Global Solution of Economic Dispatch with Valve

Point Effects and Transmission Constraints

Loı̈c Van Hoorebeeck Anthony Papavasiliou
P.-A. Absil Center for operations research and econometrics
ICTEAM Institute UCLouvain
UCLouvain Louvain-la-Neuve, Belgium
Louvain-la-Neuve, Belgium
loic.vanhoorebeeck@uclouvain.be

Abstract—Gas units are becoming an increasingly output of gas units. A typical model for a fuel cost
important component of modern power system op- function that accounts for the valve point effect is the sum
erations, due to the flexibility that they offer in a of a quadratic component with a non-smooth and non-
regime of large-scale renewable energy integration. The
valve point effect refers to the loss of efficiency when convex rectified sine function, as detailed in eq. (5). The
operating a turbine off a valve point, that is just after operational constraints of the considered problem include
the previous valve opens. The valve point effect for gas power-range restrictions and ramp constraints. We further
power plants is a primary attribute of their operation, consider transmission constraints, which are represented
however its representation raises computational chal- using the DC optimal power flow model.
lenges, due to the non-convex and non-smooth model
that is required for representing the fuel cost. In this While the valve point effect has been accounted for in
work, a new heuristic based on successive piecewise various economic dispatch models in the literature (see,
approximations of the cost function is described. This e.g. [2], [3], [4]), few studies investigate its impact in
heuristic consists in two steps: first, the optimization the context of optimal power flow (OPF) such as [5].
is run over the whole feasible set and a lower bound Besides, most approaches are based on heuristics and do
for the optimal objective is obtained, then the feasible
solutions collected in the previous step are enhanced not show any guarantees with respect to the optimality of
through local searches. The approach is tested on sev- the returned solution. In [6], a globally convergent method
eral IEEE bus systems that have been extended with relying on piecewise-quadratic approximations was devel-
generators obeying a valve point effect. oped. This method was extended in [7] to multi-periodic
Index Terms—Economic dispatch, global optimiza- cases but suffers from a long execution time.
tion, mixed-integer programming, non-convex opti- In this work, we extend this approach to handle network
mization, non-smooth optimization. constraints, and in order to remedy the long execution
I. Introduction time, we describe a heuristic able to capture solutions of
Gas units are gaining importance as components of similar quality in a reduced amount of time. We demon-
modern power system operations, as a result of the adapt- strate the effectiveness of our proposed heuristic in a
ability that they offer in systems with significant levels of transmission-constrained IEEE network, and illustrate the
renewable energy integration. Various market operators, notable impact of accounting for the valve point effect in
such as the Midcontinent ISO, have made some advances systems with significant levels of gas unit integration: the
in the representation of detailed and complex operating valve point effect substantially affects the dispatch, and
and cost constraints of gas units in their unit commitment neglecting it can lead to significant short-term operational
and economic dispatch models. The focus of the present efficiency losses. These methods can be easily extended to
paper is the valve point effect (VPE) in the context of a more complicated models that include spinning reserves,
multi-period economic dispatch model, which serves as a multiple fuels and prohibited operating zones.
starting point for the modelling of this important oper- The approach followed here, and introduced in [6], [7],
ational attribute in more advanced operational planning can be summarized as follows. First a piecewise linear
models. or quadratic under-approximation of the non-smooth and
Commonly, fuel costs in economic dispatch are modelled non-convex cost functions is constructed around a chosen
as smooth quadratic functions. Such convex models do set of knots, where the knots refer to the points where
not account for the valve point loading effect: a turbine the pieces meet. This new objective function, as well
which is loaded at a valve point, i.e., just before the as the operational and transmission constraints of the
next valve opens, operates at full efficiency and a turbine problem, constitute the initial surrogate economic dis-
operating off a valve point is working less efficiently due patch problem. This surrogate problem is then fed into
to throttling losses [1]. This effect significantly alters the a mixed-integer programming solver, e.g., Gurobi, which

21st Power Systems Computation Conference Porto, Portugal — June 29 – July , 2020
PSCC 2020
returns a solution along with a lower bound to the global readily extended as the minimization of the non-convex
objective. As the feasible set of the problem is represented and non-smooth objective,
exactly, any solution of the surrogate problem is feasible XX
(DED-DCOPF) min fg (pgt ) , (4)
for the true problem. Moreover, any lower bound of the
t∈T g∈G
surrogate economic dispatch model is a valid lower bound
of the economic dispatch problem, because the surrogate with cost functions [8],
problem works with an under-approximation of the actual
fg (pgt ) = Ag p2gt + Bg pgt + Cg + Dg sin Eg (pgt − Pg− ) .
 
