This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
1

Optimal Control of AGC Systems Considering

Non-Gaussian Wind Power Uncertainty
Xiaoshuang Chen, Student Member, IEEE, Jin Lin, Member, IEEE, Feng Liu, Member, IEEE,
and Yonghua Song, Fellow, IEEE

Abstract—Wind power uncertainty poses significant challenges SBSP Scenario-Based Stochastic Programming.
for automatic generation control (AGC) systems. It can enhance SCP Stochastic control problem.
control performances to explicitly consider wind power uncer- SDE Stochastic differential equation.
tainty distributions within controller design. However, widely
accepted wind uncertainties usually follow non-Gaussian distri-
butions which may lead to complicated stochastic AGC modeling Indices/Sets
and high computational burdens. To overcome the issue, this
paper presents a novel Itô-theory-based model for the stochastic i, j, k Indices of units.
control problem (SCP) of AGC systems, which reduces the l Indices of branches.
computational burden of optimization considering non-Gaussian m Indices of areas.
wind power uncertainty to the same scale as that for deterministic n Indices in the series expansion.
control problems. We present an Itô process model to exactly
describe non-Gaussian wind power uncertainty, and then propose t, s Time.
an SCP based on the concept of stochastic assessment functions a, b, c Parameters of typical distributions in Table
(SAFs). Based on a convergent series expansion of the SAF, the V.
SCP is reformulated as a certain deterministic control problem ΩB Index set of buses.
without sacrificing performance under non-Gaussian wind power ΩE Index set of EUs.
uncertainty. The reformulated control problem is proven as
a convex optimization which can be solved efficiently. A case ΩL Index set of branches.
study demonstrates the efficiency and accuracy of the proposed ΩL
m Index set of tie lines connected to Area m.
approach compared with several conventional approaches. Adj(i) Set of buses adjacent to Bus i.
Index Terms—Automatic generation control, non-Gaussian
distribution, stochastic control, stochastic differential equation, Variables/Vectors
wind power uncertainty, Itô theory.
Pig Output of the i-th governor.
Pim Machanical power of the i-th turbine.
N OMENCLATURE Piref Power reference of Bus i.
Abbreviations Pik Power flow from Bus i to Bus k.
ωi Rotating speed of the i-th generator.
ACE Area control error. ∆f System frequency deviation.
AGC Automatic generation control. ACEm Area control error of the m-th area.
DAF Deterministic assessment function. Wt Vector of standard Wiener process.
DC Deterministic control. Zt Vector of stochastic resources.
ESGN Extended skew generalized normal distribu- Xt Vector of all state variables.
tion. Zt Vector of stochastic resources.
MCS Monte Carlo simulation. Ut Vector of all control variables.
MPC Model predictive control. St Vector of all state variables, control variables,
ODE Ordinary differential equation. and stochastic variables.
PDE Partial differential equation. X
ft , X ct Auxiliary vector of Xt .
PDF Probability distribution function. Zt , Zt
e b Auxiliary vector of Zt .
PI Proportion-Integration controller. Set , Zbt Auxiliary vector of St .
SAF Stochastic assessment function. J Objective function.
This work was supported by National Key Research and Development J˜0 , J˜1 Deterministic counterparts of the objective
Program of China (2018YFB0905200), Key R&D Program of Sichuan function.
Province (18SXYHZ0049), and National Natural Science Foundation of China
(51577096, 51677100, 51761135015).
lt 2nd-order partial derivatives of J0 with re-
X. Chen, J. Lin, and F. Liu are with State Key Laboratory of Control spect to z0
and Simulation of Power Systems and Generation Equipment, Department
of Electrical Engineering, Tsinghua University, Beijing 100084, China(email:
linjin@tsinghua.edu.cn). Functions/Operators
Y. Song is with the Department of Electrical and Computer Engineering,
University of Macau, Macau, China, and also with the Department of µ(·) Drift function in Itô processes.
Electrical Engineering, Tsinghua University, Beijing 100084, China. σ(·) Diffusion function in Itô processes.

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
2

p(·) Probability density function. uncertainty is modeled as a deterministic input based on the
Pr {·} Operator of probability. calculated prediction, and the prediction is revised over the
var{·} Operator of variance. control period. Therefore, MPC requires only the solution of a
Ex0 ,z0 {·} Expectation under initial condition x0 , z0 . deterministic optimization problem at each time step. Although
S1,m , u1 , u2 Auxiliary function when computing CPS1. the approaches mentioned above achieve success for many
Pα,β (t, x0 ) SAF. AGC problems, a key issue is that the probability distribution
α(·), β(·) Function parameters in SAF. of the uncertainty is not explicitly considered in the controller
v(t, x0 , z0 ) A simplified notation of SAF. model, which may limit the control performance. For example,
ven (t, x0 , z0 ) DAF. MPC uses a receding-horizon implementation to cope with the
αn , βn Function parameterse in DAF. uncertainty. To achieve better performance, a larger number of
∇2 Operator of taking 2nd-order derivatives. prediction steps and a smaller time step are needed, which will
∇2z0 Operator of taking 2nd-order derivatives over result in a heavy online computational burden [13].
z0 . Problems that explicitly consider the probability distribution
of the uncertainty in control systems are usually regarded
as stochastic control problems (SCPs). Generally, optimal
Parameters
SCPs with Gaussian distributions can be solved via quadratic
Tig Time constant of the i-th governor. programming [14]. However, the wind power uncertainty in
Tit Time constant of the i-th turbine. AGC systems is usually non-Gaussian. In previous research,
Ri Droop coefficient of the i-th generator. the long-tailed characteristic of the wind power uncertainty
Hi Inertia of the i-th generator. has usually been modeled with a beta distribution or Laplace
Di Damping of the i-th generator. distribution [15]. SCPs with non-Gaussian uncertainty are usu-
bm Bias factor of the m-th area. ally solved via scenario-based stochastic programming (SBSP)
P̄iref Maximal value of Piref algorithms, in which the uncertainties are characterized by a
P̄iE Maximal value of PiE finite set of random realizations and the SCP is solved via a
ĒiE Maximal value of EiE deterministic optimization problem [16]. The SBSP approach
P̄ij Maximal value of Pi j. is widely used in intra-day applications of power systems
x0 Initial value of Xt at time t = 0. [16], [17]. Although a few studies have applied SBSP for
z0 Initial value of Zt at time t = 0. the assessment and control of AGC systems [18]–[21], it is
A, B, C Matrix coefficients of ODEs. still challenging to do so because the number of variables
φi Coefficient vector of the i-th constraint. is proportional to the number of scenarios, which is usually
γ Confidence level of the chance constraint. large to ensure good control performance [22]. Therefore, this
κγ Coefficient of inner approximation of chance approach is not typically applicable in the AGC context.
constraints under the confidence level γ. In summary, several challenges arise when dealing with
λACE , µACE Coefficients of ACE in the objective function. uncertainty in AGC systems: 1) The uncertainty is usually
ΛU Coefficients of control variables in the objec- non-Gaussian and dependent on time [15], [23]. 2) It is
tive function. challenging to explicitly consider uncertainty in AGC systems
while incurring a feasible computational burden. The authors’
I. I NTRODUCTION previous work [24] shows that a series expansion of Itô process

