See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/324584300

Data-Based Distributionally Robust Stochastic Optimal Power Flow—Part II:

Case Studies

Article in Power Systems, IEEE Transactions on · April 2018

DOI: 10.1109/TPWRS.2018.2878380

CITATIONS READS
47 1,035

5 authors, including:

Yi Guo Zechun Hu
ETH Zurich Tsinghua University
45 PUBLICATIONS 806 CITATIONS 232 PUBLICATIONS 9,122 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Yi Guo on 23 April 2018.

The user has requested enhancement of the downloaded file.

 1

Data-Based Distributionally Robust

Stochastic Optimal Power Flow
Part II: Case Studies
Yi Guo, Student Member, IEEE, Kyri Baker, Member, IEEE, Emiliano Dall’Anese, Member, IEEE,
Zechun Hu, Senior Member, IEEE, Tyler H. Summers, Member, IEEE,
arXiv:1804.06384v1 [math.OC] 17 Apr 2018

Abstract—This is the second part of a two-part paper on data- A variety of stochastic OPF methods have been recently
based distributionally robust stochastic optimal power flow (OPF). proposed to explicitly incorporate forecast errors of network
The general problem formulation and methodology have been uncertainties for modeling risk. These methods can be cate-
presented in Part I [1]. Here, we present extensive numerical
experiments in both distribution and transmission networks to gorized according to assumptions made about forecast error
illustrate the effectiveness and flexibility of the proposed method- distributions and metrics for quantifying risk. Many make
ology for balancing efficiency, constraint violation risk, and out- specific assumptions about forecast error distributions and
of-sample performance. On the distribution side, the method use chance constraints, which encode value at risk (VaR),
mitigates overvoltages due to high photovoltaic penetration using conditional value at risk (CVaR), or distributional robustness
local energy storage devices. On the transmission side, the method
reduces N-1 security line flow constraint risks due to high wind with specific ambiguity sets [2]–[10]. Others handle uncertainty
penetration using reserve policies for controllable generators. In using scenarios approaches, sample average approximation,
both cases, the data-based distributionally robust model predic- or robust optimization [11]–[24]. However, in practice all
tive control (MPC) algorithm explicitly utilizes forecast error decisions about how to model forecast error distributions must
training datasets, which can be updated online. The numerical be based on finite historical training datasets. None of the
results illustrate inherent tradeoffs between the operational costs,
risks of constraints violations, and out-of-sample performance, existing methods account for sampling errors inherent in such
offering systematic techniques for system operators to balance datasets in high-dimensional spaces.
these objectives. Our proposed data-based distributionally robust stochastic
Index Terms—stochastic optimal power flow, data-driven opti- OPF methodology, developed in Part I [1], explicitly incor-
mization, multi-period distributionally robust optimization, dis- porates robustness to sampling errors by considering sets
tribution networks, transmission systems of distributions centered around a finite training dataset. A
model predictive control (MPC) approach utilizes Wasserstein-
based distributionally robust optimization subproblems to
I. I NTRODUCTION
obtain superior out-of-sample performance. Here in Part II
S penetration levels of renewable energy sources (RESs) we illustrate the effectiveness and flexibility of the method-
A continue increasing to substantial fractions of total sup- ology in both distribution and transmission networks. On the
plied power and energy, system operators will require more distribution side, the method mitigates overvoltages due to high
sophisticated methods for balancing inherent tradeoffs between photovoltaic penetration using local energy storage devices. On
nominal performance and operational risks. This is relevant for the transmission side, the method reduces N-1 security line
both transmission and distribution networks, which have seen flow constraint risks due to high wind penetration using reserve
rapid recent increases in photovoltaic (PV) and wind energy policies for controllable generators. In both cases, the data-
sources. Here, we present extensive numerical experiments in based distributionally robust (MPC) algorithm explicitly utilizes
both distribution and transmission networks to illustrate the forecast error training datasets, which can be updated online.
effectiveness and flexibility of our data-based distributionally The numerical results illustrate inherent tradeoffs between the
robust stochastic optimal power flow (OPF) methodology for operational costs, risks of constraints violations, and out-of-
balancing efficiency, constraint violation risk, and out-of-sample sample performance, offering systematic techniques for system
performance. operators to balance these objectives.
The rest of the paper is organized as follows. Section
Y. Guo and T. Summers are with the Department of Mechanical Engi- II illustrates the data-based distributionally robust stochastic
neering, The University of Texas at Dallas, Richardson, TX, USA, email:
{yi.guo2,tyler.summers}@utdallas.edu. AC OPF for mitigating overvoltages in a modified IEEE 37-
K. Baker is with the Department of Civil, Environmental, and Archi- node distribution feeder with high PV penetration using local
tectural Engineering, The University of Colorado, Boulder, CO, USA, energy storage devices. Section III illustrates the data-based
email:[email protected].
E. Dall’Anese is with the National Renewable Energy Laboratory, Golden, distributionally robust stochastic DC OPF for reducing N-1
CO, USA, email: [email protected]. security line flow constraint risks due to high wind penetration
Z. Hu is with the Department of Electrical Engineering, Tsinghua University, using reserve policies for controllable generators. Section IV
Beijing, P.R. China, email: [email protected].
This material is based on work supported by the National Science Foundation concludes.
under grant CNS-1566127. Notation: The inner product of two vectors a, b ∈ Rm is
 2

