Escolar Documentos
Profissional Documentos
Cultura Documentos
ABSTRACT This paper presents a novel distributed secondary control method for both voltage and
frequency regulation in islanded microgrids. Firstly, the large-signal dynamic model of inverter-interfaced
distributed generation (DG) is formulated in the form of a multi-input multi-output nonlinear system, which
can be converted to a partly linear one using input–output feedback linearization. Then, the linear-distributed
model predictive controller is designated in each DG to realize the secondary voltage control by incorporating
the forecasted behaviors of the local and neighboring DG units. Through the receding optimization index
of every update process, the implementation of optimal control action accelerates the convergence rate for
voltage magnitudes to the reference value. Following, after transforming the nonlinear DG dynamics into a
first-order linear system, a distributed proportional integral algorithm is introduced in the frequency restora-
tion while maintaining the accurate active power sharing. Our approach utilizes the distributed architecture,
which indicates superior reliability and flexibility compared to the centralized manner; moreover, it can
accommodate diverse uncertainties in communication links, model parameters, and time delays. Simulation
results are provided to verify the effectiveness of the proposed control methodology.
INDEX TERMS Cooperative control, distributed model predictive control, input-output feedback
linearization, islanded microgrids, secondary control.
2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 6, 2018 Personal use is also permitted, but republication/redistribution requires IEEE permission. 50169
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
G. Lou et al.: Distributed Model Predictive Secondary Voltage Control of Islanded MGs
multi-agent system has earned much attention [7], [14]–[22] control is presented to regulate frequency while maintaining
where each DG unit exchanges information with its imme- accurate active power sharing. The study contributions are
diate neighbors upon the sparse communication topology. listed below.
In [7], a distributed-averaging proportional-integral (DAPI) (1) In this paper, a large-signal dynamic model of
controller was proposed to compensate the voltage and fre- inverter-interfaced DG is presented in the form of multi-input
quency deviations. A distributed networked approach [14] multi-output (MIMO) nonlinear system, which can be trans-
was utilized to realize the secondary voltage and frequency formed to a partly linear system using the input-output feed-
control, where each DG communicates with all DGs, and back linearization.
takes the averaged values as the nominal states to complete (2) Upon on the linear system derived via IOFL, the linear
the information update process. Based on the principle of DMPC approach is proposed for the secondary voltage con-
distributed cooperative control, the secondary voltage control trol of islanded MGs by incorporating the predictive mecha-
was converted into the linear tracker synchronization prob- nisms into DGs. Based on the forecasted information of local
lem using feedback linearization [15], [16], which however and neighboring DG units, the receding optimization index
has been geared toward the microgrid system with fixed is formulated to derive the latest optimal control action at
and accurate parameters. Further, to improve the dynamic each sampling instant. Compared with the existing literature,
performance during the voltage and frequency regulation, the proposed methodology not only accelerates the conver-
distributed secondary control based on finite-time algorithm gence process but also accommodates the uncertainties in
[17], [18], dynamic weights [19], and sliding-mode law [20] communication links and time delays.
have been researched. Alternatively, considering the uncer- (3) The proposed approach is fully distributed. Each DG
tainties of system parameters, communication links and only communicates with its immediate neighbors through a
time delays in practice, novel distributed secondary control sparse network, obviating the requirement for a supervisory
schemes [21]–[25] were proposed to accommodate the vari- centralized controller.
ations in operation conditions of islanded microgrids. The rest of the paper is organized as follows. Section II
To summarize up, the main challenges with respect to presents the large-signal dynamic model of inverter-based
the secondary control of MG lie in: 1) Distributed cooper- DG. Section III introduces the preliminary of graph theory
ative control scheme is expected to necessitate the coordi- and input-output feedback linearization. The secondary con-
nation process of voltage and frequency restoration while trol for voltage and frequency regulation based on distributed
maintaining accurate power sharing through spare commu- MPC and proportional-integral is elaborated respectively in
nication networks. 2) The convergence of both voltage and Section IV. Section V verifies the proposed secondary control
frequency regulation should be realized synchronously and using a test MG system; finally a brief conclusion is drawn in
rapidly. 3) It is desirable to have a robust control paradigm Section VI.
