Energy Reports 7 (2021) 1–9

Contents lists available at ScienceDirect

Energy Reports
journal homepage: www.elsevier.com/locate/egyr

Neural network-based integral sliding mode backstepping control for

virtual synchronous generators
∗
Qi Teng a , Dezhi Xu a , Weilin Yang a , , Jianlin Li b , Peng Shi c
a
School of Internet of Things Engineering, Jiangnan University, Wuxi, 214122, China
b
Inverter Technologies Engineering Research Center of Beijing (North China University of Technology), Beijing 100144, China
c
School of Electrical and Electronic Engineering, The University of Adelaide, Adelaide, SA 5005, Australia

article info a b s t r a c t

Article history: As an important component of microgrid system, the grid-connected inverter is composed of power
Received 30 August 2020 electronic devices, which makes the system lack physical inertia and exhibit low output impedance.
Received in revised form 17 October 2020 Especially in the transition process of grid connection, a large current impact will be generated,
Accepted 7 November 2020
which may make the system unstable. Therefore, in response to these problems, a nonlinear control
Available online 18 November 2020
method is proposed that enables the grid-connected inverter to switch freely between grid-connected
Keywords: mode and island mode. The linear control part introduces the mechanical equations of the rotor of
Virtual synchronous generators the synchronous motor, which has virtual inertia and damping, and at the same time derives the
Integral sliding mode control corresponding mathematical model. As regards the nonlinear control part, the nonlinear controller is
Backstepping control designed by integrating the integral sliding mode control method and backstepping control method.
RBF neural network
Meanwhile, considering the uncertainties including external disturbances, measurement errors and
modelling errors, an observer based on the radial basis function neural network (RBFNN) is proposed
to estimate the uncertainties to offset their influence. Then, the stability of the control system is proved
by Lyapunov stability criterion. Finally, the simulation results prove that the proposed control scheme
can realize the stability of the microgrid system in several operation modes and ensure the sound
dynamic performance with quick response comparing with other nonlinear control methods.
© 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction At the same time, in view of the lack of stability of the

microgrid, it is difficult for the inverter to provide the damping
For the purposes of economic development, technological in- effect required by the system. Therefore, the concept of virtual
novation, and environmental protection, renewable energy synchronous generators (VSG) is proposed based on traditional
sources have been rapidly developed, and the proportion of control (He et al., 2013). By adding the rotor motion equation,
electrical energy produced in the power grid is also increas- the inverter has a virtual inertia. The inverter output frequency
ing, making distributed generation (DG) optimization control, control simulates the virtual inertia and the primary frequency
operation and grid connection are particularly important, and modulation characteristics, making it have the external char-
the microgrid provides a method to apply distributed power acteristics close to the synchronous generator (Li et al., 2019;
generation systems, where the grid-connected inverter serves as Chen et al., 2020). In addition, in the island mode, VSG is used
a link between the microgrid and the large grid (Teodorescu et al., as a voltage source control method. VSG control is similar to
2011; Yang et al., 2016). On the plus side, distributed energy droop control, that is, a fixed frequency offset will be generated
access to the microgrid reduces transmission losses, improves when the load changes (Bidram and Davoudi, 2012). Therefore,
energy management, and improves efficiency (Tenti et al., 2011; the synchronization process needs to extract the grid frequency
Ieee, 2011). On the other hand, the inverter is composed of power and initial angle through the PLL to ensure a smooth connection
electronic devices, which has fast response capability but does between the inverter and the grid (Blaabjerg et al., 2006). A
not contain the inertia characteristics of traditional synchronous phase-locked loop (PLL) is required in the double-loop control of
motors, and cannot support the frequencies and voltages neces- voltage and current, in order to obtain the voltage and current
sary for the active distribution network with DG (Liu et al., 2015; components in the synchronous rotating reference system (Li,
Vandoorn et al., 2013). 2020).
In Zhong and Weiss (2011), the ‘‘Synchronverter’’ scheme is
∗ Corresponding author. proposed according to the electromagnetic inertia of the
E-mail address: [email protected] (W. Yang). synchronous generator. Proper control of the output voltage of

https://doi.org/10.1016/j.egyr.2020.11.032
2352-4847/© 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
Q. Teng, D. Xu, W. Yang et al. Energy Reports 7 (2021) 1–9

