Electric Power Systems Research 190 (2021) 106647

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

On the minimal set of controllers and sensors for linear power flow☆ T
Edwin Mora , Florian Steinke
⁎

Energy Information Networks and Systems, Technische Universität Darmstadt, Germany

A RT ICLE INFO ABSTRACT

Keywords: We consider a linear power flow model with interval-bounded nodal power injections and limited line power
Controllability flows. We determine the minimal number of power injections to control based on a minimal set of measure­
Observability ments, such that the overall system is feasible for all assignments of the non-controlled power injections. For the
Power flow important case where the possible measurements are the nodal power injections, we show that the problem can
Resilience
be solved efficiently as a mixed-integer linear program (MILP). When also line power flows are considered as
potential measurements, we derive an iterative, greedy algorithm that provides a feasible, but potentially
conservative solution. We apply the developed algorithms to both a simple microgrid and a modified version of
the IEEE 118 bus test power system. We show that in both cases a sparse solution in terms of the number of
required controllers and measurements can be obtained. Moreover, the number of required measurements can be
reduced significantly if line flow measurements are considered additionally to nodal power injections.

1. Introduction very active again in the study of complex networks, see, e.g., [4]. While
most formulations of the problem are NP-Hard due to its combinatorial
Volatile renewable energies are transforming classical power grids nature, finding only the minimum set of actuators is possible in poly­
with few large generators into complex cyber-physical networks. These nomial-time [5]. This finding is based on structural controllability
contain a large number of distributed generators and controllable loads, theory [6] and can be used to develop distributed algorithms for finding
and power lines are often operated close to their limits. In this context, the minimum number of controlled and measured nodes [4]. Structural
we ask: What is the smallest set of generators and/or loads that must be controllability theory can also be used to analyze cyber-security aspects
controlled based on the values of a minimal number of measurements, in distributed power grids [7], e.g., for evaluating the detectability and
such that (s.t.) the entire system state is feasible for all possible values identifiability of hacked nodes. Another line of research aims at de­
of the remaining elements? signing control structures that minimize the control effort, using con­
Being able to identify the (optimally small) set of critical elements in trollability metrics derived from the controllability Gramian of the
complex power grids reduces the cost and effort for their control. system [8]. Many of the related input/ouput selection problems are
Moreover, it is an important ingredient to reduce such systems’ po­ submodular which implies that greedy algorithms using these metrics,
tentially high vulnerability with respect to (w.r.t.) natural disasters or e.g., for the optimal placement of High-Voltage direct current lines in a
cyber-attacks [1], enhancing their operational resilience. An increased simplified model of the European power transmission network, have
protection status could be mandated for the identified critical elements, provable suboptimality bounds [9]. Time-varying minimal configura­
to keep the number of outages and failures in this group at a minimum, tions of sensors and actuators can be computed with the help of semi-
see [2] where the hardening of power systems to minimize system definite programming [10]. All these works are valid for linear (dyna­
damage in case of disasters is examined. mical, algebraic) systems without state or input/output restrictions.
Our research question is an instance of the well-known optimal In this contribution, we propose an alternative, novel approach
input/output selection problem, also known as the optimal actuator/sensor based on the steady-state representation of the system only, but con­
placement problem. Starting with classical work on controllability [3] sidering constrained variables. This is an important step towards real
this problem has attracted long-term research attention, in particular, applications where power injections and line flows are always subject
for linear time-invariant systems. The problem has recently become to physical limits.

☆
This work was sponsored by the German Federal Ministry of Education and Research in project AlgoRes, grant no. 01|S18066A. It has been performed in the
context of the LOEWE center emergenCITY.
⁎
Corresponding author.
E-mail addresses: [email protected] (E. Mora), [email protected] (F. Steinke).

https://doi.org/10.1016/j.epsr.2020.106647
Received 26 September 2019; Received in revised form 2 April 2020; Accepted 1 August 2020
Available online 06 August 2020
0378-7796/ © 2020 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
E. Mora and F. Steinke Electric Power Systems Research 190 (2021) 106647