to accommodate the diverse uncertainties of DG parame-
ters, communication links and time delays. Model predic- II. LARGE-SIGNAL DYNAMIC MODEL OF
tive control [26], [27] has been well-accepted in industry INVERTER-INTERFACED DG
as an effective methodology for both dynamic performance The proposed secondary control scheme is designed based
and robustness against system variations; however, the most on the large-signal dynamic model of inverter- based DG.
existing literature with regard to secondary control in MGs In general, each DG deemed as a prime dc source is inter-
neglect the predictive characteristics of individual DG agents. faced to the MG network through a voltage source inverter
To the best of the authors’ knowledge, the work concerning (VSI), a LC filter and a RC output connector [29], as shown
the forthputting of model predictive mechanism in this aspect in Fig. 1. Broadly speaking, there exist three control loops in
was only presented in [28], but it was developed based on the the primary controller for each DG unit, i.e., the power loop,
consensus tracker rule without considering the nonlinear and the voltage loop and the current control, where the power
nonidentical dynamics of individual DG units. controller provides the magnitude and frequency references
In this paper, a distributed model predictive con- for the output voltage, while the nested voltage and current
trol (DMPC) scheme is proposed for the voltage restoration
in islanded MGs. Firstly, considering the nonlinear dynamics
of heterogeneous DG units, the input-output feedback lin-
earization (IOFL) is utilized to convert the nonlinear system
to a partly linear one. Then we designate the linear model
predictive controller for each DG based on the forecasted
movements of the local and neighboring DG units. Through
the receding-horizon optimization index, the optimal control
effect allows the acceleration of synchronization process
and accommodates the variations in the operation condition.
After transforming the nonlinear model into a first-order
linear system, the distributed proportional integral frequency FIGURE 1. Block diagram of an inverter-interfaced DG.
controller is utilized to improve the voltage performance and where xi = [Pi , Qi , ωi , ildi , ilqi , vodi , voqi , iodi ioqi ]T ; ui =
reject system disturbances. With the consideration that the [ui1 ui2 ]=[uvi uωi ]T , yi = [yi1 yi2 ]T = [vodi ωi ]T and Di =
current and voltage control possess much faster dynamics [vbdi vbqi ]T represent the inputs, outputs and known distur-
than the power control [8], [17], the DG can be represented bances, respectively. The detailed expressions of fi (xi ), ki (xi ),
as a controllable voltage source, which prescribes a desired gi1 (xi ) and gi2 (xi ) can extracted from (3)-(11). Since the
relationship between active power Pi and frequency ωi , and DG output voltage holds the condition v2oi = v2odi + v2oqi ,
between reactive power Qi and output voltage v∗oi , aligning secondary control aims to select uvi and uωi such that vodi and
the magnitude on the d-axis of the reference frame, ωi approach their reference values, which are elucidated in
following sections.
