Acknowledgements

Virtual Inertia Emulation and

Placement in Power Grids
Institute for Mathematics and its Applications
Control at Large Scales: Energy Markets & Responsive Grids

Florian Dörfler Bala Kameshwar Poolla Taouba Jouini Catalin Arghir

Dominic Gross Saverio Bolognani

2 / 32

At the beginning of power systems was . . . Operation centered around bulk synchronous generation
50.02
f [Hz]
50.01 Primary Control Tertiary Control
50.00
f - Setpoint
49.99

49.98

49.97 Secondary Control

49.96
PP - Outage
49.95 Oscillation/Control
PS Oscillation
49.94
At the beginning was the synchronous machine: Pgeneration
49.93
M

d
ec

49.92

ω(t) = Pgeneration (t) − Pdemand (t) ω

ha

M
ni

49.91
dt
ca
lI

49.90
ne

change of kinetic energy = instantaneous power balance

rti

49.89
a

Pdemand 49.88
16:45:00 16:50:00 16:55:00 17:00:00 17:05:00 17:10:00 17:15:00
8. Dezember 2004
Fact: the AC grid & all of power system operation Frequency Mettlen, Switzerland
Frequency Athens
has been designed around synchronous machines. Source: W. Sattinger, Swissgrid
3 / 32 4 / 32
Distributed/non-rotational/renewable generation on the rise A few (of many) game changers . . .
synchronous generator new workhorse scaling

location & distributed implementation

Almost all operational problems can principally be resolved . . . but one (?)
Source: Renewables 2014 Global Status Report
5 / 32 6 / 32

Fundamental challenge: operation of low-inertia systems Low-inertia stability: # 1 problem of distributed generation
75 mHz Criterion Summary - Short View - Year 2001-2011
We slowly loose our giant electromechanical low-pass filter: Pgeneration 2001 2002
2003 2004 2005 2006
2007 2008
30000 2009
2010

25000

d ω
M ω(t) = Pgeneration (t) − Pdemand (t) 20000

dt Duration [s]
Events [-]
15000

10000

change of kinetic energy = instantaneous power balance 5000

Pdemand 0

Number * 10

Months of the year

# frequency violations in Nordic grid Fig. 3.2:

Number * 10 Duration

Frequency quality behaviour in Continental Europe during the last ten years. Source: Swissgrid
50
(source: ENTSO-E) same inobserved
It can clearly be Switzerland (source:
how the accumulated time Swissgrid)
continuously increases with higher
frequency deviations as well as the number of corresponding events.
49.8
3.1.2. CAUSES

M
J
f [Hz]

The unbundling process has separated power generation from TSO, imposing new
49.6 commercial rules in the system operating process. Generation units are considered as

inertia is shrinking, time-varying, & localized, . . . & increasing disturbances

simple balance responsible parties without taking dynamic behaviour into account: slow
or fast units. Following the principle of equality, the market has created unique rules for
49.4 settlement: energy supplied in a time frame versus energy calculated from schedule in
the same time frame. Energy is traded as constant power in time frame.

The market, being orientated on energy, has not developed rules for real time operation
49.2 as power. In consequence we are faced with the following unit behaviour (Figure 3.3):

Solutions in sight: none really . . . other than emulating virtual inertia

Load
evolution
which
must be A: Fast units
49 B: Slow unit
0 5 10 15 20 25 30 35 through fly-wheels, batteries, super caps, HVDC, demand-response, . . .
covered response
response Energy to be compensated - real
cause of frequency deterministic

Time t [s] Power

basepoint
scheduled
deviations

7 / 32 8 / 32
Energy
Contracted
Low inertia issues have been broadly recognized Virtual inertia emulation
by TSOs, device manufacturers, academia, etc. devices commercially available, required by grid-codes, or incentivized through markets
!""" #$%&'%(#!)&' )& *)+"$ ','#"-'. /)01 23. &)1 2. -%, 2456 5676

Massive InteGRATion of power Electronic devices !"#$%&'"'() %* +$,(-.'() /'-#%(-'

!89:;8;<=><? />@=AB: !<;@=>B >< CD!EFGBH;I
.( 0.1$%2$.3- 4-.(2 5.$)6,7 !('$).,
8.".-9 :%(.! "#$%&'# (&)*&+! ,---; :6$<,(,$,<,(, =%%77,! (&)*&+! ,---; ,(3
+><I *JK;@ E;<;@B=>J<
06>67 ?@ ?9,(3%$>,$! (&)*&+! ,--- -JLB88BI@;MB DBNLB@> -J?LBIIB8 %@B<>! "#$%&'# (&)*&+! ,---. B<I "LBO D1 ":F'BBIB<P! "&'./+ (&)*&+! ,---