denoted by ha, bi := a| b. The Ns -fold product of distribution where cos(θn ) ∈ (0, 1] is the power factor limit for RESs.
P on a set Ξ is denoted by PNs , which represents a distribution 3) Energy storage model: The state-of-charge (SOC) of the
on the Cartesian product space ΞNs = Ξ × . . . × Ξ. We use Ns energy storage device located at bus n ∈ NB ⊆ N in kWh is
to represent the number of samples inside the training dataset represented as Bnt . The dynamics of these devices are
Ξ̂. Superscript “ ˆ· ” is reserved for the objects that depend
on a training dataset Ξ̂Ns . We use (·)| to denote vector or Bnt+1 = Bnt + ηB,n PB,n
t
∆, n ∈ NB ,
matrix transpose. The operators <{·} and ={·} return the real where ∆ is the duration of the time interval (t, t + 1],
and imaginary part of a complex number, respectively. The and PB,nt
is the charging/discharging power of the storage
operator [ · ][a,b] selects the a-th to b-th elements of a vector device in kW. We assume the battery state is either charging
or rows of a matrix. t
(PB,n t
≥ 0) or discharging (PB,n ≤ 0) during each time interval
(t, t + 1]. For simplicity, we suppose the round-trip efficiency
II. OVERVOLTAGE MITIGATION IN DISTRIBUTION
of the storage device ηB,n = 1 to avoid the nonconvexity
NETWORKS
when introducing binary variables. Additionally, two common
In this section, we apply the data-based distributionally operational constraints of energy storage devices are
robust stochastic OPF methodology to mitigate overvoltages
in distribution networks by controlling set points in RESs and Bnmin ≤ Bnt ≤ Bnmax , min
PB,n t
≤ PB,n max
≤ PB,n .
energy storage devices. We provide further modeling details of
where Bnmin , Bnmax are the rated lower and upper SOC
the loads, inverter-based RESs, and energy storage devices. The min max
levels, and PB,n , PB.n are the minimum and maximum
set points of controllable devices are repeatedly optimized over
charging/discharging limits. Other constraints can be added
a finite planning horizon within a MPC feedback scheme. The
for electric vehicles (EVs), for example, a prescribed SOC
risk conservativeness of the voltage magnitude constraints and
Bnt = Bnmax at a particular time. If no energy device is
the out-of-sample performance robustness to sampling errors
connected to a certain bus, the charging/discharging power and
are explicitly adjustable by two scalar parameters. t
SOC are fixed to zero: PB,n = 0, Bnt = 0, for all n ∈ N \NB .
We define the aggregate vectors ptB := [PB,1 t t
, . . . , PB,N ]| , and
A. System model t t t |
b := [B1 , . . . , BN ] .
t
1) Loads: We use Pl,n and Qtl,n to denote the active and
reactive power demands at bus n ∈ N . We also define two
vectors ptl := [Pl,1 t t
, . . . , Pl,N ]| and qtl := [Qtl,1 , . . . , Qtl,N ]| . B. Data-based stochastic OPF implementation
t
If no load is connected to bus n ∈ N , then Pl,n = 0 and We now use the methodologies presented in Part I Section
t
Ql,n = 0. Load uncertainties are modeled based on historical IV, and the models of loads, RESs and energy storage devices
data of forecast errors. The active and reactive loads are given to develop a data-based stochastic AC OPF for solving a
by ptl = p̄tl (ut )+p̃tl (ξt ), qtl = q̄tl (ut )+q̃tl (ξt ), where p̄tl (ut ) ∈ voltage regulation problem. This stochastic OPF aims to balance
RN and q̄tl (ut ) ∈ RN are forecasted nominal loads, which can the operational costs the total CVaR values of the voltage
depend on control decisions (e.g., load curtailment control). magnitude constraints. We consider an operational cost that
The nodal injection errors p̃tl (ξt ) ∈ RN and q̃tl (ξt ) ∈ RN captures electricity purchased by customers, excessive solar
depend on the aggregate forecast error vector ξt . energy fed back to the utility, reactive power compensation
t
2) Renewable energy model: Let Pav,n be the maximum costs and penalties for active power curtailment:
availability renewable energy generation at bus n ∈ NR ⊆ N , t
JCost (αt , qtc , ptB , ξt ) =
where the set NR denotes all buses with RESs. With high X
at1,n Pl,n
 t t
− (1 − αnt )Pav,n t

RES penetration, overvoltages can cause power quality and = + PB,n +
reliability issues. By intelligently operating set points of RES n∈N
X
and energy storage, operators can optimally trade off risk of at2,n (1 − αnt )Pav,n t t t
 