Our approach extends current work on the distributed control of 3. Feasible sets of controllers and measurements
power systems [11]. For instance, the robust optimal power flow al­
gorithm by [12] allows computing set points and droop constants for 3.1. Feasibility conditions & problem statement
some generators while guaranteeing feasible grid operation for all
power injections of other uncertain producers and consumers. While we The power flow model of the previous section can be abstracted as
use similar modeling, we focus on identifying the minimal sets of follows: let x N be the variables that can be set externally. is
controllers and measurements that are required for computing such set assumed to be a product of intervals, i.e., = [ x 1, x1] × × [ x N , xN ].
points. Variables x can be partitioned into the controlled variables xc, for which
The rest of the paper is organized as follows. Section 2 introduces we will design a controller in the following, and the free variables xf,
the employed linear power flow model. The feasibility of a given set of that are left free to be determined either by other users, cooperative or
controllers and sensors is defined in Section 3. We also give a formal malicious, by fixed external conditions, such as, e.g., the weather, or at
problem statement there as well as further computationally advanta­ random. The index set of the controlled variables is denoted by and
geous conditions for testing feasibility. In Section 4, we exploit those the corresponding partitions of as c and f . We assume that the
conditions for developing two efficient algorithms that minimize the variables x determine the system state uniquely and that the set of
number of controllers and sensors. In Section 5, we apply the proposed feasible system states * can be characterized via a set of linear in­
algorithms to find the smallest number of controllers and sensors for 1) equalities,
a simple microgrid consisting of 4 buses and 2) a modified version of
the IEEE 118 bus test case. Finally, concluding remarks and an outlook * = {x : Ax b}, (4)
for future research are provided in Section 6.
where A K ×N
and b K
.
Similarly, we assume a set of possible measurements y L to be

linearly related to the system state, i.e., y = Mx with M L×N

. We
2. Linear power flow
partition these possible measurements into the monitored measurements
ym, that are used as inputs to the control law, and the unmonitored
We analyze an electrical network with N electrical buses connected
variables yu, that are not required for the controller and may or may not
by T transmission lines under the common DC power flow assumptions
be recorded in practice. The index set of the monitored variables is
[13]. The voltage phase angles N determine the nodal active
denoted by .
power injections pI N
and the active power line flows pF T
as
The defined partitions of x and y allow to partition the matrices A
pI = B I , pF = BF , (1) and M along their columns or rows as well, yielding Ax = A c x c + Af xf
and y m = Mmx = Mm c x c + Mf xf .
m

where the entries of B I N ×N

and BF T ×N
are defined element- The aim of the paper is to determine the minimal set of controllers
wise as BI, jk = bjk if j ≠ k, BI, jj = k bjk and BF, jk = bjk , with bjk the and measurements that allows for the design of a control law xc(ym)
susceptance of the line connecting buses j and k. that can guarantee a feasible system state, independently of the state of
Without loss of generality, we assume that exactly one generator or the free variables xf. This can be formalized as follows.
load is connected to each bus, with an externally defined active power Definition 1 (Condition. C1) Sets and are feasible if
set point xi. If the sum of the set points in the grid is not balanced, a
droop-based primary control scheme [13] adjusts power injections pI x c : Mmf ( f) c s.t. xf f:
under adaptation of the frequency to achieve this balance, such that in A c x c (y m ) + Af x f b,
f (5)
steady state we obtain
f = Mf xf .
where y m m
pI = x k . (2)
The idea behind this definition is that the control x c (y m
f ) chosen for
Here, k N
represents the vector of droop constants, ki ≥ 0 and ymf should be valid for the xf from which y m
f originated. Note that we
∑iki > 0, and the frequency deviation with respect to its consider only control values for the steady state of the system in this
nominal value. paper. We do not examine whether and how it is possible to get there
This common setup implies that the measurable quantities pI , pF , from arbitrary initial positions. Moreover, in order to simplify the no­
and Δω are linearly determined by the controllable quantities x. The tation of the involved sets, we have used only a part of ym as input to
kernel of the Laplacian matrix B I contains only the constant vectors for f ) . However, since y = Mc x c + y f one could
the control law x c (y m m m m
m
connected graphs, that is, a constant shift of the phase angles has no easily rewrite the controller into the form xc(y ), i.e., directly using the
impact on pI . We thus fix 1 = 0 and delete the first column of B I to measurements that are actually available to the controller.
obtain B I . The remaining dimensions of θ are denoted by ˜ . We simi­ Since xf uniquely determines y m f , we can also express the control
larly reduce BF to BF . The image of B I moreover contains all vectors law as xc(xf). The formulation x c (y m f ) implies that xc will attain the
with balanced nodal injections. To handle unbalanced set points x, we same value for all values of xf that lead to the same measurements. We
add k as the last column. This lets us compute for all x with · denoting thus obtain the following equivalent condition.
zero entries Definition 2 (Condition. C1’) Sets and are feasible if

pI BI · BI · x c: f c s.t. xf , x ’f f:
1
pF = B · = B · [BI k ] x. A c x c (x f ) + A f x f b
F F
· 1 · 1 (3) x c (xf ) = x c (x ’f ) if Mm m ’
f xf = Mf x f . (6)

