Robust Event-Driven Dynamic Simulation using Power

Flow
Aayushya Agarwal, Amritanshu Pandey, Larry Pileggi
Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, USA
{aayushya, amritanp, pileggi}@andrew.cmu.edu

Abstract— The power grid is a dynamic system encompassing incompatible with each other, that studies the behavior of the
numerous control devices with varying degrees of time constants, system in the vicinity of a specific time scale, ranging from
posing a challenge for dynamic simulators to efficiently step steady state in power flow to short term dynamics using
through time. We introduce a robust event-driven simulator to dynamic simulation. One such example is the suite of DSA
mimic dynamic simulation that exploits the inherent temporal tools [3] that offer simulation software to study stability in a
sparsity by solving for the quasi-steady state of the grid using range of different time scales. However, one key missing
frequency-dependent power flow at each time-step. In this component is studying the long-term dynamics (in tens of
methodology, the effect of fast transient control actions is
minutes or hours) and the effect on steady state of large
captured within the quasi-steady state while a time-dependent
systems. Contingency planning, black starts and online
outer loop handles slow transients that dynamically change over
time. The event-driven approach offers a fast simulation
operations are just some of the applications that benefit from
framework capable of scaling to large systems by taking time- studying this behavior [3]. In these uses-cases, an engineer
steps based on proceeding events and uses a robust power flow requires a tool to understand the effect of actions that interact
simulator, Simulation with Unified Grid Analyses and with automatic controllers over the course of minutes to hours.
Renewables (SUGAR), to solve for the quasi-steady state. We The simulation tools available today are restricted by their
demonstrate the efficacy of this approach by simulating the intended time scale to be able to simulate such a problem [4].
effects of different control actions including automatic generation
control (AGC), automatic voltage (AVR) and over excitation
limits on large systems.
Index Terms-- dynamic simulation, event driven, grid device
control, quasi-steady state simulation, temporal dependency.

I. INTRODUCTION
Simulating the power grid is widely accepted as a necessity for
operation, planning, and restoration of the grid. However,
accurate time-domain simulations often are bottlenecked by
slow simulation times for simulating large systems in real-time
due to the large range of time scales attributed to the
disturbances and control actions occurring in the power grid. A
disturbance, such as an outage, propagates the disruption in the
form of electromagnetic or electromechanical transients, and
ultimately a change in the steady state through various controls
taking place. These disturbances have a time scale ranging from Figure 1: Time scales of events and control actions occurring in grid [1].
microseconds to hourly load changes as shown by Figure 1[1]. On one side of the spectrum, power flow is a conventional fast
The range of time scale is one of the major challenges in simulation tool, that approximates the steady state of a system
developing an accurate, fast simulator capable of scaling to and has proven to be scalable to large systems [5]-[6]. Advances
large systems (up to Eastern Interconnection size systems) in governor power flow have robustly improved the accuracy
while being computationally fast [2]. The grid is inherently a of the steady state by incorporating frequency deviation along
dynamic system constantly changing due to control actions with frequency control actions into the simulation [7]-[8].
each with a wide range of time constants. Tools combat the However, it still ignores the temporal dependencies of the
challenge of simulating this dynamic system by narrowing the control actions and applies all control action at once increasing
focus to a range of time scale as well as by often compromising the likelihood of incorrect steady-state results [8]. In reality,
speed for accuracy. The result is a distinct set of tools, often various control actions within the grid are constantly being