+ − Pl,n − PB,n +
constraint violation and economic efficiency (e.g., purchase of n∈N
electricity from the main grid, active power curtailment costs, X X
+ at3,n |Qtn | + at4,n αnt Pav,n
t
.
and reactive compensation costs). The active power injections
n∈N n∈N
of RESs are controlled by adjusting an active power curtailment
t t t t
factor αnt ∈ [0, 1]. Reactive power set points of RESs can also We collect all decision variables into yt = {α , qc , pB , b },
be adjusted within a limit S̄n on apparent power as follows: and all RES and load forecast errors into the random vector
q ξt . Now the MPC subproblems take the following form:
((1 − αnt )Pav,nt )2 + (Qt )2 ≤ S̄ , n ∈ N .
n n R Data-based distributionally robust stochastic OPF:
We define aggregate vectors: αt := [α1t , . . . , αN t |
] and ptav := t+H
Xt  N
X̀ 
t t | t t t |
[Pav,1 , . . . , Pav,N ] and qc := [Q1 , . . . , QN ] . inf E[ Jˆτ
Cost ] + ρ sup E Qτ
[Q̄ τ
o ] ,
yτ ,κτo , Ns
If bus n ∈ N \NR has no RES, by convention we set $ 1,n ,$ 2,n τ =t Q τ ∈ P̂τ o=1

αnt = 0, Pav,n t
= 0 and Qtn = 0. The curtailment factor and t+H
Xt  N 
X̀ 1 X
Ns 
t t
reactive power compensation {αn , Qn } together set the inverter = inf E[ Jˆτ
] + λ ε
o τ + s τ
,
yτ ,κτo ,
Cost
Ns i=1 io
operating point and are also subject to a power factor constraint $1,n ,$2,n τ =t o=1
λτo ,sτio ,ςiko
τ

tan(θn )[(1 − αnt )Pav,n

t
] ≤ |Qtn |, n ∈ NR , (1a)
 3

subject to :
ρ(b̄ok (κτo ) + hāok (yτ ), ξˆτi i + hςiko , d − H ξˆτi i) ≤ sτio , (1b)
kH | ςiko − ρāok (yτ )k∞ ≤ λτo , (1c)
ςiko ≥ 0, (1d)
Ns  
1 X
[(1 − αnτ )P̂av,n
τ,i 2
] + (Qτn )2 − S̄n2 + $1,n
τ τ
≤ $1,n µ
Ns i=1 +
(1e)
Ns  
1 X
tan(θn )[(1 − αnt )P̂av,n
τ,i
] − |Qτn | + $2,n
τ τ
≤ $2,n µ,
Ns i=1 +
(1f)
Bnmin ≤ Bnτ ≤ Bnmax , (1g)
min τ max
PB,n ≤ PB,n ≤ PB,n , (1h)
τ +1 τ τ
Bn = Bn + ηB,n PB,n ∆, (1i)
τ
0 ≤ αn ≤ 1, (1j)
∀i ≤ Ns , ∀o ≤ N` , n ∈ NR , k = 1, 2, τ = t, . . . , t + Ht ,
τ τ
where $1,n , $2,n , and κτo are CVaR auxiliary valuables,
τ τ τ
and λo , sio , ςiko are auxiliary variables associated with the
Fig. 1. IEEE 37-node test feeder with renewable energy resource and storage
distributionally robust Wasserstein ball reformulation. For devices.
simplicity, the power factor constraints and apparent power
limitation constraints are not treated as distributionally robust
4000
constraints, and instead are handled using direct sample average Total Load Demand
3500 Total Available Solar Energy
approximation. 3000

2500
kW

2000
C. Numerical results
1500
We use a modified IEEE-37 node test feeder to demonstrate 1000

our proposed data-based stochastic AC OPF method. As shown 500

in Fig. 1, the modified network is a single-phase equivalent 0

12 AM 3 AM 6 AM 9 AM 12 PM 3 PM 6 PM 9 PM 12 AM
and the load data is derived from real measurements from
feeders in Anatolia, CA during the week of August 2012 [25].
Fig. 2. Total available solar energy and load demand.
We place 21 photovoltaic (PV) systems in the network. Their
locations are marked by yellow boxes in Fig. 1, and their
TABLE I
capacities are summarized in Table I. Based on irradiation data C APACITIES OF INVERTED - BASED SOLAR ENERGY GENERATIONS AND
from [25], [26], we utilized a greedy gradient boosting method ENERGY STORAGE DEVICES
[27] to make multi-step ahead predictions of solar injections, max
Node Sn [kVA] Node Sn [kVA] Node Bn
and then computed a set of forecast errors from the dataset. 4 150 7 300 9 100
In general forecast errors increase with the prediction horizon. 9 300 10 600 10 100
Other parameters of the network, such as line impedances 11 660 13 360 28 50
and shunt admittances, P are taken from [28]. The total nominal 16 600 17 360 29 250
t 20 450 22 150 32 250
available solar power n Pav,n and aggregate load demand 23 750 26 300 35 120
over 24 hours is also shown in Fig. 2. 28 750 29 300 36 200
The energy storage systems are placed with PV systems at 30 360 31 600
32 330 33 750
certain nodes, as shown in Fig. 1. Their locations and capacities 34 450 35 450
are listed in Table I. We select the capacities in the range 36 450
of typical commercial storage systems, or aggregate of 10-
12 residential-usage batteries (e.g., electric vehicles), which
are connected to the same step-down transformer. The lower Due to high PV penetration, overvoltage conditions can
limit of SOC, Bnmin , is set to be zero for all batteries. The emerge during solar peak irradiation. For simplicity, we model
t
charging/discharging rate PB,n is also limited by 10% of their only the upper voltage limitations over the solar peak hours
respective energy capacity Bnmax . Voltage limits V max and 6AM to 5PM as distributionally robust (1b)-(1d). The lower
V min are 1.05 p.u. and 0.95 p.u., respectively. The cost function voltage limitations and other constraints during the 6AM to
parameters are at1,n = 10, at2,n = 3, and at3,n = 3 and at4,n = 6. 5PM are handled with empirical sample averaging, as in
The decision making time period is 5 minutes. [6], [20]. The power factor PF limit is 0.9 in (1f). The risk
 4