in different situations (Yan et al., 2007; Karimi and Feliachi,

Nomenclature 2008; Yan et al., 2010). In Zhang et al. (2010), only the opera-
δ Power angle tion of the system in grid-connected mode was studied. Based
on this synchronizer, a microgrid control topology that works
ωset Mechanical rated value
simultaneously in grid connection and island mode without re-
ω Mechanical angular velocity
construction is proposed in Ashabani and Mohamed (2012), so
ψf Virtual rotor flux island detection technology is not required. At the same time, the
θ Rotor angle proposed nonlinear controller takes into account the nonlinearity
ϕ Current angle of VSG power, frequency and angle. It avoids the problem that
ẽ Electromagnetic electromotive force the previous research mainly focused on the nonlinear control of
ĩ Three-phase current VSG output current and voltage. During the operation of VSG, the
ũ Three-phase voltage effect is not obvious (Kim and Lee, 2010; Rahmani et al., 2010).
Dp Damping coefficient In Yang et al. (2019), the controller adds a command filter on the
basis of backstepping, thereby eliminating the calculation of ex-
J Virtual inertia
pansion and input saturation nonlinear problems, and simplifying
Mf Mutual inductance
controller design process. However, the influence of uncertainties
n Droop coefficient of the system was not considered in the controller design process.
Pe Active power Backstepping Control (BSC) is a kind of system which can de-
Qe Reactive power compose a high-order nonlinear systems into subsystems lower
Qref Reactive power reference value than the original system order, and then carry out the Lyapunov
R f , Lf Impedance of the rotor winding function design for each subsystem in turn, and give the virtual
R g , Lg Grid impedance control quantity, and ‘‘step back’’ to the whole system until the
design of the entire control law is completed (Boucheta et al.,
Rs , Ls Impedance of the stator winding
2010). Integral sliding mode control (ISMC) method can over-
Te Virtual electromagnetic torque
come system uncertainty, especially for nonlinear systems, it
Tm Mechanical torque has strong robustness to external disturbances and parameter
Vn Rated voltage perturbations (Xu et al., 2019; Fei and Feng, 2020).
Vr Voltage reference value Since there are switching items in sliding mode control, chat-
BSC Backstepping Control tering will occur. These high frequencies can easily excite the
DG Distributed generation unmodeled dynamics of the control system, have a serious impact
ISMBC Integral sliding mode backstepping control on system performance, and even cause the system to oscillate
ISMC Integral sliding mode control or lose stability (Fei and Chen, 2020a). Moreover, in actual en-
gineering applications, there are still measurement errors and
PLL Phase-locked loop
disturbances from the system, which will also affect the per-
PWM Pulse width modulation
formance of the system. Therefore, these uncertain disturbances
RBFNN Radial basis function neural network need to be considered.
SG Synchronous generator According to the error model of nonlinear VSG system, this
VSG Virtual synchronous generator paper proposes an integral sliding mode backstepping controller
(ISMBC). For the uncertain disturbance of the system, an RBFNN
was designed for observation using the characteristics of infinite
approximation of the neural network. The stability of RBF neural-
the inverter can make the inverter have the function of network-based integral sliding mode backstepping controler is
adjusting the system voltage and frequency like the synchronous proved by Lyapunov function (Fei and Wang, 2019). Compared
generator. More comprehensively reflects the characteristics of with some current methods, main contributions of this paper
synchronous generators. However, when a grid connection is include the following:
required, a phase-locked loop is still required for pre-
(1) A nonlinear controller is designed for the VSG system, and
synchronization operation. However, in the transient process,
supplementary signals are added, so that the system can be
especially in weak power grids, the phase-locked loop PLL ad-
stable in islanding, grid-connected and switching between
versely affects the stability of the entire system (Zhang et al.,
the two modes. Therefore, the large signal stability of MG is
2010). Therefore, a self-synchronizing inverter is improved and
guaranteed.
designed, which is a synchronization method based on virtual
(2) Considering the uncertainties existing in the system includ-
impedance, which does not require a special synchronization
ing external disturbances, measurement errors and mod-
unit, that is, does not require the participation of PLL, but it
elling errors, which is closer to the actual system operation,
still takes nearly 1s to achieve synchronization with the grid
the RBFNN observer is designed to estimate the uncertain-
voltage (Zhong et al., 2014).
ties, which solves the uncertainty of the system and ensures
The transition between grid connection and island mode and
the tracking of the given instruction.
the sudden change of load power belong to large signal in-
(3) An integral sliding mode tracking error state is adopted,
terference in MG system. In Mohamed and Radwan (2011), a
and the sliding surface can achieve the convergence of the
droop-based power sharing controller is proposed in order to
system through the constant velocity reaching law.
obtain an active damping characteristic to reduce the high power
angle fluctuation and oscillation during the transition from grid This paper is organized as follows: In Section 2, the system
connected mode to islanding mode, seamless transition can be model of VSG is introduced. In Section 3, the system mathe-
realized in transition mode, which provides robustness for mi- matical model is built and the control method is introduced. In
crogrid operation. In addition, under large disturbances, the linear Section 4, the design process of the proposed control strategy
controller is not enough to make the MG stable under different is described. At the same time, the stability of the controller is
conditions. Therefore, it is necessary to design a nonlinear con- analyzed. In Section 5, The feasibility of the controller is proved
troller with global stability to ensure stable and robust operation by simulation. In Section 6, conclusions are presented.
2
Q. Teng, D. Xu, W. Yang et al. Energy Reports 7 (2021) 1–9

Fig. 3. The rotor mathematical model of VSG.

synchronous generator is the basic concept of a virtual syn-

chronous generator. It introduces the rotor motion equation of
the synchronous generator, including the rotor’s rotational inertia
and damping coefficient, so that the frequency and the amplitude
of the voltage vector can be directly controlled in the control,
so the output characteristics of the synchronous engine can be
simulated by the inverter.
Therefore, the frequency dynamic characteristics of VSG are
Fig. 1. Simplified structure model of SG.
expressed as:
d∆ω
J = Tm − Te − Dp (ω − ωref ) (3)
dt
where Te and Tm respectively represent the virtual electromag-
netic torque and the mechanical torque, J represents the virtual
inertia, the damping coefficient is represented by Dp , ω and ωset is
the mechanical angular velocity and its rated value, respectively.
The output active power can be obtained from the stator phase
current and electromagnetic electromotive force:
Fig. 2. Power part of VSG.
Pe = ⟨ĩ, ẽ⟩ = ωψf ⟨ĩ, sin
˜ θ⟩ (4)

where ⟨, ⟩ represents the inner product operation of the vector.