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
enacted with different time constants and demands and supply efficiency offers a potential fast and lightweight solution to
are constantly in flux. While an existing steady state simulation simulating the steady state changes due to transient actions.
offers insight into ideal isolated scenarios, it doesn’t offer a In this paper, we introduce an event-driven dynamic
complete picture of the dynamic system response [9]. simulation framework that builds upon a governor power flow,
Conversely, dynamic simulation offers the accuracy to SUGAR [5], as an approximation for the quasi steady-state.
simulate transients by solving electromechanical differential The framework incorporates real-time data into the governing
algebraic equations [9] at each time step. The bottleneck for fast power flow tool to provide accurate models that are capable of
yet accurate transient simulation to long-scale dynamics and solving for fast transients at a time splice. To dictate simulation
large systems is the large degree of time constants (ranging progression in time, the tool exploits the temporal sparsity by
from milliseconds to minutes) of control actions in the grid being driven by events rather than time steps, thereby
[10]-[14]. This forces dynamic simulation to take small time
achieving high levels of efficiency and scalability that can be
steps to respect the fast transients even during latent periods in
shown to simulate large systems of 80k+ buses. The novelty in
which the slower transients dominate. A similar issue is seen in
circuits community while simulating stiff circuits such as phase this work is the event-driven time-step control incorporated
locked loops [16]. During these latent periods, dynamic with an accurate quasi-steady state solver that allows fast
simulation is unable to take larger time steps that would avoid simulation speeds. The event-driven framework has its
the unnecessary computation and is required to compute over parallels to circuit-simulators SPECS and ACES [16] that were
the latent period. Along with the non-linearity of the differential highly efficient for simulating stiff circuits with time constants
equations, this prevents traditional transient simulation from that vary with orders of magnitude, similar to the power grid.
being able to solve long-term dynamics of large systems The rest of the paper is organized as follows. Section II
without sacrificing speed of simulation [13]. Previous works describes the quasi steady-state approximation that is solved
have dynamically adjusted the time-step during periods of using a governor power flow framework, SUGAR, as described
quiescence and have proven to aid in speed, but still fall short in Section IV. Sections V describes the algorithmic framework
in scaling to real life test systems [12]. To combat the issue, real for the event-driven platform followed by results in Section VI.
time digital simulators (RTDS) exploit the spatial and temporal
sparsity in parts of the grid to parallelize the transient simulation II. QUASI-STEADY STATE
onto multiple cores [14]. While RTDS does show The framework for event-driven dynamic analysis solves for a
improvements to scalability, it often requires large number of quasi-steady state (QSS) at each time step. The QSS
cores to be effective. Other works have exploited the matrix approximation considers a specific time interval that justify the
structure of dynamic simulation to readily parallelize across assumptions regarding the dynamic system, modeled as a set
multiple nodes and thereby achieve simulation efficiency [15]. of differential algebraic equations as shown below [17]:
While these transient simulations are necessary in
𝑥̇ = 𝑓(𝑧!,# , 𝑧!,$ , 𝑧!,#̇ , 𝑧!,$
̇ , 𝑧% , 𝑥) (1)
understanding oscillations and periods of fluctuation, the real
interest in studying a system often lies in the steady states 0 = 𝑔(𝑧! , 𝑧!,$ , 𝑧!,#̇ , 𝑧!,$
̇ , 𝑧% , 𝑥) (2)
achieved throughout the span of time. The system of equations is defined by a vector of state
To bridge the gap between these two formulations, previous
variables (𝑥) that represent the voltages at each capacitor node
works have developed a quasi-steady state model of the
transient equations, which study splices of time in which the or the current through an inductor [16]. The response of state
grid attains momentary steady state. Under the assumption of variables due to different control actions is a result of the
zero time-derivative for fast transients, quasi-steady state offers governing control system equations and network
a compromise on accuracy for efficiency for long-term analysis dependencies, which are characterized in (1) and (2),
[17]-[19]. While quasi-steady formulation is unable to simulate respectively. The control actions injected by a source into the
oscillations and transient effects, it provides a reasonable network are separated into fast and slow control actions (𝑧!,$
approximation of the change in steady-state due to these and 𝑧!,# respectively) based on their time constants, and, along
transient disturbances. Previous works have capitalized on the with their time derivative (𝑧!,$ ̇ and 𝑧!,#̇ ) are responsible for a
efficiency but have not shown to scale the formulation to large network response. 𝑧!,$ and 𝑧!,# model various control actions
systems such as the Eastern Interconnection (80k+ buses),
with varying time constants including automatic voltage
partly due to long term instabilities due to the quasi-steady state
approximations [19]. A recent tool developed by PNNL regulators (AVR), automatic generation control (AGC) and
[DCAT] [20] has capitalized on the use of power flow as an generator ramping. Additionally, discrete state variables (𝑧% ),
approximation of the quasi-steady state to study the effects of such as tap changers, change discretely during the time span.
cascading failures. Further work has also used quasi-steady The QSS approximation simplifies the differential state
state approximations to simulate the risk associated with equations above by studying a particular segment in time that
cascading outages [21]. In addition, previous work However, is long enough to consider the state variables’ response to fast
the power flow formulation should also account for steady state transient control actions (𝑧!,$ ) to reach steady state due to the
changes in frequency which would better model frequency short period of oscillations and transients. However, this time
control actions in the grid [8]. Nonetheless, the concept of using scale is also short enough to approximate the slower control
quasi-steady states as a compromise between accuracy and actions to remain stationary (𝑧!,# , 𝑧% ) thereby setting 𝑧!,#̇ to