level parameter η is set to 0.01 for quantifying 1% violation profiles, as shown in Fig. 5. When ρ is small, there is almost
probability of constraints (1e)-(1f). To emphasize the effect no curtailment, causing overvoltages at several buses. As ρ
of sampling errors, the number of forecast error samples Ns increases, more active power is curtailed, and all voltages
included in the training dataset Ξ̂N t
s
is limited to 30. The move below their upper limit. Similar comments apply for
forecast errors are not assumed bounded, so the parameters varying the Wasserstein radius ε. For example, fixing ρ and
of the support polytope Ξt := {ξt ∈ RNξ : Hξt ≤ d} are set increasing ε also results in more curtailment and lower voltage
to zero in (1b)-(1c). We solved (1) using the MOSEK solver magnitude profiles, which leads to better robustness to solar
[29] via the MATLAB interface CVX [30] on a laptop with energy forecast errors.
16 GB of memory and 2.8 GHz Intel Core i7. Solving each Finally, we evaluate out-of-sample performance by im-
time step during solar peak hours with distibutionally robust plementing the full closed-loop distributionally robust MPC
constraints takes 4.84 seconds. Note that our implementation scheme over the 24 hour period with a 15 minute planning
is not optimized for speed and in principle could easily be horizon. Monte Carlo simulations with 100 realizations of
sped up and scaled to larger problems since the problem is forecast errors over the entire horizon are shown in in Fig.
ultimately convex quadratic. 6. We subsampled new solar energy forecast errors from the
In our framework, there are two key parameters, ρ and training dataset. The closed-loop voltage profiles based on
ε, that explicitly adjust trade offs between performance and MPC decisions for all scenarios at node 28 are shown (other
constraint violation risk, and robustness to sampling errors. nodes with overvoltages show qualitatively similar results).
Fig. 3 illustrates the basic tradeoffs between operational cost Again, it is clearly seen that larger values of ε and ρ yield
and CVaR values of voltage constraint violations during a more conservative voltage profiles.
24-hour operation for various values of ρ and ε. It can be In summary, we conclude that the proposed data-based
readily seen that as ρ increases, operational cost increases, but distributionally robust stochastic OPF is able to systematically
CVaR decreases since the risk term is emphasized. Similarly, assess and control tradeoffs between the operational costs, risks,
as ε increases, we see that CVaR decreases, since a larger and sampling robustness in distribution networks. The benefits
Wasserstein ball leads to higher robustness to sampling errors. of the open-loop stochastic optimization problems are also
These parameters offer system operators explicit data-based observed in the closed-loop multi-period distributionally robust
tuning knobs to systematically set the conservativeness of model predictive control scheme.
operating conditions.
Fig. 4(a)-4(c) shows the aggregated solar energy curtailment
and substation power purchases for varying risk aversion ρ and III. N-1 SECURITY PROBLEM IN TRANSMISSION SYSTEMS
Wasserstein radius ε. In order to prevent voltages over 1.05 p.u., In this section, we apply the proposed methodology from
the available solar energy must be increasingly curtailed as the Part I in a transmission system to handle N-1 line flow security
risk aversion parameter ρ increases. As a result, the network constraints. The basic DC power flow approximation and device
must import more power from the substation. The increasing modeling is discussed in Part I. Here, we also incorporate N-1
curtailment of solar energy and purchase of power drawn from security constraints and associated contingency reactions due
the substation lead to significantly higher operational cost. to uncertain wind power injections.
However, these decisions will also lead to more stable voltage

A. System model
We consider a transmission system with NG generators
(e.g., conventional thermal and wind) connected to bus subset
0.1
NG ⊆ N . There are NL loads, Nl lines, and Nb buses. The
outages included for N-1 security consist of tripping of any
0.05
single lines, generators or loads, yielding Nout = NG +NL +Nl
possible outages. We collect the outages corresponding to a
CVaR