2. System model of VSG Eq. (4) is simplified as:
3 3
As shown in Fig. 1, simplified structure model of synchronous Pe = ωiψf cos(θ − ϕ ) = ωiψf cos δ (5)
2 2
generator(SG) is described. The impedance of the stator winding
is described by Rs and Ls . Similarly, the impedance of the rotor where δ = θ − ϕ , which means the power angle.
winding is expressed by Rf and Lf . θ represents the rotor an- Then through the output power and mechanical angular ve-
gle, and the transformation of mutual inductance is related to locity, the virtual electromagnetic torque is calculated as:
it (Bidram and Davoudi, 2012). Pe 3
Three-phase voltage ũ = [ua ub uc ]T can be determined Te = = iψf cos δ (6)
ω 2
as:
It is worth noting that there is no real electromagnetic torque in
dĩ the VSG system, and the virtual electromagnetic torque Te is only
ũ = −Rs ĩ − Ls + ẽ (1)
dt designed for VSG control.
]T Similarly, the reactive power Qe is expressed as:
where ĩ = [ia ib ic ]T = i cos ϕ i cos(ϕ − 23π ) i cos(ϕ − 43π ) ,
[
ia , ib and ic respectively represent the three-phase output current 3
Qe = ωiψf sin δ (7)
of the stator, ϕ is the current angle, ẽ = [ea eb ec ]T is used to 2
express as electromagnetic electromotive force. Fig. 4 is the flux control loop. Fig. 4 shows that comparing the
ẽ is generated by rotor movement: voltage reference value generated by the reactive power/voltage
droop control with the amplitude of the feedback voltage gen-
ẽ = θ̇ if Mf sin
˜ θ = ωψf sin
˜θ (2)
erates an error, and the error is sent to the low-pass filter after
where ψf is the virtual rotor flux, which can be derived from passing through the PI controller. The following low-pass filter
˜θ =
the excitation current if and the mutual inductance Mf , sin acts to imitate the weakening of the rotor flux of the synchronous
]T
sin θ sin(θ − 23π ) sin(θ − 43π ) . motor. The magnetic flux attenuation of the synchronous genera-
[
As shown in Fig. 2, it is a simple grid-connected inverter, tor is due to the DC voltage applied to the rotor excitation, and the
which takes three-phase voltage source inverter as the research inductance of the rotor winding produces a delay for the stator.
object. It consists of three inverter arms that use pulse width In addition, The function of the low pass filter is filter out the
modulation (PWM), and then its output is filtered by an LC high-frequency harmonics output by the backstepping controller,
filter in order to weaken the voltage ripple caused by switching. which helps to model the backstepping controller.
Meanwhile, the grid impedances represented by Rg and Lg are Therefore, the transfer function Gflux (s) of the low-pass filter is
considered in the grid-connected mode. expressed as:
The rotor mathematical model of VSG is shown in Fig. 3. ψf ka
Controlling the inverter to simulate the basic behavior of a real Gflux (s) = = (8)
u τa s + 1
3
Q. Teng, D. Xu, W. Yang et al. Energy Reports 7 (2021) 1–9

Fig. 4. The flux control loop.

Fig. 6. The RBFNN structure diagram.

Fig. 5. The structure of controller.