Virtual synchronous generators: A survey and new perspectives !"#$%&'()*+,-+#'"(./#0*/1(2-33/*04($(5&*0-$1((

“The question that has to be
Hassan Bevrani a,b,⇑, Toshifumi Ise b, Yushi Miura b
a
Dept. of Electrical and Computer Eng., University of Kurdistan, PO Box 416, Sanandaj, Iran
6#+*0&$(7*/8&9+9(:"(!&;0*&:-0+9(<#+*="(20/*$=+(
0/(6;/1$0+9(7/>+*(2";0+%;((
b
Dept. of Electrical, Electronic and Information Eng., Osaka University, Osaka, Japan

examined is: how much power !""" #$%&'%(#!)&' )& *)+"$ ','#"-'. /)01 23. &)1 2. -%, 2456

?$-0@&+*(!+1&11+A(!"#$"%&'()))A(B*-#/()*$#C/&;A(*"+,-%'!"#$"%&'()))A($#9(?&11+;(D$1$*$#=+(

electronics can the grid cope !789:;< "=>?<:;@7 (@7:9@? ':9<:8AB C@9 !"#$%&#'$%()*+'",'"%-#,.%/#",012%3#*',#4%
with?” (European Commission) /'(DE/F( #9<7G=;GG;@7 'BG:8=G 5,)"16'%
H;8I8; JK>. (<=LI8?? F1 M@@:K. N9<;7 *1 %O<=. %7O98P H1 $@GQ@8. <7O (K9;G N1 M9;AK:
7898:%+1*%;'<'*=''4>?%@%58;8A8%$'%A11*?@%!"#$%&'("()"&*'+,,,@%98%/1"'21B%1*$%38%/#<<4.'"C@%
!"#$%&'("()"&'+,,,%

current controls what else?

d
M ω(t) = Pgeneration (t)−Pdemand (t) . . . essentially D-control
dt

⇒ plug’n’play (decentralized & passive), grid-friendly, user-friendly, . . .

⇒ today: where to do it? how to do it properly?

all options are on the table and keep us busy . . . 9 / 32 10 / 32

Outline

Introduction

Novel Virtual Inertia Emulation Strategy

inertia emulation
Optimal Placement of Virtual Inertia

Conclusions
Challenges in power converter implementations Averaged inverter model
1
Electrical Power and Energy Systems 54 (2014) 244–254

Contents lists available at ScienceDirect

Real Time Simulation of a Power System with iαβ R L
Electrical Power and Energy Systems VSG Hardware in the Loop ix
DC cap & AC filter equations:
ic +
journal homepage: www.elsevier.com/locate/ijepes
Vasileios Karapanos, Sjoerd de Haan, Member, IEEE, Kasper Zwetsloot + + 1
Faculty of Electrical Engineering, Mathematics and Computer Science
Delft University of Technology Cdc v̇dc = −Gdc vdc + idc − m> iαβ
Virtual synchronous generators: A survey and new perspectives
Delft, the Netherlands
E-mails: vkarapanos@gmail.com, v.karapanos@tudelft.nl, s.w.h.dehaan@tudelft.nl
idc vdc gdc Cdc vx C vαβ iload 2
Hassan Bevrani a,b,⇑, Toshifumi Ise b, Yushi Miura b
a