0
generator, a line and a load in different sets IG , Il , and IL .
The outages in total are I = {0} ∪ IG ∪ IL ∪ Il , where {0}
-0.05
indicates no outage.
The formulation of a N-1 security problem is based on
the following assumptions: 1) the power flow equations are
100 12 PM approximated with DC power flow, as described in Part I
9 AM
Operational 50 6 AM
(Section V); 2) each wind farm is connected to a single bus
Cost in Dollars of the network; 3) load forecasting is perfect; 4) a single line
3 AM
0 Time
ρ Increase outage can cause multiple generator/load failures.
The objective of the data-based distributionally robust
Fig. 3. Tradeoffs between operational cost, conditional value at risk (CVaR)
of voltage constraints, and robustness to sampling errors, illustrated by varying
stochastic DC OPF is to determine an optimal reserve schedule
the parameters ρ and ε that control relative objective function weighting and for responding to the wind energy forecast errors while taking
Wasserstein ball radius, respectively. the network security constraints into account. We define
 5

4000 4000 4000

rho = 1 rho = 1 rho = 1
3500 rho = 15 3500 rho = 15 3500 rho = 15
rho = 20 rho = 20 rho = 20
3000 rho = 30 3000 rho = 30 3000 rho = 30
rho = 35 rho = 35 rho = 35
2500 rho = 100 2500 rho = 100 2500 rho = 100
rho = 200 rho = 200 rho = 200
kW

kW

kW
2000 2000 2000

1500 1500 1500

1000 1000 1000

500 500 500

0 0 0
12 AM 3 AM 6 AM 9 AM 12 PM 3 PM 6 PM 9 PM 12 AM 12 AM 3 AM 6 AM 9 AM 12 PM 3 PM 6 PM 9 PM 12 AM 12 AM 3 AM 6 AM 9 AM 12 PM 3 PM 6 PM 9 PM 12 AM

(a) Total curtailed active power ε = 0.0000 (b) Total curtailed active power ε = 0.0005 (c) Total curtailed active power ε = 0.0010

3000 3000 3000

2000 2000 2000

1000 1000 1000

kW

kW

kW
0 0 0
rho = 1 rho = 1 rho = 1
rho = 15 rho = 15 rho = 15
-1000 rho = 20 -1000 rho = 20 -1000 rho = 20
rho = 30 rho = 30 rho = 30
rho = 35 rho = 35 rho = 35
-2000 -2000 -2000
rho = 100 rho = 100 rho = 100
rho = 200 rho = 200 rho = 200
-3000 -3000 -3000
12 AM 3 AM 6 AM 9 AM 12 PM 3 PM 6 PM 9 PM 12 AM 12 AM 3 AM 6 AM 9 AM 12 PM 3 PM 6 PM 9 PM 12 AM 12 AM 3 AM 6 AM 9 AM 12 PM 3 PM 6 PM 9 PM 12 AM

(d) Power drawn from substation ε = 0.0000 (e) Power drawn from substation ε = 0.0005 (f) Power drawn from substation ε = 0.0010

Fig. 4. Comparison of active power curtailment and power purchased from substation for various values of risk aversion ρ and Wasserstein radius ε. As these
parameters increase, more active power from PV is curtailed and more power is drawn from the substation, leading to lower constraint violation risk but higher
operating cost.

Fig. 5. Optimal network voltage profiles for varying ρ and ε. Overvoltages are reduced as ρ increases.

j
Pmis ∈ R for all j ∈ I as the generation-load mismatch PGj or PLj is the sum of the power loss caused by multiple
given by failures. If there is no power disconnection caused by a line
 j outage, or if j = 0 (no outage happens), the power mismatch
j j
 PL − PG , if j ∈ {0} ∪ Il is set to zero: Pmis = 0.
j
Pmis = +PLj , if j ∈ IL To respond to contingencies, we can also define another
−PGj , if j ∈ IG , reserve policy response matrix Rj,d j,d j,d |

mis,t := [R1 , . . . , Rt ] ∈
Rt , so that the affine reserve policy becomes
where PLj , PGj ∈ R denote the power disconnection corre-
sponding to the outage j ∈ I. Define PG ∈ RNG , and
PL ∈ RNL as nodal generation and load injection vectors. udt = Dtd ξt + Rj,d j d
mis,t Pmis + et , ∀j ∈ I, d = 1, . . . , Nd . (2)
For the generator or load failures, the power disconnection PGj
or PLj is corresponding to the components in the vector PG or The general constraint risk function of the line flow in Part I
PL . In the case of line failures j ∈ Il , the power disconnection (Section V, Equation (14c)) is then given by
 6

Fig. 6. Monte Carlo simulation results of the voltage profiles at node #28 resulting from the full distributionally robust closed-loop model predictive control
scheme. We validate that in closed-loop larger values of ε and ρ yield more conservative voltage profiles.