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
zero. For example, a slow generator linear ramp is viewed as a robustness (for large scale and ill-conditioned networks) [5]
constant during a short time period during which the state and physical accuracy (amongst others due to lack of
variables have reached quasi-steady state. The approximation frequency information) [8].
is that during the time the state variables take to settle,
particular control actions are slow enough to be viewed as
constant during that period. The approximation simplifies the
equations (1)-(2) to solve for the quasi-steady state for the state
variables as follows:
0 = 𝑓## (𝑧!,# , 𝑧!,$,&& , 𝑧!,#̇ , 𝑧% , 𝑥'&& ) (3)
0 = 𝑔(𝑧! , 𝑧!,$,&& , 𝑧!,#̇ , 𝑧% , 𝑥'&& ) (4)
𝑧!,#̇ = 0 (5) Figure 3: Steady state of induction motor through transient simulation (blue)
and power flow (red marker) [21].
where, 𝑓## is a function dictating the steady state values of the
state variables, 𝑥'&& and response of fast control actions, IV. SUGAR Power Flow
𝑧!,$,&& . These approximations enable the simulation to solve for Traditional power flow simulation relies on good initial
the states at the times shown by the red markers in Figure 2. conditions close to the final state to converge to a solution.
The figure demonstrates the primary and secondary generator However, in the dynamic event-driven framework, an event
responses on a testcase where the quasi-steady states are solved may cause the system to drastically change its state, thereby
for using the approximations in (3)-(5). making the pre-event state an inadequate initial condition.
Δf
SUGAR uses circuit heuristics with state variables as real and
Primary
Disturbance @ "'
Secondary Inertial Primary Secondary
imaginary voltages and currents to achieve a level of
Inertial
Response
Control
Response
Control
Response
Response Control
Response
Control
Response
robustness [5] capable of solving large testcases representing
"$ "% "& "' "( "* ") the Eastern Interconnection, regardless of initial conditions.
Δf$
" 1) Governor Power Flow
The robust and scalable circuit formulism in SUGAR, was
extended to model the frequency deviation which allows the
tool to solve for the steady state due to frequency control
actions such as primary and secondary control. Unlike
Figure 2: Frequency response of system with primary and secondary traditional power flow that assumes the grid is operating at
frequency control due to disturbances.
nominal frequency, a governor power flow [8] introduced a
A discrete change in 𝑧% or 𝑧! causes an impulse in the frequency deviation variable (Δ𝑓) to accurately model QSS.
differential equations (1)-(2) as well, that will lead to a new a) Primary Frequency Response
quasi-steady state. For example, discrete events such as tap Each synchronous generator model in the governor power flow
changers, over- or under- voltage tripping will cause an included an additional term Δ𝑃( representing the change in
impulse that propagates into a change in steady state. Even a active power due to primary frequency response (6) [8].
discrete change in the slower control actions will create a 𝑃)
discrete change in the function 𝑓## , thereby creating a new set ∆𝑃( = − ∆𝑓 (6)
of differential equations. We define the discrete change as an 𝑅
event, and robustly simulate the quasi-steady state achieved where 𝑃) and 𝑅 are parameters describing the inertial and
after this event once the oscillations and transients have settled. droop response of a generator. The model must also respect
generator active power limits (7)-(9) to realistically simulate
III. POWER FLOW QUASI-STEADY STATE
the behavior of primary control.
In order to model the quasi-steady state after an event, the QSS Δ𝑃(*+, ≤ Δ𝑃(- ≤ Δ𝑃(*./ (7)
dynamic equations given by (3)-(5) must be solved efficiently
and accurately. Power flow has been shown to solve for a Δ𝑃(*+, = 𝑃(*+, − 𝑃(&01 (8)
steady state of a grid where all the variables have a zero-time Δ𝑃(*./ = 𝑃(*./ − 𝑃(&01 (9)
derivative [22], as shown in Figure 3, in which power flow where 𝑃(*./ and 𝑃(*+, are a generator’s active power limits
solution, indicated by the red marker, accurately describes the and 𝑃(&01 is the set active power.
steady state of an induction motor. To incorporate the limits within the power flow framework,
The QSS equations (3)-(5) are similarly formulated to find [8] introduced a continuous function that modeled the primary
the steady state of the short-term variables (x and y) with the frequency response while respecting the active power limits,
assumption of stationary control variables (𝑧% and 𝑧! ). This as shown by Figure 4. Continuous functions have been
implies that power flow can be used to solve for the QSS given previously shown to improve convergence of power flow
stationary long term and discrete variables. However, simulations, as the underlying non-linear numerical solver,
traditional power flow simulation tools have shown to lack

