at approximately 6 s. This event illustrates the responses of the data center components, particularly the UPS, to the voltage dip and recovery according to the implemented voltage ride-through logic.

3. Workload-side disturbance (t = 12 s onward): A periodic square-wave variation is applied to the IT workload to represent power demand fluctuations associated with AI training tasks.

Figure 3. Case 1 responses: active and reactive power at the PCC, UPS input, cooling loads, and static loads, together with RMS supply voltage and UPS operating mode.

The UPS input active power falls close to zero during battery operation and ramps back up after the return to online operation. The subsequent workload changes appear in both the UPS input power and the total power at the PCC. The bottom panel shows the supply voltage and the UPS mode (0: online; 1: battery).

This case compares the responses of the power electronic modules in DC_EMT_lib.pslx at different modeling fidelity levels.

Figure 4. Case 2 configuration comparing AVM, SWM, and ROM representations of IT loads and AVM and SWM representations of cooling loads.

The average, switching, and reduced-order models are compared under the following simulation sequence:

1. Initialization (t = 0–2 s): The IT load is initialized to 0.6 pu.

2. Grid-side disturbance (t = 2 s): A voltage dip is applied at 2 s, and the voltage begins to recover at 2.5 s.

3. Workload-side disturbance (t = 6 s onward): A periodic square-wave variation is applied to the IT workload to represent power demand fluctuations associated with AI training tasks.

These disturbances allow the three model variants to be compared under both grid-side and workload-side changes.

The cooling-load model is initialized during the first 5 s. At 5 s, the voltage-dip profile used for the IT-load comparison is applied to the cooling-load supply. This test compares the responses of the average and switching models to a grid-side disturbance.

Figure 5. Case 2 responses. The upper three panels compare supply voltage and UPS input active and reactive power for SWM (green), AVM (blue), and ROM (red). The lower three panels compare supply voltage and cooling-load active and reactive power for SWM (green) and AVM (blue).

The IT-load models follow similar UPS input power trajectories in this example, with differences during transients and in steady operation. The cooling-load comparison also shows differences in ripple and power levels between AVM and SWM.

Together, these comparisons illustrate how modeling fidelity affects simulated dynamic responses and support model selection for different study objectives.

If you use this model in your research, please cite the software repository:

X. Zhao, Y. Tang, and J. Zhao, Generic EMT Modeling of Data Center Load v0 (DC_EMTv0), version v0, Dartmouth College, 2026. [Software]. Available: GitHub repository.

@misc{zhao2026dcemt,
  author       = {Zhao, Xingyu and Tang, Yingyi and Zhao, Junbo},
  title        = {Generic {EMT} Modeling of Data Center Load v0 ({DC\_EMTv0})},
  year         = {2026},
  howpublished = {GitHub repository},
  url          = {https://github.com/xingyuz-phd/Generic-EMT-Modeling-of-Data-Center-Load},
  note         = {Version v0, Dartmouth College}
}

For questions about the model, please contact Xingyu Zhao at xingyu.zhao.th@dartmouth.edu.