I N past few years, the penetration of wind generations has

been increasing rapidly [1]. However, the wind generations
are introducing much more uncertainty into power systems,
and this uncertainty has severe impacts on the automatic
generation control (AGC) systems [2], [3]. The uncertainty
brought by wind generations is usually non-Gaussian, which
is challenging for uncertainty modeling and AGC design under
this non-Gaussian uncertainty.
Techniques for handling wind power uncertainty in AGC
systems include proportional-integral (PI) control [4], [5],
robust control [6], [7], and model predictive control (MPC)
[8]–[12]. PI AGC controllers are among the most popular
controllers in practice, and there have been studies on sev-
eral variants thereof, such as fuzzy PI controllers [4] and
dynamic-gain PI controllers [5]. In robust AGC approaches,
the uncertainty is modeled using intervals or bounded sets,
and the controllers are designed to guarantee the performance
in the worst case scenario [6], [7]. In MPC AGC controllers
[8]–[12], an open-loop optimal control problem is solved
repeatedly in a receding-horizon manner. In this approach, the
can be used for accurate AGC assessment under non-Gaussian
uncertainty, and provides an Itô-Theory-Based (ITB) algorithm
to assess AGC performance. However, it is still challenging to
apply the ITB approach in the optimal control problem. On
the one hand, ITB in [24] can only deal with given control
schemes, while the control scheme is to be determined in the
optimal control scheme. On the other hand, the formulation
in [24] is non-convex, while the convexity is crucial in the
optimal control problem.
Against this background, this paper extends the results
in [24], and proposes a novel Itô-theory-based approach to
compute the optimal control problem of AGC systems under
non-Gaussian uncertainties. The key idea of this paper is to in-
troduce the concept of stochastic assessment functions (SAFs)
to express the expectation values in SCPs and then transform
the SAFs into a few deterministic assessment functions (DAFs)
to significantly reduce the computational burden. Different
from [24], this paper supports parameterized control schemes
and formulates a convex optimization model of the optimal
control problem.

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
3

The major contributions of this paper include the following: Therefore, (1)
 can be simplified as 
1) An Itô model of an AGC system considering the the 1 

1
 X
non-Gaussian stochastic characteristics of wind farms ω̇i = − Di + ωi − Piref − Z (i) + Pij 
Hi Ri
is presented, in which the wind power uncertainty is j∈Adj(i)
modeled in the form of Itô processes and the AGC (3)
system is described by a stochastic differential equation 2) Branch Flow: The branch flow Pij satisfy
(SDE) model. This formulation is applicable for various
distributions of wind power uncertainty. Ṗij = Bij (ωi − ωj ) (4)
2) Based on the Itô model of AGC systems, an SCP is with Bij being a constant defined as
formulated in which the expectation values are expressed |Vi ||Vj |
cos θi0 − θj0


as SAFs. We then reveal the SAFs into a series of Bij = (5)
xij
DAFs with the proof of convergence. Unlike in the SBSP where Vi and Vj are the nominal bus voltages, xij is the
approach, no scenario and Monte Carlo simulations are line reactance, and θi0 and θj0 are the operating points of the
needed in this approach. The unknown control scheme is phase angles. The same model can be found in [26]. Note that
embedded as a parameterized form so that the theorem (4) omits the initial deviation Pij (0) in the branch flow. In
can be used in the optimization problem. This result practice, Pij (0) cannot be arbitrary but instead must satisfy
extends the results in [24], in which the control scheme
must be given. Pij (0) = Bij (θi (0) − θj (0)) (6)
3) A convex optimization formulation of the SCP is for- for some vector θ [27].
mulated based on the series expansion of the SAFs, 3) AGC Signals: Typical AGC signals which will be used in
which consequently achieves a good trade-off between objectives of control schemes include frequency deviation and
the accuracy and the computational burden. ACE. The frequency deviation ∆f is defined as the weighted
Following this introduction, Section II presents the mod- average of the local frequencies:
P
eling of wind power uncertainty and AGC systems. Section 1 G Hi ωi
∆f = Pi∈Ω (7)
III formulates the SCP for an AGC system via the concept 2π i∈ΩG Hi
of SAFs. Section IV presents the series expansion of an SAF The area control error (ACE) of area m is defined as
and transforms the SCP into a convex optimization problem. A X
ACEm = Pij − bm ∆f (8)
case study is presented in Section V, and Section VI concludes
ij∈ΩL
the paper. m

where ACEm is the ACE of area m, ΩL m is the set of tie-lines

II. AGC M ODELING BASED ON I T Ô P ROCESSES connected to area m, and bm is the bias factor of area m.
This section models the AGC systems and stochastic re-
sources uniformly as Itô models. Similar to typical modeling
B. Wind Uncertainty Modeling Based on Itô Process
of AGC systems [25], a linear differential equation model is
used. In this section, however, the ODE model of AGC systems The distribution of the wind power uncertainty in an AGC
is embedded into an SDE model, which is consistent with the system may be non-Gaussian in practice, and thus, the model
model of stochastic resources. of the uncertainty sources should support different kinds of
probability distributions.
A. Power System Model We use an Itô process model to express the stochastic
characteristics of the wind power uncertainty. Specifically,
This subsection discusses the model of typical AGC units,
let Zt be the wind power uncertainty (as denoted in (1),
including generators, power flow models, and AGC control
but the bus number i is omited for convenience); then, its
signals.
characteristics are described by the following SDE:
1) Generators: Here we adopt a typical synchronous gen-
erator model [24], [25] dZt = µ(Zt )dt + σ(Zt )dWt (9)
 
g 1 g ωi ref where µ is the drift function that describes the determinis-
Ṗi = − g Pi + − Pi
Ti Ri tic characteristics, whereas σ is the diffusion function that
1 describes the stochastic characteristics. Wt is the integrand,
Ṗim = − t (Pim − Pig ) which is a standard Wiener process [28].
Ti (1)
  With different specifications for µ(·) and σ(·), the Itô pro-
1  X cess model can describe different probability density functions
ẇi = − Di ωi − Pim − Z (i) + PijL 
Hi (PDFs). It is proved that an Itô process model can describe
j∈Adj(i)
the uncertainties of arbitrary PDFs [24], and we provide the
In secondary frequency control, the time step is usually detailed proof and the µ(·)/σ(·) specifications for various
several seconds. Therefore, it is usually assumed that the distributions of the wind power uncertainty in Appendix A.
time constant of primary frequency control is negligible, i.e., The Itô process model can be naturally extended to cases
Tig = Tit = 0 [12]; consequently, we have where there are multiple stochastic resources with correlation.
Pig = Pim = Piref − ωi /Ri (2) We use the bold symbol Zt to represent the vector of stochastic

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
4

resources. Assume the covariance matrix of Zt is Σ, then we where x0 and z0 are the initial values of Xt and Zt ,
have √ respectively. For convenience, let Nx be the dimensionality
Zt = ΣZt0 (10) of Xt , and let Nz be the dimensionality of Zt . The major
difference between (15) and existing ODE-based AGC models
where Zt0 is the vector of independent stochastic variables, and
[8], [9], [12] is the model of the disturbance Zt , which is an
then we can model each element in Zt0 by (9).
Itô process model here. Therefore, the model used here is
The Itô process model provides an SDE form for an AGC an SDE model rather than an ODE model. This SDE model
system, which is important in the derivations throughout the describes both the stochastic characteristics and the dynamics
remainder of this paper. For this purpose, we use an Itô process of the SS in a unified way; moreover, it enables an analytical
for AGC modeling, as introduced in the next subsection. formulation of the SCP without scenario generation, as will
be discussed in the remainder of this paper.
C. AGC Modeling Based on Itô Process
III. S TOCHASTIC O PTIMAL C ONTROL OF I T Ô -AGC
1) Variable Sets: Now, we transform the models presented This section presents the SAFs to describe the objective
above into an SDE form. We first define some variable sets. function and constraints based on the model of Itô-AGC and
• Uncertain wind generations, denoted by Zt . Each element then formulates the SCP that will be studied in this paper.
of the vector Zt satisfies (9). For simplicity, we use the
following matrix form: A. Objective
dZt = µ(Zt )dt + σ(Zt )dWt (11) The objective
"Z function for AGC [12] is #
T
• Variables satisfying ordinary differential equations J =Ex0 ,z0 λACE ACEm 2
(t) + Ut> ΛU Ut dt


(ODEs), including ωi for i ∈ ΩG and Pij . These variables 0 (16)
are denoted by the vector Xt . + Ex0 ,z0 µACE ACEm
 2
(T )