problem that we wish to solve. Hence, an interval in which | {z } | {z }
the globally optimal cost of the economic dispatch problem :=fgQ (pgt ) :=fgV (pgt )
must lie, is obtained. If the prescribed tolerance is not met, (5)
the solution of the surrogate economic dispatch problem Here, pgt represents the production of generator g at time
is added to the set of knots. This defines a new piecewise t; Ag , Bg , Cg , Dg , Eg correspond to cost parameters, and
objective which is in any point superior or equal to the first Pg− is the minimum operating range of the plant. The
surrogate function. The algorithm stops either when the following constraints apply:
target tolerance is attained, or when the surrogate function • Power generation restrictions
matches the true function at the optimum of the former.
In this way, any tolerance up to the solver accuracy can Pg− ≤ pgt ≤ Pg+ , (6)
be attained. • Ramp rate limits
The structure of the paper is as follows. In section II the
main and surrogate problems are introduced. The global Rg− ≤ pgt − pg(t−1) ≤ Rg+ , (7)
method is described in section III, and two versions of a • Flow constraints, eqs. (1) to (3).
heuristic are defined. Then, the method and both heuris-
In these constraints, Pg+ is the max power generation
tics are tested on two IEEE test systems in section IV.
limits of unit g, and Rg− and Rg+ are the lower and upper
Finally conclusions are drawn in section V.
ramp limits, respectively. Note that the angles at each bus,
II. Mathematical formulation of DED-DCOPF θ, and the flows over each line, e, can be fully determined
from the production at each node n. For the sake of
A. Main problem: DCOPF based on reactance
simplicity, the main problem (DED-DCOPF) is denoted
We employ a bus angle model of the DC optimal power as (P) in the following. The surrogate problem, introduced
flow (DCOPF) problem, where the network is described hereafter, is denoted as (S).
as an oriented graph. The parameter Bk corresponds to
the reactance of a line k = (m, n). The flow of power, ekt , B. Surrogate problem
along line k at time t, is then described as a function of Due to the linearization of the flow constraints via the
the bus angle difference along the nodes that the line is DCOPF model, the feasible set of (P) is a polyhedron
joining: and the main difficulty in solving the problem comes
ekt = Bk (θmt − θnt ) , (1) from the non-convex and non-smooth objective. Thus, in a
similar fashion as [6], a surrogate problem (S) is defined by
where θmt stands for the angle of the voltage phasor of approximating the objective without changing the feasible
bus m at time step t. We fix the angle of the reference set, namely by employing the following objective function:
bus (indexed by 0), i.e., θ0t = 0 for all t. If losses are XX
neglected, the main addition to the common economic (S) min hgt (pgt ) , (8)
dispatch models, that are employed in the literature for t∈T g∈G
accounting for the valve point effect, is in the power subject to the constraints of eqs.(1) to (3), (6) and(7).
balance constraint:
X X X Given the set of knots Xgt := Xgt1 , . . . , Xgtnknotgt
, the
− pgt − ekt + Dnt + ekt = 0 . (2) terms of eq. (8) read as
g∈Gn k=(·,n) k=(n,·)
fˆg (pg ; Xgt ) if fgV 6= 0 ,
(
where Gn is the set of generators at bus n and Dnt the hgt (pgt ) := (9)
fgQ (pg ) else.
demand of bus n at time t.
This constraint must hold for each bus n, and in the Here, fˆ(• ; X) stands for the piecewise-linear interpolation
multi-period case, for each time period t. Given a flow of f , given the knots X. The smooth part of the objective,
limit T Ck for line k, the flow of power along each line is fgQ , and the non-smooth part, fgV , are defined as in eq. (5).
constrained as follows: Fig. 1 depicts the main and surrogate objective for a
−T Ck ≤ ekt ≤ T Ck . (3) given unit g and time step t. The initial objective, fg , is
plotted in solid line and the smooth part, fgQ , in dash-
With these three additional constraints, the dynamic eco- dotted line. The points where both curves meet are the
nomic dispatch (DED) problem studied in [7] can be points at which the sine from the non-smooth part, fgV ,