Therefore, the final mathematical model for expressing VSG in

a perturbed form is:
where u, ψf are used as input and output of low-pass filter
respectively. ∆δ̇ = ∆ω
The function of the reactive power/voltage droop control loop ∆ω̇ = α1 ∆ω + α2 ∆Te ( (16)
∆Ṫe = g1 (∆δ, i) u + g2 ∆δ, ∆ω, ∆Te , ψf , i + d
)
is as follows:
Dp
Vr = Vn − n(Qref − Qe ) (9) where α1 = − J
α = − 1J , g1 (∆δ, i) = 32 kτaa i cos ∆δ , g2 (∆δ, ∆ω,
, 2
∆Te
∆Te , ψf , i = τa + 23 i̇ψf cos ∆δ − 23 iψf ∆ω sin ∆δ , via α1 , α2 ,
)
where Qref is described as the reactive power reference value,
g1 (∆δ, i), g2 ∆δ, ∆ω, ∆Te , ψf , i make the equation more con-
( )
Vr is the voltage reference value, Vn is the rated voltage, n is
expressed as the droop coefficient. cise, the system control input is represented by u, which is the
amount of flux compensation.
The term ‘‘d’’ takes into account the modeling errors in the
3. Mathematical models and control methods VSG model modeling process, as well as uncertain factors such
as external disturbances and measurement errors. Therefore, the
The structure chart of the controller is described in Fig. 5. uncertain term d is introduced into the state space model to be
Considering there is no certain angle reference in the system characterized (Ruan et al., 2007).
model, ∆δ , ∆ω, ∆Te are used as the state variables of the state
equation, and they represent the deviation of δ , ω and Te from 4. Controller design and the stability analysis
their stable point, respectively, which shall be maintained tiny
with the controller. According to the built mathematical model, the controller
The dynamic change of ϕ is much slower than the dynamic is designed based on the idea of ISMBC. A RBFNN observer is
change of θ is assumed to obtain the dynamic change of the proposed to observe the uncertainties ‘‘d’’, which is used to offset
power angle δ . In other words, the deviation of ω from its rated the effects of the inaccurate measurements, modeling errors and
value is proportional to the rate of change of δ (Ashabani and system disturbance. The introduction of integral sliding mode
Mohamed, 2012). The dynamic change of frequency is controlled makes the interference to the model robust and the system has
by the frequency droop and the virtual rotor loops. Therefore, the good anti-interference ability. The purpose of control is to mini-
dynamic change of angle and frequency can be expressed as: mize the error between the actual value of the feedback and its
expected reference value.
∆δ̇ = ∆ω (10)
Dp 1 4.1. Design of RBF neural network observer
∆ω̇ = − ∆ω − ∆T e (11)
J J
3 3 3 Compared with the multi-layer feedforward BP network (Mir
Ṫe = i̇ψf cos δ + iψ̇f cos δ − iψf ∆ω sin δ (12) and Senroy, 2020), the RBF network is simple in structure and
2 2 2
has good generalization capabilities, effectively avoiding unnec-
After the change of Eq. (8), we can get: essary and cumbersome calculations, and in the case of arbitrary
ka 1 precision and compact set, basically all nonlinear functions can
ψ̇f = u− ψf (13) be estimated by RBFNN (Li et al., 2016). Therefore, an RBFNN
τa τa
observer is designed to observe uncertainties, and the observed
Substituting Eq. (6) and Eq. (13) into Eq. (12), we obtain: results are input to the controller as a feed forward compensation
3 ka Te 3 3 signal to cancel the influence of uncertainties, thereby reduc-
Ṫe = ui cos δ + + i̇ψf cos δ − iψf ∆ω sin δ (14) ing chattering generated in sliding mode control (Fei and Chen,
2 τa τa 2 2
2020b). To improve the stability of the virtual synchronous gen-
Considering that Eqs. (10) and (11) are perturbed forms, Eq. (14) erator control system, the RBFNN structure diagram is described
is rewritten as: in Fig. 6.
The quantity of nodes in the input layer of the RBFNN is 3,
3 ka 1 3 3
∆Ṫe = ui cos ∆δ + ∆T e + i̇ψf cos ∆δ − iψf ∆ω sin ∆δ which is selected as x = [x1 , x2 , x3 ] = [e1 , e2 , e3 ], where e1 , e2 ,
2 τa τa 2 2 e3 are tracking error of ∆δ , tracking error of ∆ω and tracking
(15) error of ∆Te . The quantity of nodes in the middle hidden layer
4
Q. Teng, D. Xu, W. Yang et al. Energy Reports 7 (2021) 1–9

is selected as 5, and the output of the hidden layer is h = [hj ]T = The Lyapunov function is selected as:
[h1 , h2 , h3 , h4 , h5 ]T , where hj represents the corresponding output 1( 2
e 1 + s1 2
)
of the jth neuron, Finally, the quantity of nodes in the output layer V2 = (30)
2
is set to 1.
Using the RBFNN to estimate the uncertainties d, the algorithm Substituting the value of ∆ω in Eq. (16) into Eq. (22), the
is as follows: derivative of the error e2 is:

x − cj 2 ė2 = ∆ω̇ − ∆ω̇d = α1 ∆ω + α2 ∆Te − ∆ω̇d

(  )
(31)
hj = exp (17)
2b2j Taking account of Eq. (31), the derivation of sliding surface s1
can be derived as:
where j is represented here as the jth node contained in the
hidden layer. x is the input of RBFNN, cj , bj are the standard ṡ1 = ė2 + β1 e2 = α1 ∆ω + α2 ∆Te − ∆ω̇d + β1 e2 (32)
deviation and mean of the Gaussian function, respectively. The
Thus the derivative of V2 is expressed as:
gradient descent algorithm is a good choice to obtain them. The
output of the Gaussian function is expressed as h = [hj ]T . V̇2 = e1 ė1 + s1 ṡ1 = −k1 e1 2 + s1 α1 ∆ω + α2 ∆Te − ∆ω̇d + β1 e2
( )
The weights of neurons from the hidden layer to the output
layer are: (33)

W = [w1 , w2 , w3 , w4 , w5 ]T (18) In order to ensure that V̇2 is negatively semi-fixed, that is,
V̇2 ≤ 0, the reaching law is adopted and the design is as follows:
The ideal output of the network is expressed as:

d = WTh + ε (19) −c1 sgn (s1 ) = ṡ1 = α1 ∆ω + α2 ∆Te − ∆ω̇d + β1 e2 (34)