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
Newton-Raphson, uses the first derivatives to project the be modeled as multiplying the nominal load parameters
solution step necessary to solve the set of nonlinear equations. (𝑃?'/A+? and 𝑄?'/A+? ) by a term as given below:
The continuous function in Figure 4 models the primary
𝑃> = 𝑃?'/A+? (1 + 𝐾-$ ∆𝑓) (13)
frequency response characteristics in Region 3. The two
quadratic regions, Region 2 and 4, are quadratic 𝑄> = 𝑄?'/A+? (1 + 𝐾B$ ∆𝑓) (14)
approximations that allows the function to be continuous. where (𝐾-$ and 𝐾B$ ) are load specific parameters that describe
In addition, slack generators are similarly modeled with
the linear relation between change in frequency and change in
primary frequency response (∆𝑃&- ) with the added constraint active and reactive power respectively.
-
2𝑃&&01 + ∆𝑃& 4 = 𝑉)# 𝐼)# + 𝑉+# 𝐼+# (10) b) UFLS/UVLS
Unlike the slack model in traditional power flows, (10) The frequency change term also enables SUGAR to implicitly
constrains the slack generator to deliver a finite active power model UFLS by a continuous function that describes the
and enforces realistic frequency behavior, which serve to automatic load relief when the frequency deviation is below a
improve the accuracy of the QSS. certain threshold. Similarly, UVLS is modeled as a continuous
function that relieves the load when the voltage is below a
certain threshold [22]. This changes the load active and
reactive power as shown:
𝑃> = 𝑃> #37 (1 − 𝛼 DE>& ) (1 − 𝛼 DF>& ) (15)
𝑄> = 𝑄>#37 (1 − 𝛼 DE>& )(1 − 𝛼 DF>& ) (16)
DE>&
where 𝛼 and 𝛼 DF>& ∈ [0,1] are load shedding terms that
turn off the load based on the frequency or voltage at the bus,
respectively. When the frequency deviation is below a
threshold of 𝑓#37 or the voltage at the bus has reached a state
below 𝑉!"# , 𝛼 DE>& 𝑎𝑛𝑑 𝛼 DF>& should take a value of 1, thereby
setting 𝑃> and 𝑄> to 0. The continuous function for
Figure 4: Continuous primary response model with active power limits. 𝛼 DE>& 𝑎𝑛𝑑 𝛼 DF>& are is shown in Figure 5 which is first
derivative continuous due to the additional patching regions of
b) Secondary Frequency Response Regions 2 and 4.
While the primary frequency response responds quickly, it
does not restore the grid frequency to the nominal value. The
secondary frequency response, automatic generation control
(AGC), is able to adjust generators to restore the grid to the
ideal state. The response is initiated by an error known as the
Area Control Error (ACE) (11).
Accordingly, each participating generator is assigned a
participation factor, 𝜅, that controls the adjustment in active
power required, Δ𝑃(# . Previous work modeled the interaction
between ACE and AGC as (11) and (12) respectively [9]:
𝐴𝐶𝐸 = (Σ𝑃#!23%453% − Σ𝑃6!7465 ) + 10𝛽∆𝑓8 (11)
Δ𝑃(# = 𝜅 ∗ 𝐴𝐶𝐸 (12) Figure 5: Continuously differentiable model for UFLS.
*9
where 𝛽 [:.8<=]is the frequency bias constant, and ∆𝑓8 is the 3) Incorporating Direct Measurements
frequency deviation from nominal measured when the The SUGAR framework is also capable of integrating direct
secondary control is activated [9]. Also, (Σ𝑃#!23%453% − real-time measurements from state estimation, thereby further
Σ𝑃6!7465 ) measures the deviation of the net exchange of active improving the QSS models. Generally, measurements are
power between areas from the scheduled net exchange. given in the form of injected or consumed active power, 𝑃* ,
2) Load Modeling injected reactive power 𝑄* , a voltage magnitude 𝑉* and the
In order to better approximate the QSS, the introduction of power factor, cos (𝜃* ) [24]. These measurements have been
frequency deviation as a variable is required to better model shown to be easily integrated into SUGAR power flow, which
changes in load behavior. can aid in understanding changes in the state. Any
a) Frequency Dependent Loads incorporation of direct measurement can be considered as a
The governor power flow can be extended to compute the discrete change in the network, as the network parameters such
frequency dependence on load values (𝑃> and 𝑄> ), where these as load values or network impedances may be updated. In order
values can either be interpreted as PQ or ZIP models. Previous to stay current in an online study, the event-driven framework
work [9], demonstrated that the dependence of frequency can