In real systems the nodal injections pI will be limited above and

below by the technical capabilities of the connected generator or load.
Valid set points x might be restricted to smaller intervals than the pI , to These definitions allow us to state the optimization task we aim to
leave some space for power generation scheduled by the primary con­ solve in this work.
troller. Similarly, line power flows pF and the frequency deviation Δω Problem statement. Find the set of controllers and measurements
are typically subject to upper and lower bounds. that solves

2
E. Mora and F. Steinke Electric Power Systems Research 190 (2021) 106647

xj , Aij 0
xj = , j .
x j, else (10)
Fig. 1. Relation of the desired conditions C1 and C1’ to the conditions C1*, C2 ,
which are testable with finite resources.
The optimal values thus depend only on the sign of the corresponding
elements of A. In many cases the optimal vectors for different rows of A
will therefore coincide and the cardinality of C A ( f ) is even smaller
min + than its maximum possible value K.
,

s.t. and are feasible w.r.t. C1 or C1’. (7) Definition 5 (Condition. C1*) Condition C1* is fulfilled if
| | and | | denote the cardinality of and . The cost of placing a x c: C A ( f ) c s.t. xf , xf " CA ( f ):
sensor is weighted by 0 ≤ γ ≤ 1 since it will typically be smaller than A c x c (xf ) + Af xf b
implementing a full actuator. x c (xf ) = x c (xf " ) if Mm m "
f xf = Mf xf . (11)
One could additionally incorporate into the objective the varying
efforts and costs for controlling certain elements or acquiring certain
measurements. Instead of just weighting the total number of controllers
Conditions C1/C1’ straightforwardly imply C1* since C A ( f ) f . The
and measurements we would then determine an individual weight for reverse is not always true, as can be shown by counterexample.
each element separately. While we do not follow this idea below, all
However, we will show below that this condition can be exploited for a
algorithms could straightforwardly be adapted. very efficient computation of approximate sets and , at least for the
case when the set of possible measurements consists of the power set
3.2. Related conditions points at each node, i.e., when M is an identity matrix of appropriate
dimensions, here denoted by I. For nodes with zero droop constant, e.g.,
Verifying conditions C1 and C1’ based on their definition requires typical loads or small-scale generators, the measurement of the power
checking infinitely many values of y m set points is equivalent to measuring nodal power injections.
f or xf, respectively. We therefore
derive two related conditions that are testable with finite computa­ Minimizing the objective | | + | | with respect to condition C1* or
tional resources. The relation of all derived conditions is presented in C2 will provide a lower or upper bound for the optimal solution of
Fig. 1. In the next section we then show how to exploit them to effi­ problem (7), respectively. In our experiments we found that for
ciently solve our problem. minimal sets and fulfilling C1* it was often possible to determine
Condition C1 requests the existence of a mapping x c: Mm valid affine-linear control realizations by testing C2 for such sets, i.e.,
f ( f) c
yielding valid control values. One possibility is that this mapping is the upper and lower bound coincided. In this case, and are op­
affine-linear. timal solution of (7).

Definition 3 (Condition. C2 ) Condition C2 is fulfilled if 4. Algorithms

S × ,w s.t. xf f:
x c (y m A c x c (y m The feasibility conditions formulated above enable us to develop
f ) c ) + A f xf b, (8)
two methods for addressing optimization task (7):
f