21st Power Systems Computation Conference Porto, Portugal — June 29 – July , 2020
PSCC 2020
vanishes and therefore at which the initial objective is not time and avoid timeouts. The comparison of the different
smooth due to the absolute value in eq. (5). We refer to methods in section IV shows that the heuristics reach
these points as kink points. As explained in section III, the comparable solution in a significantly reduced execution
approach considered here consists in successively adding time.
knots to the piecewise approximation to refine it. These
knots, Xgtk
, which define the piecewise approximation, hkgt , III. Methods
at step k of the algorithm are depicted as bullets and This section is devoted to the description of two meth-
include the kink points. The colour difference illustrates ods. The first one (section III-A) is largely similar to
when the points were added to the set of knots: the red the method given in [7], the difference being that, in the
ones belong to the initial set of knots, the green have been present paper, the network is accounted for. In section IV,
added during previous iterations and the purple is the – we demonstrate that network effects exhibit a rich inter-
possibly inexistant – knot to be added. Let us now consider play with the valve point effect. Therefore, it is important
the choice of initial knots and the monotonic property of to consider the representation of the valve point effect in
two successive approximations. future dispatch models. However, the size of the systems
considered in the present work renders the approach of [7]
non-viable, therefore in this paper we develop a second
method. The second method (section III-B) is a local
heuristic based on the former approach, and trades opti-

δ mality guarantees for an acceleration in the computation

time. We detail two variants of the heuristic.
Fuel cost, $

δ A. A globally convergent method

1) Method description: We refer to the globally con-
vergent method proposed in this paper as APLA. APLA
With VPE
stands for adaptive piecewise-linear approximation. The
Without VPE method starts with a set of knots X 0 (red dots in Figs. 1
Piecewise and 2) satisfying the property described in II-B1, i.e., the
approximation inclusion of the kink points in the set of knots. The first
surrogate problem, (S)0 , which is defined by the first sur-
Power Output, MW
rogate function, h0 , is then formulated as a mixed-integer
problem (MIP) and solved with a predefined tolerance γ.
Optimal solution of (S)k : (x∗ k ∗
k , h (xk )) The MIP solver returns a solution, p0 , along with a lower
Initial set of knots: X0
bound to the global optimum. The surrogate gap, δ 0 , is
Previously added knots: (x∗ ∗
l , f (xl )) for l = 1 . . . k − 1
then computed as,
Knot to be added: (x∗ ∗
k , f (xk ))

δ 0 = f (p0 ) − h0 (p0 ) . (11)

Fig. 1. Illustration of the method APLA at iteration k on a single
term of the objective. This gap is visualized in Fig. 1. Then, the optimality gap
is computed by adding the solver tolerance and surrogate
1) Choice of the initial knots: In order for the inter- gap. If the target accuracy is reached, the algorithm
polant to be lower than the original function that we wish stops. Otherwise, the obtained solution p0 is added to
to approximate, it can be shown that the kink points the set of knots, refining in this way the approximation,
should be included in the set of knots [7]. However, this and the algorithm iterates. This process is illustrated
gives a poor initial approximation: the interpolant being in Fig. 2. The red and green dots represent the initial
the zero function. Indeed, since fgV vanishes at the kink and already added knots, respectively. The purple square
points, X kink , we have fˆgV (pg ; X kink ) = 0 for all pg . is the optimal solution of the surrogate problem and the
Therefore, the maxima of fgV are also added to the initial purple dot is the evaluation of the real objective at this
set, Xgt
0
. This gives the red dots of Fig. 1. solution. It can be seen that the method locally refines
2) Monotonic property: If fg is piecewise-concave be- the approximation. In this sense, the method is adaptive
tween two kink points, we have and benefits from a lower number of knots with respect
X 1 ⊆ X 2 =⇒ fˆg (pg ; X 1 ) ≤ fˆg (pg ; X 2 ) ≤ fg (pg ) (10) to a regular meshing. The efficiency gains of the method
rely on quickly converging to a subset of the feasible space
given any feasible point pg and set of knots X 1 containing where the global optimum should lie. This follows the
the kink points of fg . philosophy of other methods used in dispatch algorithms,
The novel aspect of the present paper relative to [7] such as stochastic dual dynamic programming (SDDP) [9],
is the addition of the network constraints, as well as that also aim at computational savings by using locally
the development of heuristics which reduce the execution valid representations of the objective function that is being