ωi = ωni − mPi Pi (1)
v∗odi = Vni − nQi Qi , v∗oqi = 0 (2) III. PRELIMINARY OF SECONDARY CONTROL SCHEME
A. GRAPH THEORY AND TRACKE SYNCHRONIZATION
where mPi and nQi are the frequency and voltage droop coef- To implement the cooperative control of DGs, the microgrid
ficients, respectively; ωni and Vni are the desired frequency is considered as a multi-agent system, where each DG can
and voltage for the ith DG, respectively; v∗odi and v∗oqi are the communication with others through a spare communication
d-axis and q-axis components of v∗oi , respectively. network G = (V, E, A), with a set of n-nodes V = {v1 ,
Pi and Qi are extracted from the instantaneous powers via v2 , . . . , vn }, a set of edges E ⊂ V × V , and the associated
the two low-pass filters with the cutoff frequency ωco as adjacency matrix A = [aij ]. In a MG graph, the nodes and
Ṗi = −ωco Pi + ωco (vodi iodi + voqi ioqi ) (3) edges represent the DG units and the corresponding commu-
nication links, respectively. Each edge (vi , vj ) ∈ EG signifies
Q̇i = −ωco Qi + ωco (voqi iodi − vodi ioqi ) (4)
that node i receives information from nodej with the weight
wherevodi and voqi (resp., iodi and ioqi ) are output voltage aij > 0, otherwise aij = 0. Typically, node i obtains informa-
(resp., current) of ith DG on the d-q reference frame. tion only from its immediate neighbors, denoted as Ni , {j|(i,
Substitute (1) into (3) and introduce secondary frequency j) ∈ EG }. The Laplacian matrix of the graph, which encodes
control input uωi for frequency restoration, which yields, the information flow, P is defined as L = D − AG with D =
diag{di } and di = j∈Ni aij . Denoting the value of node i as
ω̇i − ωco (ωni + uωi − ωi ) + ωco mPi (vodi iodi + voqi ioqi ) = 0 xi , the cooperative secondary control resembles the tracker
(5) synchronization problem, where all nodes are expected to
converge to the references based on the information coupling,
Following, the differential equations for the LC filter and i.e., xi = xj = xref , ∀i, j ∈ N .
output connector while neglecting the fast dynamics of volt-
age and current controller and considering the secondary B. INPUT-OUTPUT FEEDBACK LINEARIZATION
voltage control are formulated as follows [21], [30]: Since the dynamics of DGs in a microgrid are nonlin-
Rfi 1 ear and heterogeneous, input-output feedback linearization
i̇ldi = − ildi +ωi ilqi + (Vni +uvi − nQi Qi − vodi ) (6) (IOFL) [15] can be utilized in this paper to facilitate the
Lfi Lfi
secondary controller design without linearizing the DG
Rfi voqi
i̇lqi = − ilqi − ωi ildi − (7) dynamics around the prior steady-state operating point. The
Lfi Lfi principle of feedback linearization is to transform a nonlinear
ildi − iodi system to a fully or partly linear one by means of a suitable
v̇odi = + ωi voqi (8)
Cfi coordinate conversion, of which the point of departure is the
ilqi − ioqi notation of the system relative degree.
v̇oqi = − ωi vodi (9) Define 1 [32]: The affine nonlinear system
Cfi
Rci 1 ẋ = f (x) + g(x)u
i̇odi = − iodi + ωi ioqi + (vodi − vbdi ) (10)
Lci Lci y = h(x) (13)
Rci 1
i̇oqi = − ioqi − ωi iodi + (voqi − vbqi ) (11) is said to have relative degree r at point if x0
Lci Lci
(i) Lg Lfk h(x) = 0 for all x in a neighborhood of x 0 and
where ildi , ilqi , vbdi and vbqi are the direct and quadrature terms k < r − 1;
of the inverter current ili and bus voltage vbi in Fig. 1 [15], [31] (ii) Lg Lfr−1 h(x) = 0
respectively. uvi represents the secondary voltage control where Lf h(x) is the Lie derivative of h(x) with respect to
input. f (x), defined as Lf h(x) = ∇hf (x) = (∂h/∂x)f (x), Lfk h(x) =
By combining equations (3)-(11), a large-signal dynamic (∂(Lfr−1 h)/∂x)f (x), and Lg Lfk h(x) = (∂(Lfk h)/∂x)g(x).