where y m
f = M f x f and x c (y f ) = Sy f + w .
m m m The first, derived from condition leads to a mixed-integer linear
program (MILP) that finds the smallest feasible sets and , provided
Condition C2 is obviously sufficient for C1. It is, however, not ne­
that M = I . Since C1* is necessary for C1’ but not sufficient, the obtained
cessary as can be shown by counterexample, where piecewise linear
sets and may be too small to be feasible. While we often obtained
control laws sometimes allow for fewer sensors and controllers. The
feasible results anyway, the algorithm can also be used to generate a
condition is testable with finite efforts, as we show in Section 4.
good initial solution for the second approach.
The conditions presented so far are continuous in the sense that
The second method for solving problem (7) is designed for all pos­
testing their validity requires checking an infinite set of possible rea­
sible measurement matrices M. It is a greedy procedure based on hill
lizations of xf or y m
f . However, since the possible values of xf and y f are
m
climbing (HC) and condition C2 . Recalling that condition C2 is sufficient
restricted to bounded polytopes, i.e., f and Mf ( f ), we can derive a
m
for C1 but not necessary, the obtained sets and may possibly be too
necessary condition for C1’ based only on the corners of such polytopes.
large, but are guaranteed to be feasible.
In contrast to C2 , such necessary condition will not assume the control
f ) to be affine-linear.
law x c (y m
4.1. MILP-based approach
Definition 4 (Corner). z is an extreme point or corner of the convex
set if there are no two distinct points z1, z2 and λ ∈ (0, 1) such In this section, we develop a MILP for finding the smallest feasible
that z = z1 + (1 ) z2 . sets and based on condition C1*, provided that M = I . The key is to
formulate condition C1* as a set of linear inequalities that holds for all
Denote C ( f ) as the set containing the corners of f . The number of
choices of sets and .
corners of f , denoted as |C ( f )|, is finite, but grows exponentially with
To this end, consider the binary decision variables uc ∈ {0, 1}N and
the number of free variables. A condition based on all corners of f
um ∈ {0, 1}N, defined element-wise as
would therefore be computationally prohibitive for larger dimensions
of f . Instead, we focus on a subset of corners only, namely those ones 1, j 1, j
which have the maximum impact on the constraints Ax ≤ b. Denote u cj = , u mj = ,
0, else 0, else
such subset by C A ( f ) . Let Ai be the i-th row of A, with i {1, …, K } .
Then, a point xf belongs to C A ( f ) if xf C ( f ) and if there exists for j {1, …N } . The decision variables uc and um encode the elements
i {1, …, K } such that xf is an optimal solution for of and , respectively. Finding the smallest number of elements of
and is thus equivalent to minimizing the cost u c 1 + um 1.
max Aif xf . i
xf C ( f) (9) Let x , i = 1, …, K , be defined element-wise as
x j , Aij 0
Remark 1. Note that the optimization problem (9) defining the x˜ ij = .
x j, else
elements of C A ( f ) can be solved analytically for row Ai as

3
E. Mora and F. Steinke Electric Power Systems Research 190 (2021) 106647

i
x is the Ai-optimal analytical solution of (9) for the case when all Hij ^ (S) x ,
Aij fj
variables are assumed to be free. Moreover, for any given set of con­ ^ (S) x ,
trolled variables , the elements of C A ( f ) can be identified with
Hij A ij fj (14)
i i
x f = x (1 u c), where 1 is a vector of ones of appropriate dimension ^ (S) x is then
for all i = 1, …, K^ and j = 1, …, Nf . The upper bound of A f
and ∘ represents the Hadamard product. Since we assume here that given by H1 and condition (13) is equivalent to
M = I, we can further partition the free variables into monitored and
unmonitored variables, i.e., we can write 1 uc = um + uu, with H1 + Fw l v. (15)
uu ∈ {0, 1}N being the binary vector that encodes the elements of the Putting these results together allows us to compute the minimum
unmonitored variables. The Aj-optimal corners of f can then be possible value of η for given and via the following linear program
j i
identified with x uu , j = 1, …, K . Similarly to x um , a vector of length
min
N whose non-measured entries are zero, we now consider an associated , H, w, S
i
control vector x c for which s.t. H1 + Fw l v,