21st Power Systems Computation Conference Porto, Portugal — June 29 – July , 2020
PSCC 2020
optimized, with the purpose of quickly limiting the search B. A local heuristic
to the relevant part of the feasible space by employing 1) Method description: Fig. 3 summarizes the heuristic
the information contained in the approximation of the which can be split into two steps: first a global search is
objective function. made on the whole feasible set in order to find optimum
candidates. Then, the search is refined around these can-
Start didates. In [10], the authors proposed to use the solution
of the initial surrogate problem as the initial point of an
interior point method. Here, we follow the same approach
Initialize set of knots (•)
with three main differences: a) the initial point is obtained
with a few steps of APLA, b) a list of candidates is
Solve surrogate problem
S (S)
defined with knots (• •) considered and c) the local search method is also based
on a local approximation. Let us explain more formally
these three points.
Is
no Refine: add solution
a) Search for an initial point: The first step of this
tolerance of (S), () , to knots heuristic is to select a promising candidate, around which
reached?
the local search will be initialized. In order to achieve
yes this, the APLA algorithm is used with a finite number
Stop of iterations, niter , over the whole search space Ω, i.e.,
the feasible set of (P). This number is typically chosen
Fig. 2. Flow chart of the method APLA. to be very small to avoid excessive running time. In the
experiments below, we take niter = 2.
2) Convergence guarantees: At iteration k of the b) List of candidates: During the search for an initial
method, the optimality gap can be computed as [7] point, several potential candidates for the global solution
of the surrogate problems are found by the MIP solver.
Optimality gap
z }| { Most of these points, called incumbents, are good initial
f (pk ) − f (p∗ ) ≤ δ k + γ k + k , (12) guesses for a local search. Hence, APLA is slightly mod-
ified to return a list LI of the best incumbents, which
where p∗ is the optimal solution of (P), γ k stands for the will serve as initial guesses. Using a list of initial guesses
gap of (S)k returned by the MIP solver, and k is an instead of a single starting point reduces the sensitivity of
over-approximation error. This over-approximation error the method with respect to the initial point.
is almost equal to zero for the cost functions considered c) Local search method: The local search in the neigh-
here (see [7]), hence we will neglect it in the rest of the bourhood Ωp̃i , of a specific point p̃i , at iteration i of
paper. The following result, given in [7], also still holds the inner loop, proceeds as follows: the power generation
because it only relies on the Lipschitz continuity of the constraint of eq. (6) is narrowed around the value p̃i via
objective functions and on the boundedness of the feasible the closest knots. Eq. (6) now reads
set. i
Xgtj i
≤ pgt ≤ Xgtj 0 (13)
Theorem 1. For Lipshitz continuous cost function f , the
sequence of iterations provided by APLA satisfies for j < j 0 and Xgtj i i
≤ p̃igt ≤ Xgtj 0 . In this way, the

feasible set is strongly reduced and the power ranges

lim δ k = 0 . become t−dependent. In the specific case when j 0 = j + 1,
k→∞ the (first) local problem becomes a much simpler convex
quadratic problem which can be solved efficiently.
3) Stopping criterion: The algorithm terminates when
2) Convergence guarantees: This APLA-based is con-
the optimality gap, eq. (12), is lower than a predefined
verging because every APLA call inside the loop (Fig. 3)
tolerance γ: δ k + γ k ≤ γ. This γ is chosen to be equal to
converges by theorem 1, and the number of loop calls is
the targeted MIP gap, i.e., the gap between the lower and
equal to the finite size of LI . However, the lower bound
upper objective bound of the surrogate problem, which
obtained on sub-instance Ωp̃i is not a lower bound for (P).
is given as input to the MIP solver. Theorem 1 shows
As we neglect the over-approximation error, the optimality
that the method converges but the drawback is that the
gap is computed as follows,
method requires several costly calls to the MIP solver,
which increase the computation time of the algorithm. Optimality gap

This motivates the development of a heuristic, described

z }| {
f (p ) − f (p ) ≤ f (p ) − (h
i,k̃ ∗ i,k̃ niter
(p niter
)−γ niter
) , (14)
in the following section, as well as a time criterion based
on the number of iteration and the maximum time allowed where (pi,k̃ )k̃=0, 1... and (pk )k=0,1,...,niter are the sequences
to the MIP solver. This time criterion is also used in the of solutions from APLA(Ωp̃i , ∞) and APLA(Ω, niter ).
experiments section IV. Eq. 14 shows that once the heuristic enters the inner loop