τX
+Ht  N  Ns 
X̀ 1 X
= infτ E[JˆCost
τ
]+ λτo ετ + sτio ,
Nd
X n yτ ,σo ,
τ =t o=1
Ns i=1
e t,j rdt + Gtd ξt + Cdt Adt xd0 + Btd (Dtd ξt
 λτo ,sτio ,ςiko
τ
Ef Γ d
(4a)
d=1 (3)
subject to :

j,d j d
o
+Rmis,t Pmis + et ) − p̄t ≤ 0, ∀j ∈ I,
ρ(bok (σoτ ) + haok (yτ ), ξˆτi i + hςiko
τ
, d − H ξˆτi i) ≤ sτio , (4b)
τ
where Γe t,j ∈ R2Lt×t maps the power injection of each device kH | ςiko − ρaok (yτ )k∞ ≤ λτo , (4c)
d τ
in the case of j-th outage. ςiko ≥ 0, (4d)
Nd 
Ns X
1 X n
e t,j rdt + Gtd ξ̂ i + Cdt Ad xd0 + B d (Dd ξ̂ i

Γ d t t t t t
B. Data-based stochastic OPF implementation Ns i=1 d=1
We now use the modeling here and in Part I and the affine

o
control strategy (2) to formulate a data-based distributionally + Rj,d Pj
mis,t mis
+ edt ) − p̄t ≤ 0, (4e)
[t,t]
robust stochastic DC OPF for transmission systems that also
Nd h
incorporate N-1 security constraints. The goal is to balance X i
(rdt + Cdt (Adt xd0 + Btd edt )) = 0, (4f)
tradeoffs between cost of thermal generation, CVaR values of [t,t]
d=1
the line flow constraints, and sampling error robustness. The Nd h i
generation
 with reserve policy in the cost function is given by
X
(Gtd + Cdt Btd Dtd ) = 0, (4g)
PGd,t = Dtd ξt + edt t . The operational cost of generators is

[t,t]
d=1
X ∀i ≤ Ns , ∀j ∈ I, ∀o ≤ N` , k = 1, 2, τ = t, . . . , t + Ht .
t
JCost = c1,d [PGd,t ]2 + c2,d [PGd,t ] + c3,d ,
d∈NG

which captures nominal and reserve costs of responding to C. Numerical results

wind energy forecast errors. The N-1 security reserve cost is We consider a modified IEEE 118-bus test system to
not included to simplify presentation, but this can also easily demonstrate our proposed data-based distributionally robust
be included in our framework as an additional linear cost [15]. stochastic DC OPF shown in Fig. 7. For simplicity, we only
With the proposed modeling in Part I (Section V), the show results of a single-period stochastic optimization problem.
updated data-based stochastic DC OPF is shown as follows. As with the distribution network, it is straightforward to extend
The decision variables are collected into yt = {Dt , et , Rmis,t }. to multi-period closed-loop stochastic control using MPC.
The random vector ξt comprises all wind energy forecast errors. Three wind farms are connected to bus #1, bus #9, and bus
Data-based distributionally robust stochastic DC OPF: #26, with the normal feed-in power 500 MW, 500 MW, and 800
t+H N
MW, respectively. The corresponding conventional generators
Xt  X̀ 
at bus #1 and bus #26 are removed. The wind power forecast
inf E[JˆCost ] + ρ sup
τ Q τ
E [Qo ] ,
τ
yτ ,σoτ
τ =t
Ns
Qτ ∈P̂τ o=1
errors are derived from the real wind sampling data from
 7