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
will re-compute the QSS each time there is a new state given changing significantly due to slow control actions, and thereby
by the state estimator. significantly reducing the number of model evaluations.
One of the largest inaccuracies within power flow as an To efficiently track the progression of events, we introduce
approximation of the quasi steady state is the use of PQ and an event queue, 𝐸' , that is sorted by the time at which the event
ZIP load models. Previous work [24] have shown that these will occur. Slower control actions such as large-mass nuclear
load models often result in the wrong voltage sensitivity, generator ramp ups will include two events; one for beginning
making it inaccurate during a dynamic simulation. To improve the control and one for finishing. On the other hand, scheduled
accuracy of quasi steady state, we employ the use of BIG load discrete changes in 𝑧% , are sorted in the queue by the time they
models [23] which have demonstrated a higher level of will occur. Other discrete changes that result from the previous
accuracy in transients. Comprised of a susceptance, current QSS, such as over-loading of line, are determined at which
source and conductance, the BIG load model is able to time they occur and then inserted into the event queue.
accurately predict the variation in load due to voltage during Therefore, after solving for a quasi-steady state, the simulation
the time series, unlike traditional PQ and ZIP models. The BIG framework will take a step-in time to the next event in 𝐸' . It is
model is especially effective for an event-driven dynamic vital to ensure that the framework does not take a step that
analysis, as quasi steady state is continuously being updated skips an event of interest, as the temporal dependency on the
and requires proper sensitivities to simulate the effect of skipped event may change subsequent states.
control actions and discrete network changes. To improve the simulation time required to solve for each
4) Control Actions in Power Flow QSS, the governor power flow solver considers the previous
In order to differentiate between short-term and long-term time step’s state as an initial condition for the Newton-
control actions, we introduce a user-defined maximum time Raphson power flow solver. Many events only cause a large
constant, 𝜏G6H , that is greater than the short-term dynamic change in certain areas of the grid, therefore by using the
control action time constants. By properly defining 𝜏G6H , we previous time step state as an initial condition, the voltages of
are able to quantify which variables, defined as (𝑧! ) can use the network that do not experience much change will solve
the QSS assumption of being stationary. The user defined within few iterations. Using the circuit formalism also allows
𝜏G6H , will be subject to lower bound, defined as the time guaranteed convergence through the use of homotopy methods
constant of the slowest power flow control action, to ensure in case the previous state is not a good initial condition [5].
that power flow can still be a valid QSS approximation. A A. Backtracking
governor power flow includes essential control actions such as
AVR and primary control to represent an accurate QSS. These Over the course of the dynamic simulation, slow control
control actions have an associated time constant, and we can actions are gradually changing over events. This gradual
define the maximum of those as: change may induce a discrete event such as over-loading of
line. However, the event-driven framework would have
𝜏?EG6H = max (τIJK , 𝜏LMNN- ) (17) skipped over the exact time at which this occurs, resulting in
In governor power flow simulation, SUGAR considers a user- skipping multiple events at once. This violates the temporal
defined 𝜏G6H ≥ 𝜏?EG6H . Given a 𝜏G6H , we can now include dependencies that allow the event-driven framework to
steady state models of control systems with time constants less accurately simulate long-term dynamics. To amend this, we
than 𝜏G6H into a single power flow event. step back in time till we have found the first event caused by
V. EVENT-DRIVEN STEPS the slow control system.
Input Test Case
Governor power flow is able to robustly solve for a quasi-
Build %& and insert
steady state after an impulse change in discrete or long-term scheduled events