b
Dept. of Electrical and Computer Eng., University of Kurdistan, PO Box 416, Sanandaj, Iran
Dept. of Electrical, Electronic and Information Eng., Osaka University, Osaka, Japan Abstract- The method to investigate the interaction between a
To better study and witness the effects of virtual inertia, the
hardware of a real VSG should be tested within a power
C v̇αβ = −iload + iαβ
system. Investigating the interaction between a real VSG and
Virtual Synchronous Generator (VSG) and a power system is
− − −
˙ = −Riαβ + 1 mvdc − vαβ
presented here. A VSG is a power-electronics based device that a power system is not easy as a power system cannot be
a r t i c l e i n f o a b s t r a c t emulates the rotational inertia of synchronous generators. The manipulated without disturbing customers. Building a real
Article history:
Received 31 December 2012
In comparison of the conventional bulk power plants, in which the synchronous machines dominate, the
distributed generator (DG) units have either very small or no rotating mass and damping property. With
development of such a device started in a pure simulation
environment and extends to the practical realization of a VSG.
power system for testing purposes would be too costly. By
replacing the power system with a real time simulated one,
Liαβ
Received in revised form 12 June 2013
Accepted 13 July 2013
growing the penetration level of DGs, the impact of low inertia and damping effect on the grid stability
and dynamic performance increases. A solution towards stability improvement of such a grid is to pro-
vide virtual inertia by virtual synchronous generators (VSGs) that can be established by using short term
Investigating the interaction between a VSG and a power system
is a problem, as a power system cannot be manipulated without
disturbing customers. By replacing the power system with a real
this problem can be solved. In this paper the testing of a real
hardware VSG in combination with a simulated power system
2
energy storage together with a power inverter and a proper control mechanism. time simulated one, this problem can be solved. The VSG then is described.

delays in measurement acquisition, signal processing, & actuation

Keywords:
The present paper reviews the fundamentals and main concept of VSGs, and their role to support the The power processor from Fig.1 is built from a Triphase®
1
Virtual inertia
Renewable energy
power grid control. Then, a VSG-based frequency control scheme is addressed, and the paper is focused
on the poetical role of VSGs in the grid frequency regulation task. The most important VSG topologies
interacts with the simulated power system through a power
interface. The advantages of such a laboratory test-setup are [9], [10] inverter system. The Matlab/simulink VSG
VSG
Frequency control with a survey on the recent works/achievements are presented. Finally, the relevant key issues, main numerous and should prove beneficial to the further algorithm is directly implemented on the inverter system

modulation: ix = 12 m> iαβ , vx = 12 mvdc

Voltage control technical challenges, further research needs and new perspectives are emphasized. development of the VSG concept. through a dedicated FPGA interface developed by Triphase®.
Microgrid ! 2013 Elsevier Ltd. All rights reserved.

I. INTRODUCTION
In order to test the hardware implemented VSG and to
study its effects within a power system, it is connected with a
passive: (idc , iload ) → (vdc , vαβ )
Short term frequency stability in power systems is secured real time digital simulator from RTDS® [17] through a power
1. Introduction A solution towards stabilizing such a grid is to provide addi-

2 accuracy in AC measurements (need averaging)

The capacity of installed inverter-based distributed generators
(DGs) in power system is growing rapidly; and a high penetration
level is targeted for the next two decades. For example only in Ja-
tional inertia, virtually. A virtual inertia can be established for
DGs/RESs by using short term energy storage together with a
power electronics inverter/converter and a proper control mecha-
nism. This concept is known as virtual synchronous generator
mainly by the large rotational inertia of synchronous
machines which, due to its counteracting nature, smoothes out
the various disturbances. The increasing growth of dispersed
interface (Fig 2).

generation will cause the so-called inertia constant of the

pan, 14.3 GW photovoltaic (PV) electric energy is planned to be
connected to the grid by 2020, and it will be increased to 53 GW
by 2030. In European countries, USA, China, and India significant
targets are also considered for using the DGs and renewable energy
sources (RESs) in their power systems up to next two decades.
3 constraints on currents, voltages, power, etc.
Compared to the conventional bulk power plants, in which the
synchronous machine dominate, the DG/RES units have either very
small or no rotating mass (which is the main source of inertia) and
damping property. The intrinsic kinetic energy (rotor inertia) and
damping property (due to mechanical friction and electrical losses
in stator, field and damper windings) of the bulk synchronous gen-
erators play a significant role in the grid stability.
With growing the penetration level of DGs/RESs, the impact of
low inertia and damping effect on the grid dynamic performance
4 guarantees on stability and robustness
and stability increases. Voltage rise due to reverse power from
PV generations [1], excessive supply of electricity in the grid due
to full generation by the DGs/RESs, power fluctuations due to var-
iable nature of RESs, and degradation of frequency regulation
(VSG) [3] or virtual synchronous machine (VISMA) [4]. The units will
then operate like a synchronous generator, exhibiting amount of
inertia and damping properties of conventional synchronous ma-
chines for short time intervals (in this work, the notation of
‘‘VSG’’ is used for the mentioned concept). As a result, the virtual
inertia concept may provide a basis for maintaining a large share
of DGs/RESs in future grids without compromising system stability.
The present paper contains the following topics: first the funda-
mentals and main concepts are introduced. Then, the role of VSGs
in microgrids control is explained. In continuation, the most
important VSG topologies with a review on the previous works
and achievements are presented. The application areas for the
VSGs, particularly in the grid frequency control, are mentioned. A
frequency control scheme is addressed, and finally, the main tech-
nical challenges and further research needs are addressed and the
paper is concluded.
power system to decrease. This may result in the power
system becoming instable [1]-[3]. A promising solution to
such a development is the Virtual Synchronous Generator
(VSG) [4]-[8], which replaces the lost inertia with virtual
Fig. 2. RTDS and Power Interface and VSG in a closed loop. ✓
inertia. The VSG consists of three distinctive components,
namely a power processor, an energy storage device and the
The RTDS® simulates power systems in real time and is
often used in closed loop testing with real external hardware. model of a θ̇ = ω
appropriate control algorithm [4] as shown in Fig. 1. This
system has been tested in a full Matlab/Simulink [21]
Keeping in mind that the ADCs and DACs, which are the
inputs and outputs of the RTDS, have a dynamic range of
 