to sampling errors. The conservativeness of the generator

policies are controlled by adjusting ρ and ε. By explicitly using
the forecast error training dataset and accounting for sampling
errors, risks are systematically assessed and controlled. Since
the forecast errors are non-Gaussian, existing methods may
significantly underestimate risk [32]. Increasing ε provides
better robustness to sampling errors and guarantees out-of-
sample performance. Note that CVaR values can be negative
when the worst-case expected line flow is below the constraint
boundary.
Fig. 8(b) illustrates the out-of-sample performance of the
decisions via Monte Carlo simulations. For every risk aversion
value ρ shown in Fig. 8(a), we subsampled new values from the
training dataset to operate our system based on the decisions
from Fig. 8(a), and calculated the empirical CVaR of the line
constraints violations. The number of scenarios for Monte Carlo
simulations is 1000. From the simulation results, we conclude
Fig. 7. IEEE 118-bus test network with multiple wind energy connections. that larger ε ensures smaller line constraint violation for all
lines except line 38. This happens because the risk objective is
the sum of all five CVaRs, and a lower overall risk is achieved
hourly wind power measurements provided in 2012 Global for certain values of ρ and ε by trading off against risk of
Energy Forecasting competition (GEFCCom2012) [31]. The violating the flow limit of line 38. In general, it is possible
wind energy forecast errors are evaluated based on a simple to prioritize certain important constraints by weighting their
persistence forecast, which predicts that the wind injection associated risk higher compared to lower priority constraints.
in the following period remains constant. It can be seen the Again, Monte Carlo simulations demonstrate conservativeness
forecast errors are highly leptokurtic, i.e, that the errors have of controlled explicitly by changing the Wasserstein radius
significant outliers that make the distribution tails much heavier ε and the risk aversion ρ. Overall, our simulations illustrate
than Gaussian tails. We scale the forecasting errors to have zero the flexibility of our proposed data-based stochastic OPF for
mean and the standard deviations of 200 MW, 200 MW and transmission systems.
300 MW for the wind farms at bus #1, #9 and #26, respectively.
We consider five key lines, which deliver the wind power from IV. C ONCLUSION
the left side of the system to the right, as marked by the red We have illustrated the effectiveness and flexibility of our pro-
crosses in Fig. 7. The line flow limits are shown in Table II. posed multi-period data-based distributionally robust stochastic
OPF methodology. We performed numerical experiments to
TABLE II balance overvoltages in distribution networks and N −1 security
F IVE MAIN CHANNEL LINES DATA
line flow risks in transmission networks. The flexibility of
# of line From bus To bus Line flow limitation [MW] controllable devices was exploited to balance efficiency and
7 8 9 600 risk due to high penetration renewable energy sources. In
37 8 30 500
38 26 30 500 contrast to existing work, our method directly incorporates
54 30 38 500 forecast error training datasets rather than making strong
96 38 65 300 assumptions on the forecast error distribution, which allows us
to leverage distributionally robust optimization techniques to
To simple our presentation, only these five lines flows are achieve superior out-of-sample performance. Parameters in the
handled with distributionally robust optimization (4b)-(4d) in optimization problems allow system operators to systematically
both directions; the remaining line flows are modeled by N- select operating strategies that optimally trade off performance
1 security constraints (4e) with nominal CVaR and sample and risk.
average approximation (essentially equivalent to taking ε = 0),
and no other local device constraint is included. Additionally, ACKNOWLEDGEMENTS
we assume no bounds on the wind power forecast errors ξτ ,
The authors would like to thank Prof. Jie Zhang and Mr.
hence H and d in (4b)-(4c) are set to zero. It takes 10 seconds
Mucun Sun, at the Department of Mechanical Engineering,
to solve (4) for each time step using the MOSEK solver [29]
The University of Texas at Dallas, for their support on the
via the MATLAB interface with CVX [30] on a laptop with
solar forecasting.
16 GB of memory and a 2.8 GHz Intel Core i7.
Fig. 8(a) illustrates the solutions of the proposed data-based
R EFERENCES
distributionally robust stochastic DC OPF for varying risk
aversion and Wasserstein radius. Once again, the numerical [1] Y. Guo, K. Baker, E. Dall’Anese, Z. Hu, and T. Summers, “Data-
based distributionally robust stochastic optimal power flow, Part I:
results demonstrate fundamental tradeoffs between operational Methodologies,” IEEE Transactions on Power Systems (under review),
cost, CVaR values of line constraints violations, and robustness 2018.
 8

Line 7 Line 37 Line 7 Line 37

20 20 -500 600

10 10 -600 500
CVaR

CVaR

MW

MW
0 0 -700 400

-10 -10 -800 300

6 6.5 7 7.5 8 8.5 6 6.5 7 7.5 8 8.5 0 200 400 600 800 1000 0 200 400 600 800 1000
Operation Cost in Dollars × 104 Operation Cost in Dollars × 104 weight factor ρ weight factor ρ
Line 38 Line 54 Line 38 Line 54
5 15 550 800

500 700
10
CVaR

CVaR

MW

MW
0 450
600
5
400
500
-5 0 350
6 6.5 7 7.5 8 8.5 6 6.5 7 7.5 8 8.5 0 200 400 600 800 1000 0 200 400 600 800 1000
Operation Cost in Dollars × 104 Operation Cost in Dollars × 104 weight factor ρ weight factor ρ
Line 96 Line 96
10 500

5 400
CVaR

MW
0 300

-5 200
6 6.5 7 7.5 8 8.5 0 200 400 600 800 1000
Operation Cost in Dollars × 104 weight factor ρ

(a) Predicted tradeoffs between operational cost and CVaR of line constraint (b) Out-of-sample performance is demonstrated by Monte Carlo simulation, the
violation. conservativeness is explicitly controllable.

Fig. 8. Five key line flow constraints are used to illustrate our proposed data-based distributionally robust stochastic DC OPF in a transmission system.