variables, denoted as an event. Although it is unable to capture Start with t = !"#"$

the oscillations or fluctuations before reaching the QSS, the Solve SUGAR to find

simulation platform focuses on the change in steady states as a QSS at !"#"$

result of control actions. Often in long-term dynamic studies Insert Real-Time

Measurement as
Go to next event in %&
! = !()(#$
engineers are concerned with the effect of transient control Event

actions on steady state of the grid. Solve SUGAR to find

QSS at !

Using the QSS approximation, the event-driven dynamic NO

simulation exploits the temporal sparsity of each event to Simulation Infeasible Solution Converge?

YES
simulate the response of interest. This allows the simulator to YES Create new event
overlook transient oscillations by considering fast control
Another event take * ` −!
place? Set ! = + (!()(#$ ./() )

actions to reach steady state within a single power flow NO

simulation and focus on the QSS response due to slower events ! = !1"#23 ?

such as discrete changes in the network or other frequency Solved

YES

dependent and operator-based control actions. This framework

overlooks latent periods in which the grid states are not Figure 6 Event-driven framework algorithm.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 In order to effectively find the time at which the event at t=1 to 51.8 Hz (shown by the dashed green marker), thereby
occurred we use a binary search to take a time step back equal causing a transmission line to trip. The framework backtracked
to half the time step take. This backtracking algorithm to include this event with the first event, and together caused
efficiently searches for the event, however if it finds that two an overall frequency decrease to reach a steady state at t=17s
events have occurred within a period of 𝜏G6H , then we can (indicated by the solid green marker). The markers over-layed
approximate the events to have happened simultaneously. The on the transient analysis performed by [25] in Figure 7 indicate
overall flow of the framework is shown in Figure 6. the quasi-steady states reported from the event-driven
framework matched the dynamic response.
VI. RESULTS
B. Dynamic Simulation of Large System
The algorithm described in Figure 6 offers a robust framework
Now that the approach is verified using dynamic simulation,
that simulates a series of quasi-steady states during a pre-
we can show the efficacy of the framework by simulating long-
defined time interval. The framework steps through time by
term dynamics of large systems, which is a challenge for
solving for the quasi-steady state for each event (past the
dynamic simulation tools.
minimum time constant) that causes a sufficient change in the
We demonstrate the approach on a large 8k bus system [26],
QSS dynamic equations given in (3)-(5). Along with an
with various controls and discrete changes, listed in Table 1 to
accurate representation of a QSS using governor power flow,
highlight the effectiveness of the framework. The system is
the result is a simulation platform that mimics dynamic
initialized at steady state given by the original RAW file.
simulation with a high level of efficiency.
TABLE 1: CONTROL ACTIONS IN 39-BUS TESTCASE
A. 39-Bus Comparison with Dynamic Simulation
Control Action Location/Occurrence
To justify the accuracy of the event driven framework, we AVR All generators
compare the quasi-steady states after a primary response Primary Frequency Response All Generators
control during a disturbance with a dynamic simulation. Secondary Response All Generators
Previous work created a dynamic model of IEEE testcases Generator Ramp All Generators (10MW/minute)
Update network file
[24], in particular a dynamic simulation of a 39-bus system that (with real time measurements)
Every 2 minutes
experiences a 10% increase in load resulting in a transmission
line tripping connecting bus 39. In the experiment, at t=0 sec, To initiate the dynamic sequence, all loads increased by 10%
shown in Figure 7 the system was in steady state, but after 1 to change the frequency as shown by the figure below.
second the load steps up by 10% to increase the frequency by
consuming more active power. This resulted in a quasi-steady
state, at t=4 with a frequency of 51.8 Hz, there by tripping a
line connecting bus 39 (with a load on the bus). The system
then reaches a steady state after the transient effects at t=17s.