• Variables satisfying algebraic equations, including
ACEm for area m and ∆f . These variables are denoted where T is the terminal time, λACE is the coefficient of
by the vector Yt . Note that Yt does not need to exist in the ACE, ΛU is the coefficient of the control variable, and
the final model since it can be substituted with Xt and µACE is the coefficient of the terminal state. Ex0 ,z0 {·} denotes
Zt in accordance with (7) and (8). the expectation operator under the initial conditions x0 and
• Control variables, i.e., Piref for i ∈ ΩG . These variables z0 , and the superscript “>” denotes the transpose operation.
are denoted by Ut . The objective function expressed in (16) shows the trade-off
between the control performance, as measured by ACEm , and
Thus, the ODEs in (3) (4) can be rewritten as:
the cost, as measured by Ut .
Ẋt = AXt + BUt + CZt (12) For convenience, we use the following general form to
where A, B and C are matrix coefficients. This linear- represent the objective function:
"Z #
differential-equation formulation has been widely used in T
previous studies [8], [9], [12]. min J = Ex0 ,z0 f (St )dt + g(ST ) (17)
0
2) Control Policy: Generally, there are two kinds of control
policies: one is state feedback control [29], [30], i.e., where St represents the concatenation of Xt , Ut , and Zt :
 
Ut = u(Xt , Zt ) (13) Xt
St = u(Zt ) (18)
 
and the other is disturbance feedback control [31]–[34], i.e.,
Zt
Ut = u(Zt ) (14)
Since the performances of these two types of control policies It is clear that (17) is a generalization of (16) since the ACE
have been proven to be equivalent [30], we adopt the distur- can be expressed as a linear combination of the state variables
bance feedback control policy (14) since it will be helpful for Xt , Ut and Zt . Moreover, f and g are both convex and of
establishing the convex optimization problem. Moreover, we up to quadratic order.
denote the set of control policies by U. The set U is determined B. Constraints of an AGC System
by the information that the controller can obtain. For example,
The constraints in an AGC system [12] include generator
a centralized control scheme assumes all the information of the
output constraints, ramping constraints, line flow constraints
state variables can be obtained by the controller, while in a
and frequency deviation constraints:
decentralized control scheme, the controller of each area can
only make decision based on the information in the same area. − P̄iG ≤ Pi,t ≤ P̄iG , ∀i ∈ ΩG
3) The Itô-AGC Model: According to (11) and (12), the − R̄iG ≤ Pi,t − Pi,t−1 ≤ R̄iG
stochastic system (SS) of interest can be formulated as (19)
− P̄ij ≤ Pij,t ≤ P̄ij , ∀i, j ∈ ΩG , i ∈ Adj(j)
SS:
dXt = (AXt + Bu(Zt ) + CZt )dt −∆ ¯ f ≤ ∆f,t ≤ ∆ ¯f
dZt = µ(Zt ) + σ(Zt )dWt (15) All of these constraints are linear constraints, and according
X0 = x0 , Z0 = z0 to the models presented in Section II-A, the variables in these

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
5

TABLE I
SCP: Eq. (25)
SAF F ORM OF THE O BJECTIVE F UNCTION AND C ONSTRAINTS

Type Pα,β (t, x0 , z0 ) α β

Objective J f g St : Eq. (15), (18) SAF: Eq. (17), (23), (24)
Constraint Ex0 ,z0 φ> >
i St St φi 0 φ>
i S >
t St φi
Theorem 1
Constraint Ex0 ,z0 φ>i St 0 φ>
i St

S
et : Eq. (26) S
bt : Eq. (27) DAF: Eq. (30)
constraints can be regarded as linear combinations of St ;
therefore, these constraints can be uniformly rewritten as:
φ>
i St ≤ φ̄i , ∀i ∈ Ω
C
(20) SCP-opt: Eq. (45)
C G E L
where Ω = Ω ∪ Ω ∪ Ω , φi is the coefficient vector
specifying the linear combination of St , and φ̄i is the upper Fig. 1. Flowchart of the proposed approach.
bound on the i-th constraint. Here, we highlight the differences between (25) and an MPC
The almost-sure constraint expressed in (20) can be natu- problem [9] for an AGC system. The major difference is that
rally relaxed to a chance constraint [35]: the SCP explicitly considers the probability distribution of the
Pr φ>

i St ≤ φ̄i ≥ γ (21) uncertainty, which is characterized by σ(Zt )dWt . Because the
According to [35], the chance constraint given in (21) has SCP considers the stochastic characteristics of the wind power
a tractable inner approximation: uncertainty, the controller u obtained via the SCP is a closed-
q loop controller that can be applied over time. In contrast, the
x0 ,z0
 >
κγ var{φ> i St } + E φi St ≤ φ̄i (22) MPC problem models the uncertainty as a predicted input, and
the result obtained via MPC can be used in only one time step.
where To solve the SCP, a key challenge is to evaluate the SAFs

2
var{φ> x0 ,z0 >
φi St St> φi − Ex0 ,z0 φ>

i St } = E i St (23) in (25). In conventional methods, stochastic terms are usually
evaluated by means of scenario-based approaches. In these
C. Formulation of Stochastic Control Problem approaches [16], the uncertainty is expressed as a set of
scenarios, and the assessment problem for each scenario can be
Before presenting the SCP, it is necessary to introduce an
solved in a deterministic way. Once the value for each scenario
SAF to describe the objective function and the constraints in a
has been obtained, the SAF is obtained by taking the average
unified form. Specifically, given that α and β are two functions
over all scenarios. However, scenario-based approaches are
of St , we define the SAF formulation for the SS in (15) as:
Z t  time-consuming, which limits their potential for use in online
applications [36]. This is the motivation for this paper, i.e., to
Pα,β (t, x0 , z0 ) = Ex0 ,z0 α(Ss )ds + β(St ) (24)
0
provide an effective way to evaluate the SAFs without scenario
generation to allow the SCP to be solved efficiently.
where the subscript s is used to replace the time variable t,
which has been used as the upper endpoint of the integral. IV. C ONVEX F ORMULATION OF SCP-AGC
Via the SAF concept, it is possible to express the objective
This section presents a deterministic convex optimization
function and constraints by specifying the functions α and β.
formulation of the SCP in (25). The basic idea of this section
• Objective function: According to (17), we can simply let is depicted in Fig. 1. A fundamental result presented in this
α = f and β = g, as is shown in Row 2 of Table I. section is the series expansion of an SAF, which makes it
• Constraints: According to (22) and (23), it is necessary possible to transform the stochastic variables in St into 2
to compute Ex0 ,z0 φ> >
i Xt Xt φi and E
x0 ,z0 >
φi Xt . Rows groups of deterministic variables, namely, S
et and Sbt , and the
3 and 4 of Table I show the specific SAF parameters for SAFs into DAFs. S et , S
bt , and the DAFs can be calculated
these 2 expectation values. in a deterministic way; thus, the SCP can be expressed as
Based on the SAF formulation of the objective function and a deterministic optimization problem, denoted by SCP-opt.
constraints, the SCP can be formulated as We then prove the convexity of SCP-opt and discuss the
SCP: computational burden of the convex optimization problem.
min J = Pf,g (T, x0 , z0 ) A. Series Expansion of an SAF
u∈U
s.t. Before introducing the series expansion of an SAF, we first
introduce two groups of auxiliary variables.
dXt = (AXt + Bu(Zt ) + CZt )dt
Let X
ft and Z et be defined as the solution to the following
dZt = µ(Zt ) + σ(Zt )dWt (25) system of differential equations:
X0 = x0 , Z0 = z0 dXft = (AX ft + Bu(Z et ) + C Z
et )dt
q
κγ P0,φ> >
i St St φi
(t, x0 , z0 ) − P0,φ>
i St
(t, x0 , z0 ) dZ
et = µ(Zet )dt (26)
C
+ P0,φ>
i St
(t, x0 , z0 ) ≤ φ̄i , ∀i ∈ Ω Xf0 = x0 , Z
e0 = z0