− sin(θ)
simulation environment with promising results. ±10V max rated at 5mA max and the Triphase® inverter
synchronous >
system is rated at 16kVA, it is clear that a power interface has
M ω̇ = −Dω + τm +
to come in between to make this union possible as it is shown iαβ Lm if

2. Fundamentals and concepts

in Fig. 2.
The main function of the power interface is to replicate the
voltage waveform of a bus in a network model to a level of
generator cos(θ) if
(especially in the islanded microgrids [2], can be considered as
some negative results of mentioned issue. The idea of the VSG is initially based on reproducing the dynamic
properties of a real synchronous generator (SG) for the power
400VLL at terminal 1 in Fig. 2. This terminal is loaded by the
VSG and the current flowing from/to the VSG is fed back to
the RTDS, to load the bus in the network model with that
C v̇αβ = −Gload vαβ + iαβ  
electronics-based DG/RES units, in order to inherit the advantages

˙ = −Riαβ − vαβ − ωLm if − sin(θ)

⇑ Corresponding author at: Dept. of Electrical and Computer Eng., University of current.
Kurdistan, Sanandaj, PO Box 416, Iran. Tel.: +98 8716660073. of a SG in stability enhancement. The principle of the VSG can be
E-mail address: bevrani@ieee.org (H. Bevrani). applied either to a single DG, or to a group of DGs. The first Fig. 1. The VSG Concept. The simulated power system is a transfer from the

0142-0615/$ - see front matter ! 2013 Elsevier Ltd. All rights reserved.
Matlab/Simulink environment, in which the system was
developed initially, to RSCAD [18] format. Ls iαβ
http://dx.doi.org/10.1016/j.ijepes.2013.07.009

today: use DC measurement, exploit analog storage, & passive control This work is a part of the VSYNC project funded by the European
Commission under the FP6 framework with contract No:FP6 – 038584
In section II the requirements for testing a VSG and the
principle of a VSG are discussed and in section III the test set
cos(θ)
(www.vsync.eu).

11 / 32 12 / 32

A standard power electronics control would continue by See the similarities & the differences ?
iαβ R L
DC cap & AC filter equations:
3 2 1 acquiring & processing ix
ic +
reference synthesis + + 1
tracking control
(virtual sync gen, of AC measurements Cdc v̇dc = −Gdc vdc + idc − m> iαβ
(cascaded PIs) idc vdc gdc Cdc vx C vαβ iload 2
droop/inertia, etc.) C v̇αβ = −iload + iαβ
- 2 synthesis of references − − −
Liαβ˙ = −Riαβ + 1 mvdc − vαβ
(voltage/current/power) 2
4 4 1
3 track references at
iαβ R L
modulation: ix = 12 m> iαβ , vx = 12 mvdc passive: (idc , iload ) → (vdc , vαβ )
+
ix
+
ic
+
converter terminals
idc vdc gdc Cdc vx C vαβ iload

− − −
4 actuation via emulation ✓
(inner loop) and/or via model of a θ̇ = ω  
DC source (outer loop) synchronous > − sin(θ)
M ω̇ = −Dω + τm + iαβ Lm if
generator cos(θ) if
C v̇αβ = −Gload vαβ + iαβ  
˙ − sin(θ)
let’s do something different (smarter?) today . . . Ls iαβ = −Riαβ − vαβ − ωLm if
cos(θ)