model of ith DG can be arranged in compact MIMO system, Thus, we consider the coordinate transformation as
ẋi = fi (xi ) + ki (xi )Di + gi1 (xi )ui1 + gi2 (xi )ui2 follows:
yi1 = hi1 (xi ) = vodi , yi2 = hi2 (xi ) = ωi (12) zp = [h(x), Lf h(x), · · · , Lfr−1 h(x)] (14)
Considering a nonlinear system with relative degree r = n, yi1 0 1
where ziv = , Aiv = , Biv = [0 1]T and
the coordinate transformation which is required to construct yi1 0 0
a normal form is given as zi = Lfk h(x), 1 ≤ i ≤ n, which Civ = [1 0]. In the distributed control, it is assumed that
results in an exactly linear system. If r < n holds, it is each DG exchanges information with its immediate neighbors
always possible to find zr+1 , . . . , zn such that the mapping through a sparse communication network and only a small
z = [z1 , . . . , zn ] has a jacobian matrix which is nonsingular portion of DG units have direct access to the reference value
and qualifies to be partial feedback linearized [33]. We can vref . Hence, the essence of the secondary voltage control is
further decompose the states into z = [zp zn−p ]T , where zp to find a distributed vi in (20) so that yi1 → vref , ∀i. To solve
represents the r functions zi = Lfr−1 h(x) (1 ≤ i ≤ r) that this problem, the linear distributed model predictive control
are termed as a partial set of the new coordinate form while using input-output feedback linearization is presented in this
zn−p signifies the dynamics of the remaining (n − r) orders, paper.
denoted as the system internal dynamics. Therefore, it is The foremost step for the proposed distributed control is
convenient to convert the controller design of the nonlinear to determine the discrete-time model, which is formulated
system (13) to the linear controller for the following partially to facilitate the iterative characteristics of model predictive
linearized counterpart: control-based algorithm. The resulting discrete time system
of (20) can be arranged as follows,
ż1 = z2
··· ziv (k + 1) = Aid ziv (k) + Bid vi (k)
żr−1 = zr yi1 (k) = Cid ziv (k) (21)
żr = b(z) + a(z)u (15)
where Aid , Bid and Cid are system matrices obtained through
where b(z) = Lfr h(8−1 (z)) and a(z) = Lg L1r− f h (8−1 (z)), Euler discretization.
with 8−1 (z) = x. Hence, it can be arranged in a state space Denoting HP as the prediction horizon and HU as the
form as: control horizon, the future evolution of output value yi1 can
be derived according to the available current state ziv (k) and
żp = Azp + Bv (16) the protocol (21) as follows:
where A and B are the matrices for states and input, respec- yi1 (k + 1|k) = Cid Aid ziv (k) + Cid Bid vi (k) (22)
tively; v is the auxiliary control input for the partial linearized
yi1 (k + 2|k) = Cid A2id ziv (k) + Cid Aid Bid vi (k)
system in the transformed coordinate.
+ Cid Bid vi (k + 1) (23)
IV. PROPOSED DISTRIBUTED SECONDARY CONTROL ...
HP −1
In this paper, a distributed secondary control scheme is imple- yi1 (k + HP |k) = Cid AH P
id ziv (k) + Cid Aid Bid vi (k)
mented locally by choosing the appropriate control inputs uvi HP −HU
+ · · · + Cid Aid Bid vi (k + HU − 1)
and uωi to regulate the voltage magnitude vodi and frequency (24)
ωi to the reference values vref and ωref respectively, which
can be considered as a tracker consensus problem. Define vectors Yi1 (k + 1, HP |k) = [yi1 (k + 1|k)yi1 (k +
2|k) . . . yi1 (k + HP |k)]T and Vi (k, HU |k) = [vi (k)vi (k +
A. LINEAR DMPC-BASED VOLTAGE REGULATION 1) . . . vi (k + HU − 1)]T , equations (22)-(24) can be written
For the voltage dynamics in (12), the direct relationship in a compact form:
between the output yi1 and input ui1 can be generated after
Yi1 (k + 1, HP |k) = Fi ziv (k) + Gi Vi (k, HU |k) (25)
the second derivative of yi1 , indicating that the relative degree
for voltage control is r = 2: where Fi = [Cid Aid Cid A2id . . . Cid AH P T
id ]
where |Ni | represents the number of the neighbors of ith DG; Remark 1: The selection of wi indicates whether the ith
Yr (k) = [r(k), r(k + 1), . . . , r(k + HP )]T is the desired DG has direct access to the reference value vref , similar to
output value during the predictive horizon, here Yr (k) = the pinning gain, that is, if the ith DG directly receive the
vref ∗ IHp with IHp as the HP dimensional vector of one- reference information, wi > 0, otherwise wi = 0.