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
6

Let X
ct and Z bt be defined as the solution to the following formed into the computation of a series of DAFs, the com-
system of differential equations: putation of which depends only on S et and Sbt . Therefore, the
equations for St can be removed from the SCP since they can
dXct = (AX ct + B ∂u Z b + CZ bt )dt
∂Zet t be replaced with those for S et and S
bt . Moreover, Theorem 1
extends the results in [24], in which the control scheme must
dZbt = ∂µ Z b dt (27)
be given. This extension is essential due to that in the optimal
∂Z et t
control problem, the control scheme is to be determined.
X
c0 = 0N ×N , Z b0 = IN
x z z
B. Reformulation of the Objective Function and Constraints
where INz is the identity matrix of order Nz . Note that X
ct is
This subsection reformulates the objective function and
an Nx × Nz matrix, and Z bt is an Nz × Nz matrix.
constraints in accordance with Theorem 1. Since the series
For simplicity, let
 f   c  expansion of an SAF is convergent, it can be approximated
Xt Xt by a finite series. Theoretically, this approximation can reach
 ∂u b 
St = u(Zt ) , St =  e Zt  (28) an arbitrary level of accuracy as long as sufficient high-
e  e  b
∂ Zt
Zt
e Zt
b order corrections are introduced. For convenience, we use the
1st-order approximation here as an example, but the same
It is clear that S
bt = ∂ S
et /∂z0 . technique can be conveniently applied for higher orders of
Notably, St and St are both deterministic. In contrast, St is
e b approximation.
stochastic because Zt is determined by an SDE, specifically 1) Objective Function: According to Theorem 1, we have
the term σ(Zt )dWt . In practice, ordinary differential equa-
J = Je0 + Je1 (35)
tions can be solved simply via discretization, while solving Z T
SDEs is much more difficult. The following theorem makes it
Je0 = f (S
et )dt + g(S
eT ) (36)
possible to replace St with S et and S
bt in the SCP.
0

Theorem 1. Let v(t, x0 , z0 ) = Pα,β (t, x0 , z0 ). If α and Z T

1 e >
β have continuous 2nd-order derivatives, then the following Je1 = σ(Zt ) lt σ(Z
et )dt (37)
series is convergent 0 2
∞
X where lt = ∇2z0 Je0 . lt is calculated as follows:
v(t, x0 , z0 ) = ven (t, x0 , z0 ) (29) Z t
n=0
b> ∇2 f S b> ∇2 g S
   
with lt = S s
bs ds + S
t
bt (38)
0
Z t
ven = αn (S
es )ds + βn (S
et ) (30) To formulate a convex optimization problem, as discussed
0 in Section IV-C, we rewrite (36) and (38) as inequalities:
α0 = α, β0 = β (31) Z T
J0 ≥ f (S
et )dt + g(S
eT ) (39)
es ) = 1 σ(Z
e
es )> ∇2 ven σ(Z
 
αn+1 (S es ), n ≥ 0 (32) 0
z0
2 Z t
βn+1 = 0, n ≥ 0 (33) lt ≥ b> ∇2 f S
S
 
b> ∇2 g S
bs ds + S
 
bt (40)
Z t s t
0
b> ∇2 αn S b> ∇2 βn S
 2     
∇z0 ven = Ss
bs ds + S
t
bt (34)
0
It is clear that (39) and (40) are equivalent to (36) and (38)
  when J is being minimized.
where ∇2z0 ven is the 2nd-order partial derivative of ven with
2) Constraints: For Ex0 ,z0 φ>i St , we have
respect to z0 , while ∇2 αn and ∇2 βn are the Hessian matrices
x0 ,z0 > > x0 ,z0
of αn and βn , respectively. The ven are called DAFs in this E φi St = φi E St = φ>i St
e (41)
paper, as shown in Fig. 1. Now, let us consider Ex0 ,z0 φ> >
i St St φi . According to The-
orem 1, we have
The proof of Theorem 1 is rather technical and is not closely
related to the other content of this paper. Thus, we delay its Ex0 ,z0 φ> > > e e>
i St St φi =φi St St φi
presentation to Appendix B. Z t (42)
+ es )> S
σ(Z b> φi φ> S
bs σ(Z
es )ds
According to Theorem 1, an SAF can be expressed as the s i
0
sum of a series of ven , n ≥ 0. The stochastic term σ(·) does not
in which we use the fact that ∇2 φ> > >


exist in an SDE as in (15) but rather in a deterministic integral, i Ss Ss φi = φi φi .
as shown in (32). Moreover, since S et can be obtained in a According to (23), we have
Z t
deterministic way, each term in this series is a deterministic
var φ> es )> S
b> φi φ> S

i St = σ(Z s i
bs σ(Z
es )ds (43)
function of t, x0 and z0 ; i.e., no scenarios or expectation 0
values need to be considered when computing this function. Thus, the constraint expressed in (22) can be rewritten as
Therefore, Theorem 1 provides a means of computing the SAF s
Z t
in a deterministic way. κγ σ(Z b> φi φ> S
es )> S es )ds + φ> S
bs σ(Z et ≤ φ̄i (44)
s i i
By Theorem 1, the computation of an SAF can be trans- 0

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
7

C. The Optimization Problem and Its Convexity

According to the above discussion, the optimization formu-
lation of the SCP can be summarized as

SCP-opt:
min J : (35)
u∈U
s.t. : (26)(27)(37)(39)(40)(44)
Unlike in (25), there is no stochastic term in SCP-opt. By
discretizing SCP-opt over the time t, we can solve it with
commercial optimization tools.
Moreover, regarding the convexity of the SCP-opt, if f and
g are convex and of up to quadratic order and the control policy
space U is convex, then SCP-opt is convex. Actually, according
to (26) and (27), Z
et and Z bt are constants (i.e., independent of
the function u). Therefore, (26) and (27) are convex sets of S et
and St . Since f and g are convex, (39) is convex. Moreover,
b
since f and g are convex and of up to quadratic order, ∇2 f
and ∇2 g are constant positive semidefinite matrices; thus, the
right-hand term of (40) is convex and quadratic, which means
that (40) is convex. Finally, (44) is a second-order cone of S bt
and St and thus is convex. Therefore, SCP-opt is convex.
e
Remark 1. Note that the convexity of U does not require the
convexity of each u ∈ U. Therefore, there are many ways
to parametrize U. For example, the affine disturbance control
parametrization reads
u(Zt ) = Ut0 + F1> Zt
while the quadratic disturbance control parametrization reads
u(Zt ) = Ut0 + F1> Zt + Zt> F2 Zt
In (46), U is parametrized by Ut0
and F1 , while in (47), U is
parametrized by Ut0 , F1 and F2 . Therefore, U is convex with
regard to its parameters in each case, though in the 2nd case,
u ∈ U is quadratic.
D. Discussion of the Computational Burden
SCP-opt is a convex optimization problem, which can be
efficiently solved. Here, we discuss the computational burden
of the proposed approach by showing that it is as efficient
as the corresponding deterministic control (DC) problems and
much faster than the MPC and SBSP approaches. Note that
the stochastic control problem degenerates into a deterministic
control problem if σ = 0, and therefore, the constraints (37)
and (40) and the square root in (44) do not exist. Since S bt
appears only in these expressions, (27) is not needed. However,
constraints (26), (39) and (44) are still needed to obtain a
solution to the deterministic discrete control problem. The
numbers of variables and constraints are listed in Table II,
where Nx is the dimensionality of X, Nz is the dimensionality
of Z, Nc is the number of constraints, Nt is the number of
time steps in discretization, Np is the number of prediction
steps in MPC, and Ns is the number of scenarios in SBSP.
It is shown that the deterministic control problem con-
tains (Nx + Nz )Nt variables and (Nx + Nz + Nc )Nt +
1 constraints, whereas the proposed stochastic control ap-
proach involves [(Nx + Nz )(Nz + 1) + 1] Nt variables and
(45)
Fig. 2. IEEE 118-Bus System.
[(Nx + Nz )(Nz + 1) + Nc + 1] Nt +2 constraints. In practice,
we usually have Nz  Nx , i.e., the dominant sources of
uncertainty, e.g., buses that are connected to wind generators,
are much fewer than the number of state variables. Therefore,
the numbers of variables and constraints in the considered
(46) problem are close to those in the DC problem. By contrast,
the computational burden of MPC in one control step is equal
to that of the DC approach, where Nt is replaced with the
(47) number of prediction steps Np , and the total computational
burden of MPC is equal to the one-step computational burden
multiplied by the number of control steps, i.e., Nt . The SBSP
approach requires approximately Ns (Nx + Nz )Nt variables
and Ns [(Nx + Nz + Nc )Nt + 1] constraints, as shown in
Table II. Therefore, the scale of the MPC (or SBSP) problem
is approximately Np (or Ns ) times larger than that of the
DC problem. Moreover, since the computational burden grows
superlinearly with the numbers of variables and constraints,
the computational burden of the MPC and SBSP approaches
will increase rapidly as Np and Ns increase. In practice, Np
and Ns are usually large; thus, the proposed approach is much
more efficient than the MPC and SBSP approaches.
In summary, the proposed approach allows non-Gaussian
stochastic characteristics to be explicitly considered while
incurring a computational burden similar to that of DC.
V. C ASE S TUDY
This section presents a test case involving the IEEE 118-
bus system [37]. We first evaluate the effectiveness of the
reformulation made in Section IV-B and then compare the
proposed approach with conventional approaches.
A. Settings
The case considered is the IEEE 118-bus system [37], which
contains 54 generators and 186 lines, as shown in Fig. 2. The
parameters of this system can be found in [37]. The base