where W is the ideal weight. The network proximity error is where the constant c1 > 0 is the rate at which the movement
expressed as ε . Therefore, the output expression of the designed point of the system approaches the switching plane s1 = 0. The
RBFNN disturbance observer is as follows: meaning of sgn (x) is expressed as:
d̂ = Ŵ T h 1 x>0
{
(20)
sgn (x) = 0 x=0 (35)
4.2. Integral sliding mode backstepping control strategy −1 x<0
Therefore, the optimal choice of the controller’s virtual control
First, define the tracking error of the system as follows: law ∆Te d is:
e1 = ∆δ − ∆δref (21) 1 [
∆T e d = −c1 sgn (s1 ) − α1 ∆ω + ∆ω̇d − β1 e2
]
(36)
e 2 = ∆ω − ∆ω d
(22) α2
Substituting Eq. (34) into Eq. (33), we obtain:
e3 = ∆Te − ∆Te d (23)
V̇2 = −k1 e1 2 − c1 |s1 | ≤ 0 (37)
where ∆δref is the reference value of ∆δ , ∆ωd , ∆Te d are the ref-
erence values of ∆ω and ∆Te , which are virtual control variables. Third, in order to make the torque Te stable, an integral sliding
First, in order to stabilize the work angle δ , the Lyapunov mode design is performed on the torque virtual controller. The
function is selected as: sliding surface is designed as:
1 ∫ t
V1 = e1 2 (24) s2 = e 3 + β 2 e3 dt (38)
2
0
Substituting the value in Eq. (16) into Eq. (21), the derivative
The Lyapunov function is selected as:
of the error e1 is:
1( 2
e1 + s1 2 + s2 2
)
ė1 = ∆δ̇ − ∆δ̇ref = ∆ω − ∆δ̇ref (25) V3 = (39)
2
Therefore, in light of Eq. (25), V̇1 can be derived as: Substituting the value of ∆Te in Eq. (16) into Eq. (23), the
derivative of the error e3 is:
V̇1 = e1 ė1 = e1 ∆ω − ∆δ̇ref
( )
(26)
ė3 = ∆Ṫe − ∆Ṫed = g1 (∆δ, i) u + g2 ∆δ, ∆ω, ∆Te , ψf , i + d − ∆Ṫed
( )
In order to satisfy the stability of the system, it is necessary
to make V̇1 ≤ 0. Therefore, the optimal choice of the controller’s (40)
virtual control law ∆ωd is:
Based on Eq. (40), the derivation of sliding surface s2 can be
∆ωd = −k1 e1 + ∆δ̇ref (27) obtained as:

where k1 > 0 is a positive constant. ṡ2 = ė3 +β2 e3 = g1 (∆δ, i) u + g2 ∆δ, ∆ω, ∆Te , ψf , i + d − ∆Ṫed +β2 e3
( )
Substituting Eq. (27) into Eq. (26), we obtain:
(41)
V̇1 = −k1 e1 2 ≤ 0 (28)
Thus the derivative of V3 is expressed as:
Second, in order to stabilize the angular velocity ω, design of
V̇3 = e1 ė1 + s1 ṡ1 + s2 ṡ2 = −k1 e1 2 − c1 |s1 |
the angular velocity virtual controller using integral sliding mode. [ ]
The sliding surface is designed as:
+ s2 g1 (∆δ, i) u + g2 ∆δ, ∆ω, ∆Te , ψf , i + d − ∆Ṫed + β2 e3
( )
∫ t
s1 = e 2 + β 1 e2 dt (29) (42)
0
5
Q. Teng, D. Xu, W. Yang et al. Energy Reports 7 (2021) 1–9

Similarly, in order to ensure that V̇3 is negatively semi-fixed, Table 1

that is, V̇3 ≤ 0, the sliding mode reaching law is used to fulfill the Basic parameters of VSG.