x uc x ic x u c, H ijA^ (S) x ,
ij fj i = 1, …, K^ , j = 1, …, Nf ,

i.e., x ic is a vector of length N whose non-controlled entries are zero. Hij A^ij (S) x fj , i = 1, …, K^ , j = 1, …, Nf . (16)
i
Condition C1* states that for all i = 1, …, K , the control vector x c for The above described algorithm for testing the validity of C2 for fixed
i
the Ai-optimal corner x f of f should be valid and that it should be and can now be used as a subroutine to minimize over the sets
identical to the control vector for all other corners in C A ( f ) that cannot and as well. To do this, we proceed iteratively from initial sets and
be distinguished given the measurements. We thus can consider only adapting them one element at a time. Since we want to measure the
the worst case of the unknown elements and write compactly optimization progress also for non-feasible combinations and , we
i extend the minimization objective to
Ax ic + A m um + Au uu b,
J( , )= + + µ max( , 0), (17)
i
with à m and Au defined element-wise as, k = 1, …, K ,
where η is the feasibility indicator obtained from solving problem (16).
i
A˜ m, kj = Akj x˜ij , Au, kj = Akj x˜ kj . μ > 0 is a weighting factor that penalizes the infeasibility of and .
We choose μ ≫ 1 to steer the iteration quickly towards feasible solu­
Thus, the mixed-integer linear program that solves (7) when M = I tions.
reads The cost function (17) is minimized via a greedy hill climbing
min uc 1 + um 1
procedure. In each iteration we compute the objective value for all sets
u c, um, uu, x˜ ic ’
or ’ that can be generated by adding one element to either or .
s.t. ˜ im um + Au uu
Ax˜ ic + A b, i = 1, …, K , We then choose the step which yields the largest improvement of the
x u c x˜ ic x u c , i = 1, …, K , objective value (17). As soon as the sets of controllers and measure­
ments are feasible, we stop the iteration.
u c + u m + uu = 1 . (12)
It is well known that the solution of this greedy approach depends
on the selection of the starting point. A natural option is to start with
Remark 2. Note that, particularly in large scale applications, there may
empty sets, selecting the most important controllers and measurements
be several constraints, i.e., rows of A and corresponding entries of b,
during the first iterations. Alternatively, we propose to use the
that are not violated for any realization of x. Hence, when
MILP (12) formulation as an initial guess. More specifically, we solve
optimizing (12), we only take into account the rows of A, for which a
the MILP (12) for M = I first. We then use the found controller set as
violation of (4) is possible, i.e., where Ax
i i bi > 0 . This preprocessing
a starting point for the greedy approach, while disregarding the found
is also utilized by the greedy search proposed below.
measurements. Instead, we start with an empty . This way the
measurements resulting from general M, which potentially allow for
4.2. Greedy approach more compact control systems than the identity measurements, can be
integrated well, but the critical controllers are already identified.
In this section, we first show how to check condition C2 efficiently
Remark 3. Since general MILP has exponential worst-case time
via a linear program (LP) for fixed sets and . Thereafter we de­
complexity, this is an upper bound on the complexity of our first
scribe an iterative algorithm to choose and adapt these sets in order to
approach (12). In contrast, LP as used for our second approach (16) is
find minimal feasible sets.
known to have polynomial worst-case time complexity, and the hill
For given and , condition C2 mandates to check if there exists a
climbing procedure only adds polynomial factors. However, for the
valid affine-linear control law that makes the system feasible for every
realistic examples discussed in the next section we found the MILP
possible value xf f . More precisely, there should exist an affine-
approach to be more efficient than the hill climbing procedure. The
linear control law defined via S and w such that for all xf f we have
latter’s computation time depends strongly on the starting point. For
A c SMmf + Af
the examined medium to large problem instances, it allowed finding
Ac b 1
SMm xf + small, guaranteed to be feasible solutions for and with very
f I w xc · ,
SMm I xc · reasonable efforts. For cases when M = I the MILP solution could often
f
v be verified to be feasible (and thus also optimal) by solving the small
^ (S)
A
F l
(13) LP (16) only once without further adaptation of or . We thus see
where we introduce as an indicator of how far the system is from both algorithms as an important contribution for solving real control
being infeasible. A control law is valid if η ≤ 0. design problems with state constraints.
To tackle condition (13) for all xf f we only need to consider the
maximum of the left hand side expression. Let K^ = K + 2| | be the 5. Numerical examples
number of rows of A ^ (S) and N = N | | the number of free variables.
f
We can introduce an upper bound on A ^ (S) x via a matrix H K^ × Nf , The algorithms developed in 4.1 and 4.2 are now applied to find the
f
whose entries fulfill minimal feasible configuration of controllers and measurements for two