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
8

TABLE II
NO. OF VARIABLES AND C ONSTRAINTS

Variables Constraints
Set S
bt lt (26) (27) (37) (44) (39) (40)
SCP-opt (Nx + Nz )Nt (Nx + Nz )Nz Nt Nt (Nx + Nz )Nt (Nx + Nz )Nz Nt 1 Nc Nt 1 Nt
DC (Nx + Nz )Nt 0 0 (Nx + Nz )Nt 0 0 Nc Nt 1 0
MPC (Nx + Nz )Np Nt 0 0 (Nx + Nz )Np Nt 0 0 Nc Np Nt 1 0
SBSP Ns (Nx + Nz )Nt 0 0 Ns (Nx + Nz )Nt 0 0 Ns Nc Nt Ns 0

power is 100 MVA. The parameters of the primary control TABLE III
loop are fixed, and the supplementary control output Piref is S IMULATION R ESULTS OF S YSTEM VARIABLES
√
controllable. The frequency constraint is set to 50 ± 0.1 Hz. Variable MCS ṽ1
There are two areas in the system, of which the frequency ∆f (Hz) 0.0148 0.0145
L
bias factors are −5 pu/Hz and −1.2 pu/Hz respectively. The P77−95 (pu) 0.0452 0.0448
L
time step is 1 s [36], and the control horizon is 30 minutes. Average P80−94 (pu) 0.0332 0.0325
Standard L
The SCP-opt algorithm is performed every 100 s in order to P98−100 (pu) 0.0896 0.0888
Deviation L
P99−100 (pu) 0.0889 0.0892
maintain a good performance.
Different from the model in the algorithm, which is a ACEA (pu) 0.5464 0.5341
ACEB (pu) 0.4025 0.3987
simplified and linearized model, we use a realistic model
with detailed parameters in the MATLAB/Simulink platform
during the simulation in order to verify the proposed approach × 10-3
4

on practical settings. Such method is also widely adopted simulation

ṽ0 + ṽ1
3
in existing literature, i.e., to compute the control output via ṽ0

E ∆ 2f (Hz 2 )
simplified models, and to verify the control performance via 2

detailed models [11], [38]. Specifically, in the simulation

model, we assume the deadzones of the primary control loop
is 0.01 Hz, and the communication delay of the controller is
0.2s.
The stochastic sources are 6 wind farms on buses 6, 11,
18, 32, 55 and 100, and the nominal power of each generator
is 100 MW. The power output of each wind generator is the
sum of a predicted power output Pwpred and a fluctuating power
output ∆Pw . In accordance with [15], we assume that the wind
power uncertainty satisfies the Laplace distribution L(0, 0.05).
Moreover, We use the following equation to describe the
correlation between different wind power sources,
√
∆Pw = ΣZt (48)
where Σ is the correlation coefficient matrix, as shown in
Appendix C; and Zt is the vector of independent stochastic
resources, satisfying
(i) (i) (i)
dZt = −Zt dt + (0.1|∆Zt | + 0.005)dWt
To verify the effectiveness of the proposed method, the
analytical results obtained using the proposed approach will
be compared with the results obtained through a Monte
Carlo simulation. In the Monte Carlo simulation, 1000 wind
generation scenarios satisfying (49) were generated.
B. Accuracy of 1st-Order Reformulation
In Section IV-B, we used a 1st-order reformulation of the
SAF, i.e., v ≈ ve0 + ve1 ; in this subsection, we will validate the
effectiveness of this order of reformulation for use in the SCP.
A PI controller is used here, and the test SAF is the variance of
the frequency deviation ∆f (t) under wind power uncertainty.
Therefore, α = 0 and β = ∆2f . A simulation result obtained
as an average value over 1000 Monte Carlo samples is also
provided, and the variance of ∆f is depicted in Fig. 3. To
1
0
0 10 20 30 40 50 60 70 80 90 100
time(s)
Fig. 3. Simulated results for system frequency.
be clearer, we only depict the results in the first 100 s as an
example. It is evident that the first-order reformulation ve0 + ve1
is in good agreement with the Monte Carlo simulation, while
the zeroth-order reformulation ve0 shows a larger deviation
from the simulated value. This can be interpreted in terms of
the concept of variance. In fact, the estimated value ve0 exhibits
frequency deviations based on the predicted wind power and is
exactly (E∆f )2 , and the difference between E∆2f and (E∆f )2
is the variance of ∆f , i.e., var (∆f ). Therefore, the difference
between the first-order form and the zeroth-order form, i.e.,
ve1 , can be interpreted as the variance.
(i) In addition, we test the accuracy of SAF with respect to
(49)
other important variables, including the tie line flow and ACE.
For simplicity, we only list the average√standard deviation of
these variables in Table III, in which ṽ1 means the square
root of the variance ṽ1 , which is the standard deviation. In
summary, the first-order reformulation achieves a good approx-
imation of the SAF and captures the different characteristics
of a stochastic system by means of different terms, i.e., ve0 and
ve1 .
C. Performance and Benchmarks
This section verifies the SCP-opt represented by (45).
The objective function is given in (16), where λACE =
1000, ΛU = 7000I, and µACE = 5000. These weights are set
to large values to ensure numerical convergence. Moreover,
the affine control policy represented in (46) is used. An initial