entries; 1Vi (k, HU |k) = [1vi (k), 1vi (k + 1), . . . , 1vi (k + Remark 2: The computational complexity to obtain the
HU − 1)]T with 1vi (k + h) = vi (k + h) − vi (k + h−1), explicit solution mainly lies in the inversion of Hi , which can
h = 0, . . . , HU − 1. In this optimization formulation, the first be performed offline for the online calculation of the control
term penalizes the state disagreement between the ith DG and law vi (k).
its |Ni | neighboring counterparts; the secondary term stresses Therefore, the routine procedure of IOFL-based linear
the local voltage deviation from the reference value; while the DMPC algorithm can be depicted as Fig. 2:
third signifies the control energy over the future HU steps. Qi ,
Ri and Wi are the corresponding positive definite symmetric
weighting matrices for the three terms, with ||∗||2Qi = ∗T Qi ∗.
For simplicity, these matrices can be selected as Qi = qi I ,
Wi = wi I , and Ri = ri I (qi , wi , ri > 0), respectively.
Considering thatP the control effect can be obtained as vi (k+
h) = vi (k − 1) + hr = 01vi (k + r), the future evolution
of system (25) can be expressed as a function of the control
incremental sequence 1Vi (k, HU |k),
Yi (k + 1, HP |k) = Ni 1Vi (k, HU |k) + Mi (k) (27)
where Ni = Gi 0̄i , Mi (k) = Fi ziv (k) + Gi 0i0 vi (k − 1), and
1 1 ··· 0
0i0 = ... 0̄i = ... ..
.
..
.
1 1 ··· 1
Substitute (27) into the local cost function Ji (26), which FIGURE 2. The routine process of IOFL-based DMPC algorithm.
is formulated into an equivalent unconstrained quadratic pro-
gram (QP) problem and expected to be solved at each sam- Step 1: Each DG determines the discrete-time model (21),
pling instant, and acquires the local states xi and the output voltages of its
neighboring DGs over the current time k.
min Ji = 1ViT (k, HU |k)Hi 1Vi (k, HU |k)
1Vi (k,HU |k) Step 2: Each DG computes the optimal control sequence
− LiT (k)1Vi (k, HU |k) (28) 1Vi (k, HU |k) according to the rolling optimization index (26)
and applies the first element to generate the current control
where the positive definite matrix Hi has the form effort vi (k) according to (30).
Hi = NiT (Qi + Wi )Ni + Ri Step 3: Calculate the control input ui1 (k) through the
feedback law (19) and implement it on the local primary
and controller.
Li (k) = 2NiT [Wi (Mi (k) − Yr (k)) + Qi (Mi (k) Step 4: Increment the sampling time index k → k + 1, and
1 X repeat the above steps.
− Yj (k + 1, HP |k))] Owing to the line impedance effect, voltage restoration
|Ni |
j∈Ni
and accurate reactive power sharing could not be achieved
Let ∂Ji /∂1Vi (k, HU |k) = 0, and the explicit solution can simultaneously; therefore, a trade-off is established between
be derived with the form the voltage regulation and reactive power sharing [7], [18].
In this section, we have focused on the voltage regulation
1Vi (k, HU k) = (1 2)Hi−1 Li (k)
(29)
rather than reactive power because the latter can be realized
Following the receding horizon principle of model pre- due to the capacitive compensation of both loads and trans-
dictive control, only the first term of the optimal sequence mission lines [28].