design as follows. Parameters value Parameters value

Ug 510 (V) τa 0.01
−c2 sgn (s2 ) = ṡ2 = g1 (∆δ, i) u+g2 ∆δ, ∆ω, ∆Te , ψf , i +d−∆Ṫed +β2 e3
( )
Udc 1100 (V) ka 0.0005
Pm 10 (kW) kp 50
(43) ω 314 (rad/s) ki 200
n 0.002
where the constant c2 > 0.
According to Eq. (43), the control law u is designed as:
Table 2
[ ]
1 Parameters of controller and RBF observer.
−c2 sgn (s2 ) − g2 ∆δ, ∆ω, ∆Te , ψf , i + ∆Ṫed − β2 e3
( )
u=
g1 (∆δ, i) Parameters value Parameters value
k1 25
(44) β1 0.1 β2 0.1
c1 0.15 c2 2
Substituting Eq. (43) into Eq. (42), we obtain:
γ 0.002
V̇3 = −k1 e1 2 − c1 |s1 | − c2 |s2 | + ds1 (45)
According to Lyapunov stability criterion, the control system
can be maintained stable if V̇3 ≤ 0. Considering the adverse effect The above equation shows that e1 , s1 and s2 are uniformly
of uncertainties on the stability of the system, it is necessary bounded (Djeghali and Ghanes, 2013), and −k1 e1 2 , −c1 |s1 | and
to adjust c1 , c2 as large as possible, which may aggravate the −c2 |s2 | are negative. According to Eq. (51), by using the proposed
chattering of the system. Thus, the disturbance observer is put control method, the entire system will gradually stabilize. The
forward to estimate the existing uncertainties. working flow of the designed controller is shown in Fig. 7.
For the above situation, the characteristics of the neural net-
work’s infinite approximation to the unknown model part are 5. Simulation
used to select the RBFNN observer to observe the uncertainties.
After introducing the RBFNN observer, a new Lyapunov function The corresponding model is established in the MATLAB/
is designed, and The neural network adaptive law is obtained Simulink simulation platform to prove the feasibility of the pro-
during its derivation. posed control strategy. In order to verify the effectiveness of
The new Lyapunov function is selected as: the proposed control method, the model was built by MAT-
1( 2 ) W̃ T W̃ LAB/Simulink simulation platform for verification. The main anal-
V = e 1 + s1 2 + s2 2 + (46) ysis is the active power output by the VSG and the system
2 2γ
frequency change in the island mode, the grid connection, and the
where γ > 0, W̃ = Ŵ − W , the derivative of V is as follows: transition to the island mode. Meanwhile, another two control
methods, namely, BSC and ISMBC are applied to the VSG as
˙
W̃ T Ŵ comparison.
V̇ =e1 ė1 + s1 ṡ1 + s2 ṡ2 +
γ The selection of controller parameters and observer param-
[ eters is very important. The control effect is closely related to
2
= − k1 e1 − c1 |s1 | + s2 g1 (∆δ, i) u the selection of appropriate parameters, the selection of specific
simulation parameters is shown in Tables 1 and 2.
˙
W̃ T Ŵ The simulation of the system containing 10 kW load is selected
]
+ g2 ∆δ, ∆ω, ∆Te , ψf , i + d − ∆Ṫed + β2 e3
( )
+ to prove the feasibility of the controller, and the performance
γ
parameters at the moment when the VSG is connected and dis-
(47)
connected from the power grid are observed. so the damping
Design the new control law u as: coefficients and virtual inertia are selected as Dp = 20.26 and
[ ] J = 0.4 kg m2 .
1 The VSG system first operates in the island mode, and the VSG
−c2 sgn (s2 )−g2 ∆δ, ∆ω, ∆Te , ψf , i −d̂+∆Ṫed −β2 e3
( )
u=
g1 (∆δ, i) system supplies 10 kW of active power to the 10 kW local load.
(48) Then, when t = 0.3 s, since the grid is operating normally and
there is no fault, the system will be connected to the grid. Finally,
Let f˜ = f − fˆ , substituting Eq. (48) into Eq. (47), we obtain: when t = 0.6 s, due to the failure of grid operation, the system
˙ needs to be operated off-grid and enter island mode operation.
[ ] W̃ T Ŵ
V̇ = −k1 e1 2 − c1 |s1 | + s2 −c2 sgn (s2 ) − d̃ + Fig. 8 is a comparison of the active power changes of the three
γ methods. It can be seen that in the island mode, the active power
T ˙ can quickly and smoothly reach the required active power. In
W̃ Ŵ
= −k1 e1 2 − c1 |s1 | − c2 |s2 | − W̃ T hs2 + the active output power switching process, it can be seen that
γ the maximum oscillation range of the proposed controller is 9.86
˙
( )
Ŵ ∼ 9.965 kW, the maximum oscillation range of BSC is 9.82 ∼
= −k1 e1 2 − c1 |s1 | − c2 |s2 | − W̃ T hs2 − (49) 9.96 kW, and the maximum oscillation range of ISMBC is 9.855
γ
∼ 9.97 kW. Compared with BSC, ISMBC adds the sliding mode
Then, the weight adaptive law is selected as: control, which expands the application range of the BSC method.
Therefore, it makes the interference to the system robust. This can
˙
Ŵ = γ hs2 (50) be seen from this comparison result, when ISMBC is connected to
the grid at 0.3 s, the oscillation amplitude is smaller. The proposed
Substituting Eq. (50) into Eq. (49) can be expressed as:
control method uses RBFNN to observe the uncertainties d, and
V̇ = −k1 e1 2 − c1 |s1 | − c2 |s2 | ≤ 0 (51) compensates for the disturbance in sliding mode control, which
6
Q. Teng, D. Xu, W. Yang et al. Energy Reports 7 (2021) 1–9

Fig. 7. Structure of integral sliding mode backstepping controller based on neural network.

Fig. 8. The situation of variation of active power. Fig. 10. The situation of variation of three-phase voltage.

Fig. 9. The situation of variation of frequency.

Fig. 11. The situation of variation of three-phase current.

can effectively reduce the switching gain, thereby effectively re-

ducing chattering, the accuracy of the controller is improved, and
the data in the simulation It can also be seen that the oscillation
range of the proposed control method during grid connection is
the smallest and the transition is smoother.
Fig. 9 shows the frequency change. From the comparison of the
frequency changes of several methods, it can be seen that when
the t = 0.3 s system is connected to the grid, the transition of the
controller using ISMBC is obviously smaller than the frequency
oscillation of BSC alone, and the process is smoother, and After
being connected to the grid, it tends to 50 Hz faster, which reflects
the robustness of sliding mode control. Compared with ISMBC,
the proposed control method has a smoother dynamic process Fig. 12. The situation of variation of three-phase current without nonlinear
and almost no frequency oscillation, which reflects the observa- controller.
tion of uncertainties by the RBFNN observer, and compensates for
disturbances in sliding mode control, thereby effectively reducing
chattering, Improve the accuracy of the controller. At t=0.6 s, in when the system runs from island to the grid and transitions
the process of switching to island mode, the system frequency to the island again, the voltage and current waveforms can ba-
change under the proposed controller is more stable. sically remain unchanged. Among them, in the process of grid
Fig. 10 and Fig. 11 shows the output situation of three-phase connection, the effective value of the line voltage of the three-
voltage and three-phase current, respectively. It is not difficult to phase voltage can be maintained at about 510 V, which can be
observe that under the control of the proposed control strategy, consistent with the voltage of the grid.