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
control series Ut0 and a coefficient matrix F1 are computed
and are then used to determine the control output in accordance
with the actual wind power.
We test SCP-opt for two different cases, i.e. the centralized
control (SCP-opt-c), in which a centralized controller deter-
mines all the control variables; and the decentralized control
(SCP-opt-d), in which the controller of each area determines
the corresponding control variables.
Here, we use the DC, MPC and PI approaches as bench-
marks for conventional approaches. In the DC approach, an
initial control series is computed that is not subsequently
adjusted over the horizon. By contrast, in the MPC approach,
the control policy is adjusted at each step by re-solving a finite-
horizon deterministic control problem in accordance with the
actual wind power. The MPC was designed to generate 10
steps (i.e., 10 s) of predicted values. In the PI controller,
the parameters are tuned via a grid search. To illustrate the
advantages of the proposed approach over the SBSP approach,
SBSP schemes with 10 scenarios and 100 scenarios, denoted
by SBSP(10) and SBSP(100), are also considered as bench-
marks. In the SBSP approach, the control policy is the same as
that in the SCP-opt approach, but the optimal control problem
is solved via the scenario-based approach.
1) Performance of Different Control Schemes: One thou-
sand scenarios were generated to verify the performance of the
proposed approach. The control performance is summarized
in Table IV, which shows that the SCP-opt-c and SBSP(100)
achieve the best control performance. The performance of
SCP-opt-d is a little worse than that of SCP-opt-c due to that a
centralized controller can obtain the information of the entire
system, while the decentralized controller can only obtain
partial information. In comparison to the SCP-opt, the PI and
DC approaches yield worse control performance in terms of
both the objective function value and the probability that the
frequency constraint will be violated. Specifically, since this
case features high wind penetration, the PI controller cannot
avoid constraint violation. The MPC approach achieves a
performance similar to that of the proposed approach, although
it is slightly worse because the number of prediction steps
cannot be too large. SBSP(10) does not perform as well as
the SCP-opt, whereas SBSP(100) performs slightly better. In
fact, the models used in the SCP-opt and SBSP approaches are
similar, but the SCP-opt uses a series expansion to solve the
SCP, while the SBSP approach uses a scenario-based approach
to solve it.
Fig. 5 depicts the results of the SCP-opt, PI, DC and MPC
approaches for one scenario in 10 minutes. The results of the
SBSP approach are not depicted because they are similar to
those of the SCP-opt except for the computation time. Under
the DC approach, the frequency exceeds the upper limit at 37
s because the actual wind power is higher than the predicted
wind power. By contrast, in the SCP-opt and MPC approaches,
the control output is adjusted to avoid this frequency violation.
The control outputs of the SCP-opt and MPC approaches are
similar, with the difference being that the control policy is
recalculated just before 37 s in the MPC approach, whereas
the control policy and the affine disturbance coefficient are
calculated at the very beginning in the SCP-opt approach.
0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
9
7500
7000
6500
6000
5500
Objective
5000
4500
DC
4000
SCP-opt-c
3500 SCP-opt-d
PI
3000 MPC
2500
0 50 100 150 200 250 300 350 400 450 500
Time Delay (ms)
Fig. 4. Impacts of communication delay.
2) Impacts of Communication Delay: Here we discuss the
impacts of communication delay. The communication delay
of AGC systems is usually within seconds [39]. We evaluate
the proposed approach under different communication delay.
The results are shown in Fig. 4. It is shown the impact of the
communication delay on the objective is approximately linear.
When the communication delay is close to the control step, i.e.
1s, the objective function, which measures the ACE and the
controller output, increases due to that the controller cannot
response to the uncertainty immediately. Moreover, although
the communication delay influences the control performance,
the proposed approach still performs better than the others.
It is due to that the proposed approach computes a closed-
loop control scheme, which corrects the controller output to
mitigate the impacts of communication delay.
3) Computational Time: Now, we discuss the computa-
tional time, which are shown in Table IV. Due to the different
frequencies of these alrogithms being performed, we compare
the per-step computational time in order to be fair, where the
time step is 1s. The computation time of the PI controller
is not reported because this controller does not perform an
optimization process and thus is much faster than the other
approaches. The computation time of the SCP-opt is approx-
imately twice that of the DC approach. Although the average
per-step computation time of the MPC approach is only 1.53 s
because the number of prediction steps in the MPC scheme is
smaller than the control horizons of the DC and SCP-opt, the
total computation time of the MPC scheme is much greater
than that of the SCP-opt because of its receding-horizon
implementation. Moreover, the computation times for both
SBSP(10) and SBSP(100) are unacceptable. This comparison
clearly demonstrates the advantages of the proposed approach
in terms of both performance and computational efficiency.
In summary, the proposed approach achieves good AGC
performance in a computationally efficient manner. In contrast,
the DC and PI approaches are computationally efficient but
show worse performance, whereas the MPC and SBSP incur
high computational burdens. Therefore, the proposed approach
has attractive potential in online applications.
VI. C ONCLUSION
This paper presents a systematic framework for the efficient
optimal control of AGC systems. The Itô model proposed
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
10

TABLE IV where p(·) is a given probability density function (PDF). Then,

S IMULATION R ESULTS the stationary PDF of Zt is p.
Computation Objective Constraint Proof : According to (50),
Approach
Time (s per step) Value Violation (%) Z z
1 2
SCP-opt-c 0.09 4348.8 0.7 σ (z)p(z) = µ(z 0 )p(z 0 )dz 0 (51)
SCP-opt-d 0.06 4457.0 0.7 2 ∞
DC 0.04 4599.0 29.7 Taking the 2nd-order derivative with respect to z on both sides
PI —– 5087.6 100.0 yields
MPC 1.53 4522.8 0.5


1 ∂ 2 σ 2 (z)p(z) ∂ (µ(z)p(z))
SBSP(10) 2.16 4489.1 1.2 = (52)
2 ∂z 2 ∂z
SBSP(100) 10.42 4361.2 0.7
which is the steady-state Fokker-Planck equation [28] for
the diffusion process of (9). Therefore, p(·) is the stationary
distribution of Zt . 
0.2
predicted
actual Corollary 1. The Itô coefficients of the Gaussian distribution,
Power (pu)

0.15
the beta distribution, the gamma distribution, and the Laplace
0.1 distribution are as listed in Table V.
0.05
0 100 200 300 400 500 600
(a)
frequency deviation (Hz)

0.1
0.05 A PPENDIX B
DC
0
-0.05
SCP-opt-c P ROOF OF T HEOREM 1
SCP-opt-d
-0.1 PI
MPC
-0.15
0 100 200 300 400 500 600
(b)
The proof of Theorem 1 is based on the partial differential
0.04 equation (PDE) formulation of the SAF given in (24).
control output (pu)

0.02

0 DC
Lemma 1. Let v(t, x0 , z0 ) = Pα,β (t, x0 , z0 ). Given a
-0.02
SCP-opt-c
SCP-opt-d
stochastic system St defined by (15) and (18), define the
PI

-0.04
MPC following auxiliary variables:
0 100 200 300 400 500 600  
(c) " # x
time (s)
Ax0
b(x0 , z0 ) = , s0 = u(z)
 
Fig. 5. Comparison of different control approaches: (a) wind curves, (b) Bu(z0 ) + Cz0
frequency deviations, and (c) control outputs. z
Then, the SAF v satisfies the following equation:
in this paper is able to describe AGC systems with vari-
∂v ∂v ∂v
ous non-Gaussian wind power uncertainty. Based on the Itô =b(x0 , z0 )> + µ(z0 )>
model, a theorem of convergent series expansion of SAFs is ∂t ∂x0 ∂z0
1 >
 2  (53)
proven, which enables the design of a highly efficient optimal + σ(z0 ) ∇z v σ(z0 ) + α(s0 )
stochastic control algorithm of Itô AGC systems, and allows 2
SCPs to be solved with a computational burden comparable v(0, x0 , z0 ) =β(x0 , z0 )
to that for deterministic control problems without sacrificing This lemma is the famous Feynman-Kac formula. The proof
performance. Computational applications show that the pro- is omitted here; instead, the reader is referred to [28].
posed approach outperforms some popular control algorithms
while incurring affordable computational burden, and thus has Corollary 2. ven , as defined in (30), satisfies the following
attractive potential in the online applications of AGC systems. PDE:
∂e
vn ∂e
vn ∂evn
= b(x, z)> + µ(z)> + αn (s)
∂t ∂x ∂z (54)
A PPENDIX A ven (0, x, z) = βn (s)
T HE I T Ô C OEFFICIENTS OF AN A RBITRARY D ISTRIBUTION
Proof : According to (30), the expression for ven is similar to
This section summarizes the results in [24, Sec. II.A]. that for Pα,β except that St is replaced by S et . According to
Suppose that the PDF of Zt is p(z). Then, we have the (26) and (28), the equations for Set differs from those for St
following proposition: only by the removal of the stochastic term σ(·); therefore, by
removing σ from (53), we obtain (54). 
Proposition 1. Suppose that the functions µ and σ satisfy
Rz According to Lemma 1 and Corollary 2, it suffices to prove
µ(z 0 )p(z 0 )dz 0 that the v described by (53) can be expressed as a sum of a
σ (z) = 2 −∞
2 (50) series of the ven described by (54). To achieve this goal, we
p(z)