is applied to the process. Therefore, the control action is
expressed as vi (k) = vi (k − 1) + 0i 1Vi (k, HU |k) with B. DISTRIBUTED PROPORTIONAL-INTEGRAL
0i = [1 0 . . . 0], which gives the final explicit control effort FREQUENCY REGULATION
as follows, In this section, a distributed cooperative control is elaborated
vi (k) = vi (k − 1)+0i Hi−1 NiT [Wi (Mi (k)−Yr (k))+Qi (Mi (k) for selecting an appropriate control input uωi to synchronize
individual frequency ωi to the reference ωref . In the view of
1 X
− Yj (k + 1, HP |k))] (30) nonlinear dynamics of ith DG (12), we can obtain the relation-
|Ni | ship between frequency yi2 and input ui2 using input-output
j∈Ni
feedback linearization:
ẏi2 = LFi hi2 + Lgi2 hi2 ui2 (31)
Equation (31) can be further written as the following first-
order system:
ẏi2 = ω̇i = (ω̇ni + u̇ωi ) − mPi Ṗi (32)
As ωni is a constant value, it can be obtained from (32)
u̇ωi = ω̇i + mPi Ṗi (33)
It should be noted that the secondary frequency con-
trol is desired to guarantee accurate active power sharing
among individual DG units, i.e., Pi /Pk = Pmax,i /Pmax,k , ∀i,
k ∈ n. Because the frequency droop coefficient mpi are
usually chosen in inverse proportion to the power rating,
i.e., mpi /mpk = Pmax,k /Pmax,i , the following equality must be
satisfied:
mP1 P1 = mP2 P2 = . . . = mPn Pn (34)
FIGURE 3. Block diagram of the proposed secondary voltage control
Therefore, the secondary frequency control for tracker based on IOFL-DMPC and secondary frequency control based on the
synchronization yi2 → ωref and regular synchronization distributed proportional integral control.
mPi Pi → mPk Pk , ∀i, k can be written as follows:
Z
uωi = (eωi + ePi ) (35)
X
eωi = cω aij (ωj − ωi ) + gi (ωref − ωi ) (36)
j∈Ni
[27] M. Moradzadeh, R. Boel, and L. Vandevelde, ‘‘Voltage coordination in WEI GU (M’06–SM’16) received the B.S. and
multi-area power systems via distributed model predictive control,’’ IEEE Ph.D. degrees in electrical engineering from
Trans. Power Syst., vol. 28, no. 1, pp. 513–521, Feb. 2013. Southeast University, China, in 2001 and 2006,
[28] G. Lou, W. Gu, Y. Xu, M. Cheng, and W. Liu, ‘‘Distributed MPC-based respectively.
secondary voltage control scheme for autonomous droop-controlled micro- From 2009 to 2010, he was a Visiting Scholar
grids,’’ IEEE Trans. Sustain. Energy, vol. 8, no. 2, pp. 792–804, Apr. 2017. with the Department of Electrical Engineering,
[29] Y. Wang, Z. Chen, X. Wang, Y. Tian, Y. Tan, and C. Yang, ‘‘An estimator- Arizona State University, Tempe, AZ, USA. He is
based distributed voltage-predictive control strategy for AC islanded
currently a Professor with the School of Electri-
microgrids,’’ IEEE Power Electron., vol. 30, no. 7, pp. 3934–3951,
cal Engineering, Southeast University. He is the
Jul. 2015.
[30] G. Lou, W. Gu, L. Wang, B. Xu, M. Wu, and W. Sheng, ‘‘Decentralised Director of the Institute of Distributed Generations
secondary voltage and frequency control scheme for islanded microgrid and Active Distribution Networks. His research interests include distributed
based on adaptive state estimator,’’ IET Gener., Transmiss. Distrib., vol. 11, generations and microgrids, and active distribution networks.
no. 15, pp. 3683–3693, Nov. 2017.
[31] F. A. Rengifo, L. Romeral, J. Cusidó, and J. J. Cárdenas, ‘‘New model
of a converter-based generator using electrostatic synchronous machine
concept,’’ IEEE Trans. Energy Convers., vol. 29, no. 2, pp. 344–353,
Jun. 2009. WANXING SHENG is currently a Professor and Senior Engineer with the
[32] J.-J. E. Slotine and W. Li, Applied Nonlinear Control. Upper Saddle River, China Electric Power Research Institute.
NJ, USA: Prentice-Hall, 2009. His research interests include power system planning, distribution
[33] A. Isidori, Nonlinear Control Systems. Berlin, Germany: Springer-Verlag, automation, distributed generations, and power system informatization.
1995.