4
E. Mora and F. Steinke Electric Power Systems Research 190 (2021) 106647

Fig. 2. Minimal sets of and for a simple microgrid. The gray squares represent potential controller/measurement locations. The selected controllers and
measurements are highlighted in red and green, respectively. Scenarios (a) and (b) have M = I, whereas line flows and frequency deviation can also be measured in
(c) and (d). For each scenario, the resulting (non-unique) affine-linear control realization is provided below together with the behavior of the potentially active
constraints for all x f f on the right. For scenario (c) with multiple, equivalent optimal solutions, the colored frames denote alternative optimal solutions. The
rationale behind the scenarios is as follows: In (b) the primary control droop is reduced compared to (a). In (c) we allow for additional measurements. In (d) we
reduce the transfer capacity of the middle link to form an additional active constraint.

exemplary power systems. We first demonstrate our setup and typical transmission lines have a power transfer capacity of ± 10 MW, which
effects on a simple microgrid of 4 buses connected in a line. Subse­ is adequate to avoid grid limitations. In scenario (d) we add an active
quently, a modified version of the IEEE 118 bus test case is addressed. line constraint in the middle.
The experiments were performed using an i5 notebook with 8 GB of In scenario (a) where only the power set point at each bus may be
RAM. The algorithms were implemented in Matlab R2018b, using measured, it is sufficient to control the large generator located at bus 4
YALMIP [14] as modeling language and CPLEX 12.9 as LP and MILP for achieving feasible grid operation. The set points of the remaining
solver. smaller generators can be chosen freely and no additional measurement
devices are required.
In scenario (b) we reduce the droop of the generator at bus 4 to 4
5.1. Simple microgrid MW/Hz. This makes the measurement of the power injections at buses 1
and 2 necessary. Although the power injections at buses 1 and 2 can be
Fig. 2 shows the considered microgrid consisting of three generators chosen arbitrarily, they must be monitored so that the power produced
supplying a demand of 5 MW. It gives the topology of the grid together by the generator located at bus 4 can be set appropriately to balance the
with the capacity limits of each transmission line and each generator/ system within the given frequency tolerance.
load. The generator located at bus 4 provides primary reserve, initially In scenarios (a) and (b), where M = I, the solutions of the MILP
with a droop of 12 MW/Hz and later with 4 MW/Hz. The maximum were feasible (and optimal) without further adaptation of and and
allowed frequency deviation is ± 0.1 Hz. We first assume that all