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
11

TABLE V
S PECIFIC F ORMULATIONS OF S OME T YPICAL D ISTRIBUTIONS
Distribution PDF µ(z) σ 2 (z)
√1 exp (z − a)2 /2b
 
Gaussian 2πb
−(z − a) 2b
Beta 1
B(a,b)
z a−1 (1 − z)b−1 a
−(z − a+b ) 2
a+b
z(1 − z)
ba
Gamma Γ(a)
z a−1 exp(−bz) −(z − a/b) 2z/b
Laplace 1
2b
exp [−|z − a|/b] −(z − a) 2b|z − a| + 2b2
a−1 a ab2 z a
a
b2 Γ 1 + k2 − Γ 1 + k1
 a−1
Γ 1√+ k2 − Γ 1 + k1



 


Weibull b
(z/b) exp [−(z/b) ] Iz>0 z a
− b
d
ESGN [40] ESGN(z; a, b, c) −z + dz log Φ(az/ 1 + bz 2 + cz 4 ) 2

present a parametrized PDE of w(t, x0 , z0 ; ): By applying (60) to (59) and taking  = 0, we obtain
∂ n+1 w 
 
∂w ∂w ∂w

∂ 1
=b(x0 , z0 )> + µ(z0 )>
∂t ∂x0 ∂z0 ∂t (n + 1)! ∂n+1 =0
1 >
 2  (55) ∂

1 ∂ n+1 w 
 
+ σ(z0 ) ∇z w σ(z0 ) + α(s0 ) =b(x0 , z0 )>
2 ∂x0 (n + 1)! ∂n+1 =0
w(0, x0 , z0 ) =β(x0 , z0 ) (61)
∂ n+1 w 
  
∂ 1
Regarding w, we have Lemmas 2 and 3: + µ(z0 )>
∂z0 (n + 1)! ∂n+1 =0
1 ∂ n w 
  
Lemma 2. Consider the Taylor expansion of w in : 1
+ σ(z0 )> ∇2z0 σ(z0 )
∞ 2 n! ∂n  =0
1 ∂ n w 
  
X
w(t, x0 , z0 ; ) = n (56)
n! ∂n 
=0
Moreover, by applying (32) and (33) to (54), we obtain
n=0
∂e
vn+1 ∂e
vn+1 ∂e
vn+1
then this Taylor expansion is convergent. =b(x0 , z0 )> + µ(z0 )>
∂t ∂x0 ∂z0 (62)
Proof. This lemma has been proved in [24], but here we 1
+ σ(z0 )> ∇2z0 ven σ(z0 )
 
provide a more concise proof. To achieve this, we introduce 2
an auxiliary vector δ = ∇z w, then (55) can be rewritten as:
By comparing (62) with (61) and recalling that Lemma 3
∂w ∂w ∂w holds for n, we conclude that
=b(x0 , z0 )> + µ(z0 )>
∂t ∂x0 ∂z0
∂ n+1 w 

1
1 (57) ven+1 = (63)
+ σ(z0 )> [∇z δ] σ(z0 ) + α(s0 ) (n + 1)! ∂n+1 
2 =0
0 =δ − ∇z w Therefore, Lemma 3 also holds for n + 1, which completes
which is a group of first-order PDEs. Then, according to the proof. 
Cauchy-Kovalevskaya Theorem, the solution of (57) can be Proof of Theorem 1: it is clear that v(t, x0 , z0 ) =
expanded into the power series of its independent variables. w(t, x0 , z0 ; 1); therefore, by setting  = 1 in (56), we obtain
Therefore, we have directly finished the proof. ∞
1 ∂ n w 
X   
v(t, x0 , z0 ) = (64)
n! ∂n =0
Lemma 3. The following equality holds for ven and w: n=0
P∞
1 ∂ n w  Then, according to Lemma 3, we have v = n=0 v
en , and

ven = (58) according to Lemma 2, the series is convergent.
n! ∂n  =0
Formula (34) remains to be proven. According to (30),
Proof : when n = 0, this lemma is trivial. Now, we assume Z th
that the lemma holds for ven ; it is necessary to prove that the  2 i h i
∇2z0 αn (S
es ) ds + ∇2 βn (S

∇z0 ven = z0
et ) (65)
lemma holds for ven+1 . 0
Taking the (n + 1)-th derivative with respect to  in (55)
When µ is affine, we have
yields
" #> " #
∂ ∂ n+1 w ∂ ∂ n+1 w ∂ ∂ n+1 w h
2
i ∂S
es  2  ∂S es
n+1
=b(x0 , z0 )> n+1
+ µ(z0 )> ∇z0 αn (Ss ) =
e ∇ αn (66)
∂t ∂ ∂x0 ∂ ∂z0 ∂n+1 ∂z0 ∂z0
n+1
 
1 ∂
+ σ(z0 )> ∇2z0 n+1 (w) σ(z0 ) Therefore,
2 ∂ h i
(59) ∇2z0 αn (S b> ∇2 αn S
es ) = S
 
bs (67)
s
Consider the following equality:
∂ n+1 (w) ∂nw ∂ n+1 w A similar procedure can be applied to ∇2z0 βn (S
et ). Thus,
= (n + 1) +  (60) we complete the proof of Theorem 1. 
∂n+1 ∂n ∂n+1

0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2019.2893512, IEEE
Transactions on Power Systems
A PPENDIX C
C ORRELATION C OEFFICIENT OF W IND P OWER
The Σ in (48) is
 
1 0.3 0.2 0 0 0
0.3 1 0.4 0 0 0
 

0.2 0.4 1 0 0 0
 frequency regulation by electric vehicles,” in 2012 North American
Σ=
 
0 0 0 0 0 0
 [20] G. Gross and J. Lee, “Analysis of load frequency control performance
 