21st Power Systems Computation Conference Porto, Portugal — June 29 – July , 2020
PSCC 2020
 Start variables and further motivate the use of an APLA-based
heuristic.
Run APLA(Ω, niter ) IV. Test case study
LI
In this section, the comparison is made between the two
Select p̃i = arg min f (LI )
and remove it from LI
approaches detailed in section III and the benefit from
taking the VPE into account is estimated. In this work,
Run APLA(Ωp̃i , ∞) Gurobi 8.1 [11] has been used for solving the MIP problem
associated with the surrogate problem. The optimization
has been run on a computer with Intel-i7 3.6 GHz CPU
Is tolerance
and 16 GB of RAM. Two variants of the heuristic from
no section III-B are tested: H-local, which restricts the feasi-
reached or Update best solution
LI = ∅? ble region to a single segment, i.e., j 0 = j +1 from eq. (13),
and H-full, which restricts the feasible region to up to
yes three segments. The impact of the VPE is highlighted by
comparing the solution obtained via the aforementioned
Stop
methods with respect to the solution from a method which
Fig. 3. Flow chart of the APLA-based heuristic. does not take the VPE into account.
The data and algorithm implementations are available
on GitLab [12].
in Fig. 3, it reduces the optimality gap only by decreasing
the objective. A. Data set creation
An expression similar to eq. (12) can also be obtained, As no data set with transmission constraints and VPE is
openly available in the literature, we analyse the IEEE test
f (pi,k̃
) − f (p ) ≤ f (p
∗ i,k̃
) − (h niter
(p niter
)−γ niter
),
systems, by introducing additional generators that obey
= f (pi,k̃ ) − f (p∗,i ) + γ niter VPE. We use IEEE test system data from the PSTCA
+ f (p∗,i ) − hniter (pniter ) , and MatPower ([13], [14]), while VPE generator data can
| {z } be found in [15]. The VPE generators are added randomly
:=ζ i
in buses of the IEEE networks which are represented
≤ δ i,k̃ + γ i,k̃ + i,k̃ + γ niter + ζ i , in Figs. 4 and 7. In order to limit the uncertainty from
≈ δ i,k̃ + γ niter + ζ i , (15) this random selection, 100 trials of the algorithm are made
with different network configurations. This allows for a
where p∗,i := arg minp∈ Ωp̃i f (p). The second inequality more robust analysis of the methods. We further assess
holds because the application of APLA on the restricted the scalability of the method by increasing the number of
domain Ωp̃i is a valid instance. Hence, eq. (12) can be used. VPE units. In particular, we double the number of VPE
Furthermore, the over-approximation error is negligible units, and introduce a 5% variation on the parameters of
here and the restricted instances APLA(Ωp̃i , ∞) are solved these VPE units. We include this variation in order to
to optimality, i.e., with a tolerance γ i,k̃ ≈ 0. avoid symmetry, which creates unnecessary computational
Eq. 15 shows that at iteration k̃ of APLA(Ωp̃i , ∞), complexity, since it is rarely the case that different physical
there are three main contributions to the optimality gap. assets share an identical set of economic and technical
Firstly, δ i,k̃ , which is the gap between the surrogate and parameters. We denote as experiment the simulation of
true function. This gap goes to zero as k̃ goes to ∞ by 100 different instances of the problem.
theorem 1. Secondly, γ niter , the MIP gap obtained at the
end of APLA(Ω, niter ) and finally, ζi , which captures the B. IEEE cases with 10 added VPE units
fact that the optimal solution p∗ may lie outside of Ωp̃i . 1) IEEE 57-bus system: The original case study with
By construction pniter ∈ LI , and in the special case seven generators [13] is extended by including ten ad-
where p̃i = pniter , we have f (p∗,i ) ≤ f (pniter ), which ditional generators that obey a valve point effect. The
implies that ζ i ≤ δ niter : the heuristic either improves pniter network topology is sketched in Fig. 4. The tolerance of
or shows that it is globally optimal. the MIP solver and the maximal MIP solver time are set
3) Stopping criterion: The heuristic terminates either to 0.1% and 45 seconds when the surrogate problem is
if it reaches the targeted optimality gap, eq. (14), or after defined over the whole domain Ω, and to 0.01% and 45
iteration over the whole list LI . seconds when the problem is restricted to a subdomain
4) Extension of the methods: Similarly as APLA, the Ωp̃i . Figs. 5 and 6 (blue boxes) present a comparison
heuristic can be easily extended to account for spinning between the execution time and the final optimality gap
reserves, multiple fuels, and prohibited operation zones of the three methods: H-full, H-local and APLA. Whereas
(POZ). This will incur an increase in the number of integer some instances of the initial APLA algorithm can be