5
E. Mora and F. Steinke Electric Power Systems Research 190 (2021) 106647

the greedy approach, starting from empty sets and , produced the
same results.
In scenario (c) the measurement of the line flows and the grid fre­
quency is added to the set of potential measurements, when performing
the greedy optimization. This allows to reduce the number of mea­
surements to only one. For this scenario the solution is not unique: one
possibility is to take the measurement of the frequency deviation as
controller input, yielding an adapted primary control scheme. An al­
ternative solution that is shown in the figure is to monitor the sum of
the outputs of generators 1 and 2 by measuring the line flow between
bus 2 and 3 for controlling the set point of the generator at bus 4. This
situation will be very common in future active distribution grids, where
individual small scale loads or generators are not able to violate local
grid constraints, but their aggregated effect is important to the system.
Since the load is fixed, measuring the line between buses 3 and 4 would
be equally informative. The feasibility of all these solution candidates
was verified via LP (16), obtaining valid affine-linear control realiza­
tions in all cases.
In scenario (d), we constrain the capacity of the transmission line
connecting buses 2 and 3 to the interval [ 1, 1] MW. This represents an
active grid constraint if the generators at buses 1 and 2 produce at
maximum power. The solution obtained via hill climbing optimization
consists of additionally controlling the power injection at bus 1. Again,
several alternative solutions are possible.
In scenarios (a), (b), and (c), the frequency deviation represents an
active constraint to the operation of the system. Observe in Fig. 2 how
in each case the resulting affine-linear control law keeps the frequency
deviation inside the feasible region for all values of the non-controlled
injections. In scenario (d), the designed controller also ensures feasible
system operation despite the limited power capacity of the middle line.
While for the demonstrated example all solutions can readily be
verified manually, it shows that the situation may become much more
complex in larger grids. The topological location of generators and
loads in the grid is important as well as their capacity and their
neighborhood. An automated algorithm for selecting critical elements
to control and/or measure is thus very beneficial for complex networks
with distributed generation and transmission lines that are operated
close to their technical limits.
In scenarios (c) and (d) the use of the greedy approach is required to
deal with M I . Using the MILP solution as an initial guess for or
starting with empty sets led to the same optimal objective function
value. The solutions for and did not always agree exactly, but Fig. 3. Minimal sets of controllers and measurements for the modified IEEE 118
bus test case. The selected controllers and measurements are highlighted in red
could be shown to be equally optimal.
and green, respectively. (a) Only the nodal power set points may be measured.
The total solver time for all scenarios is shown in Table 1. As ex­
In this scenario, only 12 controllers and 20 sensors are required to guarantee
pected, the MILP optimization performs faster than the hill climbing feasible grid operation. (b) The measurement of line power flows and grid
optimization for the same instances. When computing the optimal sets frequency deviation are additionally considered as possible. In this case only 3
for scenarios (a) and (b), the MILP algorithm was more than 2 times sensors are required. (For interpretation of the references to colour in this figure
faster than the hill climbing with empty sets. It was also 1.2 times faster legend, the reader is referred to the web version of this article.)
than the hill climbing that uses the MILP solution for as initial guess,
which corroborates the benefits of such concatenated optimization see Fig. 3. This power system is composed of 54 generators, 99 loads,
procedure. and 186 transmission lines. The topology of the power system, the load
values and the line and generator capacities were taken from [15]. We
assume that each generator can be scheduled in the range of 10%-90%
5.2. IEEE 118 Bus test case of its available capacity. In addition, we admit 10% of uncertainty for
each load in both directions. The maximum allowed frequency devia­
We now analyze the modified version of the IEEE 118 bus test case, tion is taken as ± 0.2 Hz.
We first consider the case when only the power set points may be
Table 1
measured, i.e., M = I, see Fig. 3a. We obtain an optimal set of 12
Total solver time, in milliseconds, for the proposed optimization algorithms
controller and 20 measurement devices to guarantee feasible grid op­
applied to the simple microgrid.
eration. The remaining 96 injections can be left operating free and/or
Scenario MILP HC (empty sets) HC ( from MILP) be manipulated deliberately and do not require any monitoring
equipment.
(a) 77 160 100
(b) 78 233 135 To obtain this result, we first use the MILP algorithm and then va­
(c) – 277 171 lidate its solution via LP (16). The obtained η is smaller than zero,
(d) – 330 176 thereby proving the feasibility and optimality of the MILP solution.