Figure 8 Event-driven dynamic simulation of 8k bus system [26] with various

control actions and disturbances.
After the initial disturbance, the framework calculated the
quasi steady state due to the primary response at t=5s (red
marker), after an inertial nadir, at which time, the ACE was
calculated causing generators to contribute more using AGC.
The QSS of the AGC response was established at t=16.7s
(green marker) at which the frequency deviation was 0 Hz.
The system did not see a disturbance of any kind until an
update in the network (load increase, generator decrease, and
incorporation of real-time measurements) caused another
primary control response that resulted in a QSS at t=62s (blue
marker). The network change used in this simulation was a
Figure 7 Comparison of dynamic simulation and event-driven simulation of
39-bus system [25] (© [2017] IEEE). Manuscript results superimposed on uniform increase in load by 1%. At that point AGC ramped up
top of reference results. certain generators to offset the ACE calculated at t=65s.
However, during this ramp up a branch trip at t=80s (purple
The event-driven approach is able to accurately simulate the
marker) caused another event that enacted primary and
quasi-steady states reached after the oscillations settled, shown
secondary response. The frequency control actions brought the
in Figure 7. Using the governor power flow, the framework
system to a state with zero frequency deviation at t=164s
predicted the increase in frequency due to the increase in load
( ). There was no updated in the network until

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
t=240s at which the loads increased by 1% resulting in primary [3] J.Tong, L. Wang, “Design of a DSA Tool for Real Time System
Operations," IEEE International Conference on Power System
response that resulted in a QSS at 250s. The latent between t = Technology, Oct. 2016.
164s and t =240s, the black arrow, demonstrates the [4] Dan Yang, Venkataramana Ajjarapu, "A decoupled time-domain
advantage of event-driven dynamic analysis as no QSS was simulation method via invariant subspace partition for power system
calculated for that time period. analysis," IEEE Transactions on Power Systems, vol. 21, 2006.
[5] A. Pandey, et. al., "Robust Power Flow and Three Phase Power Flow
C. Scalability Analyses," in IEEE Transactions on Power Systems, 2018.
The advantage of the approach is being able to scale to large [6] H.W. Dommel, W.F. Tinney, W.L. Powell, “Further developments in
Newton's method for power system applications,” IEEE Winter Power
systems within a reasonable simulation time. The framework Meeting, Conference Paper No. 70 CP 161-PWR, January 1970.
can increase the time scale without a large penalty by using [7] Guilherme O. Alves, et. al., “A new governor power flow formulation
previous states as initial conditions. We demonstrate the based on the current injections method,” International Journal of
scalability of the framework by simulating 10 minutes of Electrical Power & Energy Systems, 2019.
[8] A. Agarwal, A. Pandey, M. Jereminov, L. Pileggi, “Implicitly Modeling
dynamic response for various network up to 80k buses, Frequency Control within Power Flow,” IEEE ISGT Europe, 2019.
representing a realistic Eastern Interconnection system, with a [9] P. Kundur, “Power System Stability and Control,” McGraw-Hill,
simulation time shown in Figure 9, including multiple Inc.,1994.
synthetic US grid testcases of varying sizes [27]. Each of these [10] G. Gurrala, et. al., “Parareal in Time for Fast Power System Dynamic
Simulation,” IEEE Transcation on Power System, vol. 31, no. 3, 2015.
simulations are run on a single core of a 2.6 GHz Quad-Core [11] Renke Huang, et. al, "Faster than real-time dynamic simulation for large-
Intel Core i7 Macbook Pro without any level of parallelization. size power system with detailed dynamic models using high-
performance computing platform", Power & Energy Society General
1000
900 Meeting 2017 IEEE, pp. 1-5, 2017.
800 [12] A. Kurita, et. al, “Multiple time-scale power system dynamic
Simulation Time (s)

700 simulation,” IEEE Transaction on Power Systems, vol.8, no.1, 1993.