21st Power Systems Computation Conference Porto, Portugal — June 29 – July , 2020
PSCC 2020
resolved in a few seconds, others require significantly more
Case57
time, due to timeouts. The timeout limit for APLA is set 0.7 Case118
at a maximum of ten iterations of 45 seconds.
The two heuristics, H-full and H-local, perform much 0.6

Optimality gap (%)

faster, and the majority of instances is executed within 0.5
20 seconds (H-local) and 100 seconds (H-full). This time
improvement is at the cost of an increase in the optimality 0.4
gap, with about half of the instances being solved at 0.3
the target tolerance of 0.1%. The bottom magnification
in Fig. 5 shows that the median execution time of H-local is 0.2
approximately four times lower than the median of H-full, 0.1
while the optimality gap, depicted in Fig. 6, is comparable.
We discuss the implication of these observations in detail 0.0
in section IV-E. H-full H-local APLA

Fig. 6. Comparison between the final optimality gap of the meth-

ods proposed in the paper for 100 random instances with targeted
optimality gap of 0.1%.

same methods and parameters as in section IV-B1. The

comparison of the execution time, Fig. 5 (green boxes),
shows that APLA performs faster in the IEEE 118 case but
it is still outperformed by the two heuristics. However, the
final optimality gap (Fig. 6) and final objective (table I)
is slightly better with APLA. We discuss the implication
of these results in detail in section IV-E.

Fig. 4. Single-line diagram of IEEE 57 bus system [16].

Case57
Execution time (s)

400 Case118
300

200

100

120
Execution time (s)

100
80
Fig. 7. Single-line diagram of IEEE 118 bus system [17].
60
40
20
C. IEEE cases with 20 added VPE units
0
H-full H-local APLA
As observed in the previous section, increasing the size
of the network while keeping the number of VPE units
Fig. 5. Comparison between the execution time of the methods for constant does not pose significant computational compli-
100 random instances with targeted optimality gap of 0.1%. The cations. On the contrary, a larger network may result in
orange lines and green triangles represent the medians and means,
respectively. The bottom frame is a magnification of the upper one. reduced run times. In this section, we increase the size of
the problem by adding ten additional VPE units to the
2) IEEE 118-bus system: In a similar way as the pre- systems.
vious case, we extended the 54-generator IEEE 118-bus 1) IEEE 57-bus system: We perform tests on an ex-
system (Fig. 7) with the same ten generators obeying tension of the 57-bus case using the same methods and
a valve point effect [13]. Tests are performed with the parameters as in section IV-B1. We report the results in

21st Power Systems Computation Conference Porto, Portugal — June 29 – July , 2020
PSCC 2020
Figs. 8 and 9. Compared to the previous tests, all three D. Impact of the VPE
methods exhibit a similar relative increase in run time and In order to assess the benefit of accounting for the VPE,
final optimality gap. However, the absolute values of the each instance of both case studies is solved, while ignoring
running time and optimality gap remain acceptable for the VPE, by specifically removing the rectified sine term.
real time applications. The performance of the methods As a result of this simplification, the objective function be-
relative to each other is similar to that of the previous case, comes quadratic and the problem becomes much simpler.
and the discussion is then analogous to section IV-B1. The computation time is less than a second, but there is
no guarantee with respect to the optimal solution. The
mean of the final objective is reported in table I, for this
Case57
method (QP) as well as for APLA and both heuristics.
Execution time (s)

400 Case118
300 The cost of neglecting the VPE can be readily estimated
200
at 5% and 1% for the IEEE 57 and IEEE 118-bus system
case with 10 additional VPE units, respectively. The
100
second experiment, including 20 additional VPE units,
0
shows larger discrepancies with 8% and 3.3% differences.
120
Execution time (s)