6
E. Mora and F. Steinke Electric Power Systems Research 190 (2021) 106647

Table 2 consumers. Unlike previous work, we take specific power limitations of

Total solver time, in minutes, for the proposed optimization algorithms applied lines, generators, and loads into account. This step strongly improves
to the modified IEEE 118 bus test case. the applicability in practice, where our approach will help reducing
Scenario MILP HC (empty sets) HC ( from MILP) control costs and efforts and increasing power systems’ resilience.
While we have only considered active power in this work, the ap­
(a) 2.57 13 5.78 proach can straightforwardly be applied to linearized power flow
(b) – 154 28
models taking into account also reactive power and voltages.
Developing a MILP formulation for condition C2 is also possible, but our
When we initialize the greedy search with empty sets, we obtain a experiments so far have not yielded satisfying run times.
feasible solution consisting of 23 controllers and 9 measurements. As
expected, the obtained solution in this case is larger than the one pro­ Declaration of Competing Interest
vided via MILP optimization. This confirms that taking the MILP solu­
tion as initial guess is beneficial for the greedy search. The authors declare that they have no known competing financial
We now add the measurements of the line flows and the frequency interests or personal relationships that could have appeared to influ­
deviation into the set of potential measurements, see Fig. 3b. This yields ence the work reported in this paper.
in total 305 possible sensor devices. We first apply the MILP algorithm
and then the greedy one, starting with the controllers identified via the References
MILP. As expected, the solution is much sparser than before. The total
number of required sensors is reduced from 20 to 3. The selected line [1] A. Abedi, L. Gaudard, F. Romerio, Review of major approaches to analyze vulner­
ability in power system, Reliabil. Eng. Syst. Saf. 183 (2019) 153–172, https://doi.
flows confer a large amount of information that help avoiding grid
org/10.1016/j.ress.2018.11.019.
capacity violations. [2] W. Yuan, J. Wang, F. Qiu, C. Chen, C. Kang, B. Zeng, Robust optimization-based
It is insightful to observe the progress of the hill climbing procedure: resilient distribution network planning against natural disasters, IEEE Trans. Smart
Grid 7 (6) (2016) 2817–2826, https://doi.org/10.1109/tsg.2015.2513048.
buses with major generators connected are selected as controlled nodes
[3] R.E. Kalman, A new approach to linear filtering and prediction problems, Trans.
first. The procedure is thus initially reducing the impact of the free ASME 82 (Series D) (1960) 35–45.
variables on the system by controlling the highest uncertain injections [4] G. Li, L. Deng, G. Xiao, P. Tang, C. Wen, W. Hu, J. Pei, L. Shi, H.E. Stanley, Enabling
first. When enough controlled nodes were selected, the selection of controlling complex networks with local topological information, Nature 8 (4593)
(2018), https://doi.org/10.1038/s41598-018-22655-5.
measurements starts to be significant for the minimization of the cost. [5] S. Pequito, S. Kar, A.P. Aguiar, A framework for structural input/output and control
Selected measurements are often related to nodes connected either to configuration selection in large-scale systems, IEEE Trans. Autom. Control 61 (2)
large non-controlled generators or to high uncertain loads. Remaining (2016), https://doi.org/10.1109/tac.2015.2437525.
[6] , C. T. Lin, Structural controllability, IEEE Trans. Autom. Control 19(3) (1974)
buses with smaller injections are mostly left unobserved. 201–208. 10.1109/TAC.1974.1100557.
Table 2 shows the obtained solver time for all studied cases. The [7] F. Pasqualetti, F. Dörfler, F. Bullo, Control-theoretic methods for cyberphysical
solution for case (a) using MILP optimization was found in about 2.57 security: geometric principles for optimal cross-layer resilient control systems, IEEE
Control Syst. Mag. 35 (1) (2015) 110–127, https://doi.org/10.1109/MCS.2014.
min. Observe that the solution for case (b) was computed in about 2364725.
28 min for the concatenated execution of both algorithms, compared to [8] G. Lindmark, C. Altafini, Minimum energy control for complex networks, Nature 8
the ca. 154 min needed by the solver when starting hill climbing with (1) (2018), https://doi.org/10.1038/s41598-018-21398-7.
[9] T.H. Summers, F.L. Cortesi, J. Lygeros, On submodularity and controllability in
empty sets. A single verification step using LP (16) took less than a
complex dynamical networks, IEEE Trans. Control Netw. Syst. 3 (1) (2016) 91–101,
second. https://doi.org/10.1109/TCNS.2015.2453711.
The computation time could further be improved, e.g., by testing [10] A. Taha, N. Gatsis, T. Summers, S. Nugroho, Time-varying sensor and actuator se­
lection for uncertain cyber-physical systems, IEEE Trans. Control Netw. Syst. 6 (2)
not all possible set extensions in each step of the greedy search but
(2019), https://doi.org/10.1109/tcns.2018.2873229.
using only a representative subset, selected by proximity in the graph. [11] D.K. Molzahn, F. Drfler, H. Sandberg, S.H. Low, S. Chakrabarti, R. Baldick,
Another idea would be to add more than one element in each iteration. J. Lavaei, A survey of distributed optimization and control algorithms for electric
For the control design task described in this paper, however, the power systems, IEEE Trans. Smart Grid 8 (6) (2017) 2941–2962, https://doi.org/
10.1109/tsg.2017.2720471.
achieved computation time seemed acceptable even without these ex­ [12] A. Mešanović, U. Münz, C. Ebenbauer, Robust optimal power flow for mixed AC/DC
tensions. transmission systems with volatile renewables, IEEE Trans. Power Syst. 33 (5)
(2018) 5171–5182, https://doi.org/10.1109/TPWRS.2018.2804358.
[13] P. Kundur, N. Balu, M. Lauby, Power System Stability and Control, EPRI Power
6. Outlook System Engineering Series, McGraw-Hill Education, 1994.
[14] J. Löfberg, Yalmip : A toolbox for modeling and optimization in Matlab,
The theoretic framework and the algorithms developed in this work Proceedings of the CACSD Conference, (2004). Taipei, Taiwan
[15] A.R. Al-Roomi, Power Flow Test Systems Repository, 2015.
allow for the efficient identification of critical controllers and mea­
surements in complex power systems with uncertain producers and