13 / 32 14 / 32
Model matching (6= emulation) as inner control loop Further properties of machine matching control
0.2 50
iαβ R L
DC cap & AC filter equations:
1 quadratic nose curves: 150
40
0.15

Frequency (Hz)
ix

g load(Ω-1(V)
ic + stationary P vs. (|V |, ω)

)
1 30

Amplitude
+ + 100

Cdc v̇dc = −Gdc vdc + idc − m> iαβ 0.1

20
vdc gdc Cdc vx vαβ 2 2 /4G
idc C iload
C v̇αβ = −iload + iαβ ⇒ P ≤ Pmax = idc dc
50
0.05
10

− − − 00 0

˙ = −Riαβ + 1 mvdc − vαβ

00 0.1 0.2
0.5 0.3 10.4 0.5 0.6
1.5 0.7 20.8 0.9 1

Liαβ ⇒ (P, ω)-droop ≈ 1/η time(s) ×10 4

2 200

⇒ (P, |V |)-droop ≈ 1/µ 100

 

Vx
− sin(θ)
0

matching control: θ̇ = η · vdc , m =µ· with η, µ > 0 2 eye candy: load step -100
cos(θ)
-200
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
3 remains passive: time(s)
50

⇒ pros: is balanced, uses natural storage, & based on DC measurement (idc , iload ) → (vdc , vαβ ) 150
40

Frequency (Hz)
Amplitude (V)
30
100
Cdc Gdc idc µ
⇒ virtual machine with M = η2
, D= η2
, τm = η , if = ηLm 4 solid plug’n’play base 20
50
10
for outer control loops
⇒ base for outer controls via idc & µ: virtual governor, PSS, & inertia 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0

Active power P x
15 / 32 16 / 32

Linearized & Kron-reduced swing equation model

mi θ̈i + di θ̇i = pin,i − pe,i

Pgeneration
Pgeneration+η
generator swing equations

P ω
pe,i ≈ j∈N bij (θi − θj )

optimal placement linearized power flows

Pdemand
demand
likelihood of disturbance at #i: ti ≥ 0
of virtual inertia state space representation:
      
θ̇ 0 I θ 0
= + T 1/2 η
ω̇ −M −1 L −M −1 D ω M −1
| {z } | {z }
A B

where M, D, & T are diagonal & L = LT (Laplacian)

17 / 32
Performance metric for emulation of rotational inertia Coherency performance metric & H2 norm
Energy expended by the system to return to synchronous operation:
f
restoration time Z ∞ X Xn
aij (θi (t) − θj (t))2 + si ωi2 (t) dt
nominal frequency 0 {i, j}∈ E i=1

effort

max deviation H2 norm interpretation:

" # 
1/2
ROCOF Q1 0 θ
1 associated performance output: y= 1/2
f 0 Q ω
2
nominal frequency
System norm:
amplification of R∞
2 impulses (faults) −→ output energy 0 y (t)T y (t) dt
disturbances: impulse (fault), step (loss of unit), white noise (renewables)
to

3 white noise (renewables) −→ output variance lim E y (t)T y (t)
performance outputs: integral, peak, ROCOF, restoration time, . . . t→∞

18 / 32 19 / 32

Algebraic characterization of the H2 norm Problem formulation

Lemma: H2 norm via observability Gramian minimize kG k22 = Trace(B T PB) → performance metric
P , mi
Xn
kG k22 = Trace(B T PB) subject to mi ≤ mbdg → budget constraint
i=1
R∞ T
where P is the observability Gramian P = 0 e A t C T Ce At dt mi ≤ mi ≤ mi , i ∈ {1, . . . , n} → capacity constraint
P A + AT P + Q = 0 → observability Gramian
I P solves a Lyapunov equation: P A + AT P +Q =0
P [1 0] = [0 0] → uniqueness
I A has a zero eigenvalue → restricts choice of Q
" # 
1/2
Q1 0 θ 1/2 Insights
y= 1/2 Q1 1 = 0 
0 Q2 ω 1 m appears as m−1 in system matrices A , B 


large-scale &
2 product of B(m) & P in the objective ⇒
I P is unique for P [1 0] = [0 0] 
 non-convex