100 TABLE I
80 Objective mean of each method.
60
H-full H-local APLA QP
40
20 Case57 (10 VPE-units) 621622 621634 621547 654535
Case118 (10 VPE-units) 2447475 2447540 2447383 2471132
0
H-full H-local APLA Case57 (20 VPE-units) 587695 587713 587669 634653
Case118 (20 VPE-units) 2227536 2227592 2227087 2301490
Fig. 8. Comparison between the execution time of the methods for
100 random instances of IEEE test cases with 20 additional VPE
units and a targeted optimality gap of 0.1%. E. Discussion
The application of the APLA method on a network-
constrained case study is compromised by the fact that a
significant number of instances terminate due to timeouts.
1.0 Case57
Case118 These timeouts push up the mean of the execution time.
However, the heuristics based on the former method are
0.8 able to reach at least 0.3% optimality gap for 75% of the
Optimality gap (%)

instances in reduced time, relatively to APLA.

0.6 Also, APLA reduces the optimality gap by increasing
the lower bound, on the one hand, and by decreasing the
0.4 optimal objective, on the other hand. On the contrary,
the heuristic approaches compute a single lower bound
and then focus on the objective; for the same optimality
0.2
gap, the solution of the heuristic is therefore lower. This
phenomenon is highlighted by the fact that even if the final
0.0 optimality gap depicted in Figs. 6 and 9 is approximately
H-full H-local APLA 10% better for APLA, the final APLA objective reported
in table I is lower than the heuristics by only ∼ 0.01% on
Fig. 9. Comparison between the final optimality gap of the methods average. Of the two variants of the heuristic considered, H-
for 100 random instances with 20 additional VPE units and a local, which restricts the local search on a smaller subset
targeted optimality gap of 0.1%.
than H-full, performs slightly faster.
The scalability of the method is assessed in two stages.
2) IEEE 118-bus system: We perform tests on an ex- Firstly, the number of buses and generators is increased,
tension of the 118-bus case using the same methods and while keeping the number of VPE units constant, and then
parameters as in section IV-B1. In this case, the median this number is doubled. The first experiments (Figs. 5
final optimality gap doubles. However, the heuristics still and 6) have shown that the problem does not become
outperform APLA execution time and the difference be- more difficult when increasing the number of nodes and
tween both heuristics is reduced, with a faint dominance usual thermal units; the problem becomes even easier
of H-local. Further comments on this increase are made when the number of nodes is increased from 57 to 118.
in section IV-E. The second experiments (Figs. 8 and 9) show that this is