[2] D. Bienstock, M. Chertkov, and S. Harnett, “Chance-constrained optimal reserves for power systems,” IEEE Transactions on Power Systems,
power flow: Risk-aware network control under uncertainty,” SIAM Review, vol. 28, pp. 4427–4437, Nov. 2013.
vol. 56, no. 3, pp. 461–495, 2014. [17] R. A. Jabr, S. Karaki, and J. A. Korbane, “Robust multi-period OPF with
[3] Y. Zhang and G. B. Giannakis, “Robust optimal power flow with storage and renewables,” IEEE Transactions on Power Systems, vol. 30,
wind integration using conditional value-at-risk,” in IEEE International pp. 2790–2799, Sept. 2015.
Conference on Smart Grid Communications, pp. 654–659, Oct. 2013. [18] M. Lubin, Y. Dvorkin, and S. Backhaus, “A robust approach to chance
[4] M. Perninge and C. Hamon, “A stochastic optimal power flow problem constrained optimal power flow with renewable generation,” IEEE
with stability constraints, Part II: The optimization problem,” IEEE Transactions on Power Systems, vol. 31, pp. 3840–3849, Sept. 2016.
Transactions on Power Systems, vol. 28, pp. 1849–1857, May 2013. [19] L. Roald, T. Krause, and G. Andersson, “Integrated balancing and con-
[5] L. Roald, F. Oldewurtel, T. Krause, and G. Andersson, “Analytical refor- gestion management under forecast uncertainty,” in IEEE International
mulation of security constrained optimal power flow with probabilistic Energy Conference, pp. 1–6, Apr. 2016.
constraints,” in IEEE Grenoble PowerTech, pp. 1–6, June 2013. [20] E. Dall’Anese, K. Baker, and T. Summers, “Chance-constrained AC
optimal power flow for distribution systems with renewables,” IEEE
[6] T. Summers, J. Warrington, M. Morari, and J. Lygeros, “Stochastic
Transactions on Power Systems, vol. 32, pp. 3427–3438, Sept. 2017.
optimal power flow based on conditional value at risk and distributional
[21] G. Hug-Glanzmann and G. Andersson, “Decentralized optimal power
robustness,” International Journal of Electrical Power & Energy Systems,
flow control for overlapping areas in power systems,” IEEE Transactions
vol. 72, pp. 116–125, Nov. 2015.
on Power Systems, vol. 24, pp. 327–336, Feb. 2009.
[7] B. Li, R. Jiang, and J. L. Mathieu, “Distributionally robust risk- [22] G. Hug-Glanzmann, Coordinated power flow control to enhance steady-
constrained optimal power flow using moment and unimodality infor- state security in power systems. PhD thesis, ETH Zurich, 2008.
mation,” in IEEE Conference on Decision and Control, pp. 2425–2430, [23] M. Vrakopoulou, B. Li, and J. L. Mathieu, “Chance constrained reserve
Dec. 2016. scheduling using uncertain controllable loads, Part I: Formulation and
[8] H. Zhang and P. Li, “Chance constrained programming for optimal power scenario-based analysis,” IEEE Transactions on Smart Grid, 2017.
flow under uncertainty,” IEEE Transactions on Power Systems, vol. 26, [24] B. Li, M. Vrakopoulou, and J. L. Mathieu, “Chance constrained
pp. 2417–2424, Nov. 2011. reserve scheduling using uncertain controllable loads, Part II: Analytical
[9] M. Hedayati-Mehdiabadi, K. W. Hedman, and J. Zhang, “Reserve policy reformulation,” IEEE Transactions on Smart Grid, 2017.
optimization for scheduling wind energy and reserve,” IEEE Transactions [25] J. Bank and J. Hambrick, “Development of a high resolution, real time,
on Power Systems, vol. 33, pp. 19–31, Jan. 2018. distribution-level metering system and associated visualization, modeling,
[10] M. Bazrafshan and N. Gatsis, “Voltage regulation in electricity distribution and data analysis functions,” tech. rep., National Renewable Energy
networks using the conditional value-at-risk,” in IEEE Global Conference Laboratory (NREL), Golden, CO., 2013.
on Signal and Information Processing, pp. 909–913, Dec. 2014. [26] “National solar radiation databased (NSRDB),” 2014.
[11] T. Yong and R. H. Lasseter, “Stochastic optimal power flow: Formulation [27] J. H. Friedman, “Greedy function approximation: a gradient boosting
and solution,” in IEEE Power Engineering Society Summer Meeting, machine,” The Annals of statistics, pp. 1189–1232, Oct. 2001.
vol. 1, pp. 237–242, July 2000. [28] R. D. Zimmerman, C. E. Murillo-Sánchez, and D. Gan, “MATPOWER:
[12] F. Capitanescu, S. Fliscounakis, P. Panciatici, and L. Wehenkel, “Cautious A MATLAB power system simulation package,” Manual, Power Systems
operation planning under uncertainties,” IEEE Transactions on Power Engineering Research Center, Ithaca NY, vol. 1, 1997.
Systems, vol. 27, pp. 1859–1869, Nov. 2012. [29] M. ApS, The MOSEK optimization toolbox for MATLAB manual. Version
[13] A. J. Conejo, M. Carrión, and J. M. Morales, Decision making under 8.1., 2017.
uncertainty in electricity markets, vol. 1. Springer, 2010. [30] I. CVX Research, “CVX: Matlab software for disciplined convex
programming, version 2.0.” http://cvxr.com/cvx, Aug. 2012.
[14] M. Vrakopoulou, M. Katsampani, K. Margellos, J. Lygeros, and
[31] “Global energy forecasting competition 2012 wind energy competition,”
G. Andersson, “Probabilistic security-constrained AC optimal power
2012.
flow,” in IEEE Grenoble PowerTech, pp. 1–6, June 2013.
[32] Y. Guo, K. Baker, E. Dall’Anese, Z. Hu, and T. Summers, “Stochastic
[15] M. Vrakopoulou, K. Margellos, J. Lygeros, and G. Andersson, “A prob- optimal power flow based on data-driven distributionally robust optimiza-
abilistic framework for reserve scheduling and N-1 security assessment tion,” arXiv preprint arXiv:1706.04267, June 2017.
of systems with high wind power penetration,” IEEE Transactions on
Power Systems, vol. 28, pp. 3885–3896, Nov. 2013.
[16] J. Warrington, P. Goulart, S. Mariéthoz, and M. Morari, “Policy-based

View publication stats