3 product of A(m) & P in the constraint

20 / 32 21 / 32
 Building the intuition: results for two-area networks
Fundamental learnings
1 explicit closed-form solution is rational function
2 sufficiently uniform (t/d)i → strongly convex & fairly flat cost
3 non trivial in the presence of capacity constraints
where would you place the inertia?
Dissimilar and Identical t/d ratios Budget, Sum, Inertia1, Inertia2
6 25

uniform, max capacity, near disturbance? 5

dissimilar t/d m1 ∗
m2 ∗

Optimal inertia allocation

20
identical t/d mbdg
4 m1 ∗ + m2 ∗

f(m1)
15
3
10
2
1 5

0
0 2 4 6 8 10 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
m1 t1 =1-t 2

performance metric optimal inertia allocation

22 / 32

Closed-form results for cost of primary control Primary control . . . cont’d

𝜔
Theorem: the primary control effort optimization reads equivalently as
P/θ̇ primary droop control
Xn ti
(ωi − ω ∗ ) ∝ (Pi ∗ − Pi (θ)) minimize
𝜔*
m mi i=1 mi
𝜔sync Xn
Di θ̇i = Pi ∗ − Pi (θ) subject to mi ≤ mbdg
i=1

(can also model effect of PSS control) P mi ≤ mi ≤ mi , i ∈ {1, . . . , n}

P1 P2

Primary control effort → accounted for by integral quadratic cost Key take-aways:
Z I optimal solution independent of network topology
∞
T √
θ̇(t) D θ̇(t) dt I optimal allocation ∝ ti or mi = min{mbdg , mi }
0

1/2 1/2
which is the H2 performance for the penalties Q1 = 0 and Q2 =D
Location & strength of disturbance are crucial ⇒ suggests min max
23 / 32 24 / 32
 Taylor & power series expansions
Key idea: expand the performance metric as a power series in m
kG k22 = Trace(B(m)T P(m)B(m))

Motivation: scalar series expansion at mi in direction µi :

1 1 δµi
numerical method for mi + δµi
=
mi
− 2 + O(δ 2 )
mi

the general case Expand system matrices as Taylor series in direction µ:

(0) (1)
A(m + δµ) = A(m,µ) + A(m,µ) δ + O(δ 2 )
(0) (1)
B(m + δµ) = B(m,µ) + B(m,µ) δ + O(δ 2 )

Expand the observability Gramian as a power series in direction µ:

(0) (1)
P(m + δµ) = P(m,µ) + P(m,µ) δ + O(δ 2 )
25 / 32

Explicit gradient computation

Expansion of system matrices & Gramian ⇒ match coefficients . . .

Formula for gradient at m in direction µ

1 nominal Lyapunov equation for O(δ 0 ):
P(0) = Lyap(A(0) , Q)

2 perturbed Lyapunov equation for O(δ 1 ) terms:

results
(1) (0) (0) (1) (1) T (0)
P = Lyap(A ,P A +A P )

3 expand objective in direction µ:

kG k22 = Trace(B(m)T P(m)B(m)) = Trace((. . .) + δ(. . .)) + O(δ 2 )
T T
4 gradient: Trace(2 ∗ B (1) P (0) B (0) + B (0) P (1) B (0) )

⇒ use favorite method for reduced optimization problem

26 / 32
Modified Kundur case study: 3 regions & 12 buses Heuristics outperformed by H2 - optimal allocation
transformer reactance 0.15 p.u., line impedance (0.0001+0.001i) p.u./km Original, Optimal, and Capacity allocations Cost
160 0.25

Scenario: disturbance at #4 m
m∗
m 0.2
120
10 11 9
I locally optimal solution 0.15

inertia

trace
80
outperforms heuristic 0.1

12
max/uniform allocation 40
0.05

1 5
I optimal allocation ≈ 0
1 2 4 5 6
node
8 9 10 12
0

3 4 8 7 matches disturbance allocation subject to capacity constraints

Original, Optimal, and Uniform allocations Cost

I inertia emulation at all 0.25
m
2 6 undisturbed nodes is 160 m∗
muni 0.2

actually detrimental 120

0.15

inertia

trace
⇒ location of disturbance &
80
0.1

  inertia emulation matters 40 0.05

In − n1 11
uniform deviation from sync as performance metric: Q = 0
1 2 4 5 6 8 9 10 12
0

In node

allocation subject to the budget constraint

27 / 32 28 / 32

Eye candy: time-domain plots of post fault behavior

Original, Optimal, and Uniform allocations
Angle Diff. Freq #4 Freq #5 Control Effort
(a) (b) ×10 -3 (c) (d)
0.05 0.15 5 2
m
0.04 1.5
muni
4 m∗