21st Power Systems Computation Conference Porto, Portugal — June 29 – July , 2020
PSCC 2020
not always true: with 20 additional VPE units, the 57- With the aim of improving the practicality of the model,
bus system becomes easier than the 118-bus system. In some direct extensions could be contemplated. Firstly, the
general, the execution time increases and the quality of method presented here can be adapted to account for
the solution worsens when increasing the number of VPE quadratic power losses, through a piecewise relaxation of
units. However, despite this time and gap increase, the the losses. Secondly, a more realistic power flow model,
values of these quantities remain under two minutes and such as a convex ACOPF, could also be used. The main
0.25 %, respectively, for most instances. change would be to resort to a mixed-integer second order
As shown in the experiments, the core difficulty of the cone programming (MISOCP) solver instead of a MIP one.
problem lies rather in the number of VPE units than in the
Acknowledgment
total number of units. This motivates the application of
the heuristics on real test cases where only a small fraction This work was supported by the Fonds de la Recherche
of the units obeys a valve point effect. Scientifique - FNRS under Grant no. PDR T.0025.18.
Finally, the cost of the valve point effect computed for References
both test cases demonstrates the interplay of the network
[1] G. L. Decker and A. D. Brooks, “Valve point loading of tur-
with the VPE; ignoring the VPE costs on average 5% bines,” Electrical Engineering, vol. 77, no. 6, pp. 501–501, Jun.
for the 57-bus case and 1% for the 118-bus case. This 1958.
difference can be explained by the distinct level of VPE- [2] K. P. Wong and C. C. Fung, “Simulated annealing based eco-
nomic dispatch algorithm,” IEEE Proceedings C - Generation,
units in the two test cases: more than half of the units Transmission and Distribution, vol. 140, no. 6, pp. 509–515, Nov
obey a valve point effect for the modified IEEE 57 case, 1993.
while this number is reduced to fifteen percent in the [3] L. S. Coelho and V. C. Mariani, “Combining of chaotic differen-
tial evolution and quadratic programming for economic dispatch
second case. These numbers increase for the case with optimization with valve-point effect,” IEEE Transactions on
20 additional VPE units. The average relative cost of power systems, vol. 21, no. 2, pp. 989–996, 2006.
neglecting the VPE units is estimated at 8% and 3% for [4] M. J. Mokarram, T. Niknam, J. Aghaei, M. Shafie-khah, and
J. P. Catalao, “Hybrid optimization algorithm to solve the non-
the 57-bus and 118-bus case, respectively. convex multiarea economic dispatch problem,” IEEE Systems
Journal, 2019.
V. Conclusion [5] T. Niknam, M. R. Narimani, and R. Azizipanah-Abarghooee,
“A new hybrid algorithm for optimal power flow considering
In this paper, a global method for the solution of the prohibited zones and valve point effect,” Energy Conversion and
multi-period economic dispatch problem with transmission Management, vol. 58, pp. 197 – 206, 2012.
constraints has been studied. This method, the adaptive [6] P. A. Absil, B. Sluysmans, and N. Stevens, “MIQP-based al-
gorithm for the global solution of economic dispatch problems
piecewise-linear approximation (APLA), suffers from long with valve-point effects,” in 2018 Power Systems Computation
execution time when applied to larger problems with Conference (PSCC), June 2018, pp. 1–7.
network constraints, and a ten-minute run time threshold [7] L. Van Hoorebeeck, A. Papavasiliou, and P.-A. Absil, “MILP-
based algorithm for the global solution of dynamic economic
(which is a reasonable upper bound on economic dispatch dispatch problems with valve-point effects,” IEEE Power and
models in European market operations) is attained several Energy Society General Meeting, 2019.
times. This motivates the development of a faster method, [8] D. Walters and G. Sheble, “Genetic algorithm solution of eco-
nomic dispatch with valve point loading,” IEEE Transactions
which still targets the global solution of this non-convex on Power Systems, vol. 8, no. 3, pp. 1325–1332, Aug. 1993.
problem and provides guarantees with respect to the final [9] M. V. F. Pereira and L. M. V. G. Pinto, “Multi-stage stochastic
solution. optimization applied to energy planning,” Mathematical Pro-
gramming, vol. 52, no. 1, pp. 359–375, May 1991.
Our main contributions are i) to demonstrate that VPE [10] S. Pan, J. Jian, and L. Yang, “A hybrid MILP and IPM ap-
matters in test instances with network constraints and ii) proach for dynamic economic dispatch with valve-point effects,”
to develop local heuristics which accelerate the execution International Journal of Electrical Power & Energy Systems,
vol. 97, pp. 290 – 298, 2018.
time of the problem. [11] Gurobi Optimization Inc, “Gurobi Optimizer Reference
The APLA method, as well as the heuristics, are tested Manual,” 2018. [Online]. Available: http://www.gurobi.com
on several instances of the IEEE 57-bus and IEEE 118- [12] L. Van Hoorebeeck, “Adaptive piecewise approximation,”
https://gitlab.com/Loicvh/apla, 2019.
bus system. We show that the heuristics produce solutions [13] R. D. Christie, “Power systems test case
of comparable quality to those obtained by APLA, at a archive,” 1999, accessed: 2019-08-07. [Online]. Available:
fraction of the computation time required by the latter. http://labs.ece.uw.edu/pstca
[14] R. D. Zimmerman and C. E. Murillo-Sanchez, “Matpower (ver-
The additional cost of ignoring the VPE effect is computed sion 7.0),” https://matpower.org.
for the different instances of the two IEEE systems and is [15] P. Attaviriyanupap, H. Kita, E. Tanaka, and J. Hasegawa,
measured on average as up to eight percent. “A hybrid EP and SQP for dynamic economic dispatch with
nonsmooth fuel cost function,” IEEE Transactions on Power
Further work may include the extension of the method Systems, vol. 17, no. 2, pp. 411–416, May 2002.
to unit commitment and an analysis of the interactions of [16] A. R. Al-Roomi, “Power Flow Test Systems Repository,”
the valve point effect with the scheduling decisions. The Halifax, Nova Scotia, Canada, 2015. [Online]. Available:
https://al-roomi.org/power-flow
impact of the position of the VPE-units in the network [17] B. Lu and M. Shahidehpour, “Unit commitment with flexi-
may also be of interest, as suggested by the various range ble generating units,” Power Systems, IEEE Transactions on,
of solutions obtained here for different network topologies. vol. 20, pp. 1022 – 1034, 06 2005.

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