7
Q. Teng, D. Xu, W. Yang et al. Energy Reports 7 (2021) 1–9

Fig. 13. The tracking error e1 , e2 , e3 .

Fig. 14. The sliding surface s1 , s2 .

inconsistency between the output voltage and the grid voltage

increases, the inrush current will further increase. Therefore, in
order to ensure the stable operation of the system, it is necessary
to avoid the occurrence of inrush current.
Fig. 13 shows the comparison of the tracking error e1 , e2 ,
e3 of the virtual control variables of the three methods. It can
be observed by comparison that the proposed control strategy
can obtain better error control. In particular, e3 can follow the
adjustment of the system more accurately and reach a stable
state.
In Fig. 14, The sliding mode surfaces s1 and s2 of ISMBC
and RBFNN-observer-based ISMBC are compared respectively. It
Fig. 15. The output of disturbance observer. can be seen that the chattering of the sliding mode surface
combined with ISMBC of the RBFNN disturbance observer is re-
duced. Especially, on the sliding mode surface s2 , it can be seen
that the sliding mode chattering under RBFNN-observer-based
In Fig. 12, the waveforms show the changes in the three- ISMBC is less than that under ISMBC. This also shows that the
phase current without a nonlinear controller. It can be seen RBFNN observer observes the uncertainties d, and compensates
from the figure that when the grid is connected in 0.3 s, the the disturbance in sliding mode control, which effectively reduces
three-phase current has a current impact and maintains 0.17 s chattering.
before it stabilizes under the VSG control strategy. Through the Fig. 15 is the system disturbance observed by the neural net-
comparison of Figs. 12 and 11, it is not difficult to see that the work observer. It can be seen that the observer can better observe
current waveform is seriously unstable when connected to the the disturbance received by the system during operation, thereby
grid without the proposed nonlinear controller. Moreover, as the maintaining the stability of the system.