0.03 0.1
1

0.02 3
0.5

0.01 0.05
Control effort

2
∆θ 1 -∆θ 4

0
∆ω4

∆ω5

conclusions
-0.5
1
-0.01 0

-1
-0.02 0

-1.5
-0.03 -0.05
-1
-0.04 -2

-0.05 -0.1 -2 -2.5

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
Time(s) Time(s) Time(s) Time(s)

Take-home messages:
best oscillation smallest peak undisturbed sites minimal control
performance frequency at #4 are irrelevant effort mi · θ̈i
29 / 32
Conclusions on virtual inertia emulation Recall: operation centered around (virtual) sync generators
Where to do it? 50.02
f [Hz]
50.01 Primary Control Tertiary Control
1 H2 -optimal (non-convex) allocation 50.00
f - Setpoint
2 closed-form results for cost of primary control 49.99

49.98
3 numerical approach via gradient computation 49.97 Secondary Control
49.96
PP - Outage
How to do it? 49.95 Oscillation/Control
PS Oscillation
49.94
1 down-sides of naive inertia emulation
49.93

In
2 novel machine matching control

er
49.92

tia
&
49.91

Em
49.90

ul
What else to do? Inertia emulation is . . .

at
io
49.89

n
, decentralized, plug’n’play (passive), grid-friendly, user-friendly, . . . 49.88
16:45:00 16:50:00 16:55:00 17:00:00
8. Dezember 2004
17:05:00 17:10:00 17:15:00

/ suboptimal, wasteful in control effort, & need for new actuators

Frequency Mettlen, Switzerland
Frequency Athens
Source: W. Sattinger, Swissgrid

30 / 32 31 / 32

A control perspective of power system operation

Conventional strategy: emulate generator physics & control
Z t
M ω̇(t) = Pmech − Dω(t) − ω(τ ) d τ − Pelec
0
| {z } | {z } | {z } | {z }
(virtual) inertia tertiary control primary control secondary control
appendix
Essentially all PID + setpoint control (simple, robust, & scalable)
Z t
M ω̇(t) = P − Dω(t) − ω(τ ) d τ − Pelec
0
| {z } |{z} | {z } | {z }
D set-point P I

Control engineers should be able to do better . . .

32 / 32
Some properties & different viewpoints Eye candy: response to a load step
1 quadratic nose curves: 50

150
40
stationary P vs. (|V |, ω)

Frequency (Hz)
Amplitude (V)
30
100 50 iαβ R L

⇒ P ≤ Pmax = 2 /4G
idc 20 1500.2
40
ix
+ + ic iload
dc

Frequency (Hz)
50

(V)
10 30

Amplitude
0.15
100 idc Gdc Cdc vx vαβ C Gload

⇒ reactive power not

(Ω-1)
0 0 20
0 0.5 1 1.5 2 0.1 − −

g load
50
×10 4
directly affected 0.05
10

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

⇒ (P, ω)-droop ≈ 1/η inverter 0

0 0.1 0.2 0.3
Active power P x
0.4 0.5
time(s)
0.6 0.7 0.8 0.9 1

(idc , iload ) 1 (vdc , vαβ )

Cdc v̇dc = −Gdc vdc + i∗dc − m⊤ iαβ
⇒ (P, |V |)-droop ≈ 1/µ 2
200
0.2

C v̇αβ = −iload + iαβ 100

_ 0.15

˙ = −Riαβ + 1 mvdc − vαβ

g load(Ω-1)
Liαβ

Vx
0 0.1

2
2 reformulation as virtual -100
0.05

& adaptive oscillator -200

0
0 0.1 0.2 0.3 0.4 0.5
time(s)
0.6 0.7 0.8 0.9 1

modulation 0 0.1 0.2 0.3 0.4 0.5

time(s)
0.6 0.7 0.8 0.9 1
200

! "
m 0 1 vdc
3 remains passive: µ ξ˙ = vdc η ·
−1 0
ξ η
100

Vx
0

(idc , iload ) → (vdc , vαβ ) -100

-200
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
time(s)

Spectral perspective on different inertia allocations The planning problem

sparse allocation of limited resources
Cone, Original, Optimal, and Uniform allocations
3
`1 -regularized inertia allocation (promoting a sparse solution):
cone