600 [13] C. Dufour, J. Mahseredjian, J. Belanger, “A combined state-space nodal
500
400
method for the simulation of power system transients,” IEEE 2011 PES
300 General Meeting, 2011.
200 [14] C. Dufour, V. Jalili-Marandi, J. Bélanger, L. Snider, “Power system
100 simulation algorithms for parallel computer architectures,”IEEE 2012
0 PES General Meeting, 2012.
0 20000 40000 60000 80000 100000
Test case size
[15] D. Fabozzi, A. S. Chieh, B. Haut, T. Van Cutsem, "Accelerated and
Localized Newton Schemes for Faster Dynamic Simulation of Large
Figure 9 Simulation time for various sizes test cases running event-driven. Power Systems," IEEE Transactions on Power Systems, Nov. 2013.
[16] L. Pileggi, R. Rohrer, C. Visweswariah, Electronic Circuit & System
VII. CONCLUSION Simulation Methods, McGraw-Hill, Inc., New York, NY, USA, 1995.
[17] T. V. Cutsem, “Voltage Stability of Electric Power Systems”.
The temporal sparsity of the control actions in the power grid Boston/London/Dordrecht: Kluwer Academic Publishers, 1998.
is a bottleneck for dynamic simulation. We introduce an event- [18] T. V. Cutsem, M. E. Grenier, D. Lefebvre, “Combined Detailed and
based dynamic simulation that captures the quasi-steady state Quasi Steady-State Time Simulations for Large-disturbance Analysis.
due to each control using a governor power flow. The governor International Journal of Electrical Power and Energy Systems,” 2006.
[19] X. Z. Wang, H. D. Chiang, “Numerical Investigations on Quasi Steady-
power flow incorporates frequency information and real-time State Model for Voltage Stability: Limitations and Nonlinear Analysis,”
measurements, thereby giving an accurate quasi-steady state of International Transactions on Electrical Energy Systems, 2014.
the system. By using events to dictate time steps, the event- [20] AB. Vyakaranam, et. al., “A study of the impact of peak demand on
driven approach successively solves for QSS, thereby skipping increasing vulnerability of cascading failures to extreme contingency
events,” IEEE 2017 PES General Meeting, 2017.
over latent periods. This approach is shown to scale to systems [21] Emanuele Ciapessoni, et. al., “Benchmarking Quasi-Steady State
as large as the Eastern Interconnection for long-term time scale Cascading Outage Analysis Methodologies,” IEEE PMAPS, 2018.
while solving the simulation within minutes. The tool is [22] A. Pandey, M. Jereminov, X. Li, G. Hug, L. Pileggi, “Unified Power
capable of simulating the long-term behavior of system due to System Analyses and Models using Equivalent Circuit Formulation,”
IEEE 2016 PES ISGT USA, 2016.
a sequence of control actions by the operator while considering [23] A. Pandey, A. Agarwal, M. Jereminov, M. R. Wagner, D. M. Bromberg,
the response of a multitude of automatic control mechanism. L. Pileggi, “Robust Sequential Steady-State Analysis of Cascading
Outages,” IEEE 2019 PES ISGT Europe Conference, 2019.
ACKNOWLEDGMENT [24] B. Hooi, H. A. Song, A. Pandey, M. Jereminov, L. Pileggi, C. Faloutsos,
This work was supported in part by the Defense Advanced Research "StreamCast: Fast and online mining of power grid time
Projects Agency (DARPA) under award no. FA8750-17-1-0059 for sequences", Proc. SIAM Int. Conf. Data Mining, pp. 531-539, 2018.
RADICS program, and the National Science Foundation (NSF) under [25] P. Demetriou, et. al.,” Dynamic IEEE Test Systems for Transient
contract no. ECCS-1800812. Analysis,” IEEE Systems Journal, 2015.
[26] PSSE34 Example Test Case “bench2.raw”. Available with download of
REFERENCES PSSE34.
[1] Chirapongsananurak, P.; Santoso, "Multi-Time-Scale Simulation Tool [27] A. B. Birchfield; T. Xu; K. M. Gegner; K. S. Shetye; T. J. Overbye,
for Renewable Energy Integration Analysis in Distribution “Grid Structural Characteristics as Validation Criteria for Synthetic
Circuits". Inventions 2017. Networks,” IEEE Transactions on Power Systems, 2017.
[2] J. Shu, W. Xue, W. Zheng, “A parallel transient stability simulation for
power systems,” IEEE Transactions on Power Systems, vol. 20, 2005.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020