8
Q. Teng, D. Xu, W. Yang et al. Energy Reports 7 (2021) 1–9

The simulation results largely prove that the controller built Fei, J., Chen, Y., 2020b. Dynamic terminal sliding-mode control for single-phase
based on the proposed control strategy can meet the require- active power filter using new feedback recurrent neural network. IEEE Trans.
Power Electron. 35 (9), 9904–9922.
ments of smooth switching during the transition process. In par-
Fei, J., Feng, Z., 2020. Fractional-order finite-time super-twisting sliding mode
ticular, before the grid connection, the system does not need to control of micro gyroscope based on double-loop fuzzy neural network. IEEE
be pre-synchronized, and the stability is improved. Trans. Syst. Man Cybern.: Syst. PP (99), 1–15.
Fei, J., Wang, H., 2019. Experimental investigation of recurrent neural network
6. Conclusions fractional-order sliding mode control of active power filter. Circuits Syst. II:
Express Briefs IEEE Trans. PP (99), 1.
He, J., Li, Y.W., Bosnjak, D., Harris, B., 2013. Investigation and active damping of
In this paper, the problem of large voltage angle and frequency multiple resonances in a parallel-inverter-based microgrid. IEEE Trans. Power
deviation between traditional VSG and grid is solved by the pro- Electron. 28 (1), 234–246.
posed RBFNN-observer-based ISMBC strategy so that the stability Ieee, B.E., 2011. IEEE Guide for Design, Operation, and Integration of Distributed
Resource Island Systems with Electric Power Systems.
and working performance of the system are guaranteed. Con-
Karimi, A., Feliachi, A., 2008. Decentralized adaptive backstepping control of
cretely, the uncertainties including external disturbances, mea- electric power systems. Electr. Power Syst. Res. 78 (3), 484–493.
surable errors and modelling errors are considered and estimated Kim, D., Lee, D., 2010. Feedback linearization control of three-phase UPS inverter
via the proposed RBFNN based observer with high accuracy and systems. IEEE Trans. Ind. Electron. 57 (3), 963–968.
opportune response, which further helps release the influence Li, S., 2020. Control strategies of grid-connection and operation based on active
disturbance rejection control for virtual synchronous generator. Int. J. Electr.
of uncertainties in the controller design process. In terms of Power Energy Syst. 123.
the system’s performance, the power oscillation is reduced, the Li, P., Hu, W., Xu, X., 2019. A frequency control strategy of electric vehicles in
frequency transition is smoother, and the dynamic performance microgrid using virtual synchronous generator control. Energy 189.
and control accuracy are improved under the designed controller Li, Y., Li, T., Tong, S., 2016. Adaptive neural networks output feedback dynamic
surface control design for MIMO pure-feedback nonlinear systems with
when comparing other nonlinear and robust control methods,
hysteresis. Neurocomputing 58–68.
which verifies the good control performance of the proposed Liu, N., Aljankawey, A., Diduch, C., Chang, L., Su, J., 2015. Passive Islanding
control strategy on the microgrid. detection approach based on tracking the frequency-dependent impedance
In the future work, we will build an experiment platform of change. IEEE Trans. Power Deliv. 30 (6), 2570–2580.
small-scale microgrid to verify the practical effectiveness of our Mir, A.S., Senroy, N., 2020. Self-tuning neural predictive control scheme for
ultrabattery to emulate a virtual synchronous machine in autonomous power
proposed controller. systems. IEEE Trans. Neural Netw. 31 (1), 136–147.
Mohamed, A.R.I., Radwan, A.A., 2011. Hierarchical control system for robust
Declaration of competing interest microgrid operation and seamless mode transfer in active distribution
systems. IEEE Trans. Smart Grid 2 (2), 352–362.
Rahmani, S., Mendalek, N., Alhaddad, K., 2010. Experimental design of a nonlin-
The authors declare that they have no known competing finan-
ear control technique for three-phase shunt active power filter. IEEE Trans.
cial interests or personal relationships that could have appeared Ind. Electron. 57 (10), 3364–3375.
to influence the work reported in this paper. Ruan, S., Li, G., Jiao, X., Sun, Y., Lie, T.T., 2007. Adaptive control design for VSC-
HVDC systems based on backstepping method. Electr. Power Syst. Res. 77
Acknowledgment (5), 559–565.
Tenti, P., Paredes, H.K., Mattavelli, P., 2011. Conservative power theory, a frame-
work to approach control and accountability issues in smart microgrids. IEEE
This work was partially supported by the National Natural Sci- Trans. Power Electron. 26 (3), 664–673.
ence Foundation of China (61973140, 61903158), National First- Teodorescu, R., Liserre, M., Rodriguez, P., 2011. Grid Converters for Photovoltaic
Class Discipline Program of Food Science and Technology (JUF- and Wind Power Systems. Wiley.
Vandoorn, T.L., De Kooning, J.D., Meersman, B., Guerrero, J.M., Vandevelde, L.,
STR20180205) and the National Science Foundation of Jiangsu
2013. Voltage-based control of a smart transformer in a microgrid. IEEE
Province of China (BK20180595). Trans. Ind. Electron. 60 (4), 1291–1305.
Xu, D., Wang, G., Yan, W., Yan, X., 2019. A novel adaptive command-filtered
References backstepping sliding mode control for PV grid-connected system with energy
storage. Sol. Energy 222–230.
Ashabani, S.M., Mohamed, Y.A., 2012. A flexible control strategy for grid- Yan, R., Dong, Z.Y., Saha, T.K., Majumder, R., 2007. Power system transient
connected and Islanded microgrids with enhanced stability using nonlinear stability enhancement with an adaptive control scheme using backstepping
microgrid stabilizer. IEEE Trans. Smart Grid 3 (3), 1291–1301. design. In: IEEE Power Engineering Society General Meeting. IEEE, pp. 1–8.
Bidram, A., Davoudi, A., 2012. Hierarchical structure of microgrids control Yan, R., Dong, Z.Y., Saha, T.K., Majumder, R., 2010. Brief paper: A power
system. IEEE Trans. Smart Grid 3 (4), 1963–1976. system nonlinear adaptive decentralized controller design. Automatica 46
Blaabjerg, F., Teodorescu, R., Liserre, M., Timbus, A.V., 2006. Overview of control (2), 330–336.
and grid synchronization for distributed power generation systems. IEEE Yang, F., Sun, Q., Han, Q.L., Wu, Z., 2016. Cooperative model predictive control
Trans. Ind. Electron. 53 (5), 1398–1409. for distributed photovoltaic power generation systems. IEEE J. Emerg. Sel.
Boucheta, A., Bousserhane, I.K., Hazzab, A., Mazari, B., Fellah, M.K., 2010. Adaptive Top. Power Electron. 4 (2), 414–420.
backstepping controller for linear induction motor position control. COMPEL- Yang, C., Yang, F., Xu, D., Huang, X., Zhang, D., 2019. Adaptive command-filtered
Int. J. Comput. Math. Electr. Electron. Eng. 29 (3), 789–810. backstepping control for virtual synchronous generators. Energies 12 (14).
Chen, J., Liu, M., Milano, F., 2020. 100% Converter-Interfaced generation using Zhang, L., Harnefors, L., Nee, H., 2010. Power-synchronization control of
virtual synchronous generator control: A case study based on the irish grid-connected voltage-source converters. IEEE Trans. Power Syst. 25 (2),
system. Electr. Power Syst. Res. 187, 106475. 809–820.
Djeghali, N., Ghanes, M., 2013. Sensorless fault tolerant control for induction Zhong, Q., Nguyen, P., Ma, Z., Sheng, W., 2014. Self-synchronized synchronvert-
motors. Int. J. Control Automat. Syst. 11 (3), 563–576. ers: Inverters without a dedicated synchronization unit. IEEE Trans. Power
Fei, J., Chen, Y., 2020a. Fuzzy double hidden layer recurrent neural terminal Electron. 29 (2), 617–630.
sliding mode control of single-phase active power filter. IEEE Trans. Fuzzy Zhong, Q., Weiss, G., 2011. Synchronverters: Inverters that mimic synchronous
Syst. PP (99), 1. generators. IEEE Trans. Ind. Electron. 58 (4), 1259–1267.