2
m
muni minimize Jγ (m, P) = kG k22 + γkm − mk1
m∗ P , mi
1
Xn
subject to mi ≤ mbdg
Imaginary Axis

i=1
0
mi ≤ mi ≤ mi i ∈ {1, . . . , n}
P A + AT P + Q = 0
-1

-2
P[1 0] = [0 0]
-3
where γ ≥ 0 trades off sparsity penalty and the original objective
-0.18 -0.16 -0.14 -0.12 -0.1 -0.08 -0.06 -0.04 -0.02 0 0.02
Real Axis

Highlights:
• m = m → best damping asymptote & best damping ratio 1 regularization term is linear & differentiable
• Spectrum holds only partial information !! 2 gradient computation algorithm can be used with some tweaking
Relative performance loss (%) as a function of γ Uniform disturbance to damping ratio
0% → optimal allocation, 100% → no additional allocation power sharing → d ∝ P ∗ , assuming t ∝ source rating P ∗

Localized and Uniform disturbances

10 100 Theorem: for ti /di = tj /dj the allocation problem reads equivalently as
Xn si
8 80 minimize

Relative Performance Loss (%)

mi i=1 mi
Card
Cardinality-Loc
Cardinality-Uni
Xn
6 Perf. %
Performance-Loc 60 subject to mi ≤ mbdg
Cardinality

Performance-Uni i=1
mi ≤ mi ≤ mi , i ∈ {1, . . . , n}
4 40

2 20
Key takeaways:
optimal solution independent of network topology
0 0
0 0.5 1 1.5 2 2.5 3 √
γ ×10 -4 allocation ∝ si or mi = min{mbdg , mi }
1 uniform disturbance ⇒ ∃ γ 1.3% loss ≡ (9 → 7) What if freq. penalty ∝ inertia? → norm independent of inertia
2 localized disturbance ⇒ (2 → 1) without affecting performance

Taylor & power series expansions Taylor & power series expansions cont’d
Key idea: expand the performance metric as a power series in m Expand the observability Gramian as a power series in direction µ

kG k22 = Trace(B(m)T P(m)B(m)) (0) (1)

P(m) = P(m + δµ) = P(m,µ) + P(m,µ) δ + O(δ 2 )

Motivation: scalar series expansion at mi in direction µi : Formula for gradient in direction µ

1 1 δµi 1 nominal Lyapunov equation for O(δ 0 ): P(0) = Lyap(A(0) , Q)
= − 2 + O(δ 2 )
(mi + δµi ) mi mi 2 perturbed Lyapunov equation for O(δ 1 ) terms:
T
Expand system matrices in direction µ, where Φ = diag(µ): P(1) = Lyap(A(0) , P(0) A(1) + A(1) P(0) )
   
(0) 0 I (1) 0 0 3 expand objective in direction µ:
A(m,µ) = , A(m,µ) =
−M −1 L −M −1 D ΦM −2 L ΦM −2 D
    kG k22 = Trace(B(m)T P(m)B(m)) = Trace((. . .) + δ(. . .)) + O(δ 2 )
(0) 0 (1) 0
B(m,µ) = , B(m,µ) =
M −1 T 1/2 −ΦM −2 T 1/2 T T
4 gradient: Trace(2 ∗ B(1) P(0) B(0) + B(0) P(1) B(0) )
Gradient computation Heuristics outperformed also for uniform disturbance
Original, Optimal, and Capacity allocations Cost
Algorithm: Gradient computation & perturbation analysis 150
m
m∗
0.15
Scenario: uniform disturbance
m

Input → current values of the decision variables mi

100
0.1
Heuristics for placement:

inertia

trace
Output → numerically evaluated gradient ∇f of the cost function 50 0.05
1 max allocation in case of
capacity constraints
Evaluate the system matrices A(0) , B(0) based on current inertia
0 0
1 1 2 4 5 6
node
8 9 10 12
2 uniform allocation in case
allocation subject to capacity constraints of budget constraint
2 Solve for P(0) =Lyap(A(0) , Q) using a Lyapunov routine
Original, Optimal, and Uniform allocations Cost
Results & insights:
90

3 For each node- obtain the perturbed system matrices A(1) , B(1) m
m∗
muni
0.15

60
1 locally optimal solution
T 0.1

Compute P(1) =Lyap(A(0) , P(0) A(1) + A(1) P(0) ) outperforms heuristics

inertia
4

trace
T T
30
0.05 2 optimal solution 6= max
5 Gradient ⇒ Trace(2 ∗ B(1) P(0) B(0) + B(0) P(1) B(0) ) inertia at each bus
0 0
1 2 4 5 6 8 9 10 12
node

allocation subject to the budget constraint