0 0 0 0 0 0
 
0 0 0 0 0 0
where the orders of the wind farms are Buses 6, 11, 18, 32,
55, 100.
R EFERENCES
[1] Y. Markorov, R. Diao, P. Etingov, S. Malhara, N. Zhou, R. Guttromson,
J. Ma, P. Du, N. Samaan, and C. Sastry, “Analysis Methodology
for Balancing Authority Cooperation in High Penetration of Variable
Generation,” tech. rep., Pacific Northwest National Library, 2010.
[2] Z. Wang, W. Wu, and B. Zhang, “A fully distributed power dispatch
method for fast frequency recovery and minimal generation cost in
autonomous microgrids,” IEEE Trans. Smart Grid, vol. 7, pp. 19–31,
Jan 2016.
[3] C. Wan, Z. Xu, P. Pinson, Z. Y. Dong, and K. P. Wong, “Optimal
prediction intervals of wind power generation,” IEEE Trans. Power Syst.,
vol. 29, pp. 1166–1174, May 2014.
[4] B. K. Sahu, S. Pati, and S. Panda, “Hybrid differential evolution particle
swarm optimisation optimised fuzzy proportional integral derivative
controller for automatic generation control of interconnected power
system,” IET Generation, Transmission Distribution, vol. 8, no. 11,
pp. 1789–1800, 2014.
[5] Y. Xu, F. Li, Z. Jin, and M. H. Variani, “Dynamic gain-tuning control
(dgtc) approach for agc with effects of wind power,” IEEE Trans. Power
Syst., vol. 31, pp. 3339–3348, Sept 2016.
[6] E. Yao, V. W. S. Wong, and R. Schober, “Robust frequency regulation
capacity scheduling algorithm for electric vehicles,” IEEE Trans. Smart
Grid, vol. 8, pp. 984–997, March 2017.
[7] G. Sharma, I. Nasiruddin, K. R. Niazi, and R. C. Bansal, “Robust
automatic generation control regulators for a two-area power system
interconnected via ac/dc tie-lines considering new structures of matrix
q,” IET Generation, Transmission Distribution, vol. 10, no. 14, pp. 3570–
3579, 2016.
[8] M. Ma, C. Zhang, X. Liu, and H. Chen, “Distributed model predictive
load frequency control of the multi-area power system after deregula-
tion,” IEEE Trans. Ind. Electron., vol. 64, pp. 5129–5139, June 2017.
[9] A. Ersdal, L. Imsland, and K. Uhlen, “Model predictive load-frequency
control,” IEEE Trans. Power Syst., vol. 31, pp. 777–785, Jan 2016.
[10] S. R. Cominesi, M. Farina, L. Giulioni, B. Picasso, and R. Scattolini,
“A two-layer stochastic model predictive control scheme for microgrids,”
IEEE Trans. Control Syst. Technol., vol. 26, pp. 1–13, Jan 2018.
[11] P. McNamara and F. Milano, “Model predictive control-based agc
for multi-terminal hvdc-connected ac grids,” IEEE Trans. Power Syst.,
vol. 33, pp. 1036–1048, Jan 2018.
[12] H. Jiang, J. Lin, Y. Song, S. You, and Y. Zong, “Explicit model predictive
control applications in power systems: an agc study for an isolated
industrial system,” IET Generation, Transmission Distribution, vol. 10,
no. 4, pp. 964–971, 2016.
[13] E. F. Camacho and C. B. Alba, Model predictive control. Springer
Science & Business Media, 2013.
[14] M. J. Hadjiyiannis, P. J. Goulart, and D. Kuhn, “An efficient method
to estimate the suboptimality of affine controllers,” IEEE Trans. Autom.
Control, vol. 56, pp. 2841–2853, Dec 2011.
[15] H. Bludszuweit, J. a. Dominiguez-Navarro, A. Llombart, and J. a.
Dominguez-Navarro, “Statistical Analysis of Wind Power Forecast Er-
ror,” IEEE Trans. Power Syst., vol. 23, no. 3, pp. 983–991, 2008.
[16] L. Yang, M. He, V. Vittal, and J. Zhang, “Stochastic optimization-based
economic dispatch and interruptible load management with increased
wind penetration,” IEEE Trans. Smart Grid, vol. 7, pp. 730–739, March
2016.
0885-8950 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
12
[17] Y. Fu, M. Liu, and L. Li, “Multiobjective stochastic economic dispatch
with variable wind generation using scenario-based decomposition and
asynchronous block iteration,” IEEE Trans. Sustain. Energy, vol. 7,
pp. 139–149, Jan 2016.
[18] J. Engels, B. Claessens, and G. Deconinck, “Combined stochastic
optimization of frequency control and self-consumption with a battery,”
IEEE Transactions on Smart Grid, vol. PP, no. 99, pp. 1–1, 2017.
[19] J. Donadee and M. Ilić, “Stochastic co-optimization of charging and
Power Symposium (NAPS), pp. 1–6, Sept 2012.
(68)
assessment criteria,” IEEE Trans. Power Syst., vol. 16, pp. 520–525,
Aug 2001.
[21] L. R. Chang-Chien, C. C. Sun, and Y. J. Yeh, “Modeling of wind farm
participation in agc,” IEEE Trans. Power Syst., vol. 29, pp. 1204–1211,
May 2014.
[22] M. C. Campi and S. Garatti, “The exact feasibility of randomized
solutions of uncertain convex programs,” SIAM Journal on Optimization,
vol. 19, no. 3, pp. 1211–1230, 2008.
[23] M. D. Tabone and D. S. Callaway, “Modeling variability and uncertainty
of photovoltaic generation: A hidden state spatial statistical approach,”
IEEE Trans. Power Syst., vol. 30, pp. 2965–2973, Nov 2015.
[24] X. Chen, J. Lin, F. Liu, and Y. Song, “Stochastic assessment of agc
systems under non-gaussian uncertainty,” IEEE Trans. Power Syst.,
vol. 34, no. 1, 2019.
[25] H. Jiang, J. Lin, Y. Song, and D. J. Hill, “Mpc-based frequency control
with demand-side participation: A case study in an isolated wind-
aluminum power system,” IEEE Trans. Power Syst., vol. 30, pp. 3327–
3337, Nov 2015.
[26] N. Li, C. Zhao, and L. Chen, “Connecting automatic generation control
and economic dispatch from an optimization view,” IEEE Trans. Control
Netw. Syst., vol. 3, pp. 254–264, Sept 2016.
[27] C. Zhao, U. Topcu, N. Li, and S. Low, “Design and stability of load-
side primary frequency control in power systems,” IEEE Trans. Autom.
Control, vol. 59, pp. 1177–1189, May 2014.
[28] E. Pardoux and A. Răşcanu, Stochastic Differential Equations, Backward
SDEs, Partial Differential Equations. Springer, 2014.
[29] M. Cannon, B. Kouvaritakis, and X. Wu, “Probabilistic constrained
mpc for multiplicative and additive stochastic uncertainty,” IEEE Trans.
Autom. Control, vol. 54, pp. 1626–1632, July 2009.
[30] J. Skaf and S. P. Boyd, “Design of affine controllers via convex
optimization,” IEEE Trans. Autom. Control, vol. 55, no. 11, pp. 2476–
2487, 2010.
[31] D. Chatterjee, P. Hokayem, and J. Lygeros, “Stochastic receding horizon
control with bounded control inputs: A vector space approach,” IEEE
Trans. Autom. Control, vol. 56, pp. 2704–2710, Nov 2011.
[32] H. Ding, P. Pinson, Z. Hu, and Y. Song, “Optimal offering and operating
strategies for wind-storage systems with linear decision rules,” IEEE
Trans. Power Syst., vol. 31, pp. 4755–4764, Nov 2016.
[33] F. Oldewurtel, C. N. Jones, and M. Morari, “A tractable approxima-
tion of chance constrained stochastic mpc based on affine disturbance
feedback,” in 47th IEEE Conference on Decision and Control, pp. 4731–
4736, Dec 2008.
[34] M. Korda, R. Gondhalekar, J. Cigler, and F. Oldewurtel, “Strongly
feasible stochastic model predictive control,” in 50th IEEE Conference
on Decision and Control and European Control Conference, pp. 1245–
1251, Dec 2011.
[35] G. C. Calafiore and L. El Ghaoui, “On distributionally robust chance-
constrained linear programs,” Journal of Optimization Theory and Ap-
plications, vol. 130, no. 1, pp. 1–22, 2006.
[36] X. Chen, J. Lin, C. Wan, Y. Song, and H. Luo, “A unified frequency-
domain model for automatic generation control assessment under wind
power uncertainty,” IEEE Trans. Smart Grid, early access.
[37] “IEEE 118-Bus System.” Available at http://icseg.iti.illinois.edu/
ieee-118-bussystem/.
[38] A. N. Venkat, I. A. Hiskens, J. B. Rawlings, and S. J. Wright, “Dis-
tributed mpc strategies with application to power system automatic gen-
eration control,” IEEE Trans. Control Syst. Technol., vol. 16, pp. 1192–
1206, Nov 2008.
[39] D. Ganger, J. Zhang, and V. Vittal, “Forecast-based anticipatory fre-
quency control in power systems,” IEEE Trans. Power Syst., vol. 33,
pp. 1004–1012, Jan 2018.
[40] K. Choudhury and M. A. Matin, “Extended skew generalized normal
distribution,” Metron, vol. 69, no. 3, pp. 265–278, 2011.