Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: This paper presents a neural network predictive controller for the UPFC to improve the transient stability
Received 27 August 2010 performance of the power system. A neural network model for the power system is trained using the back-
Received in revised form 13 May 2011 propagation learning method employing the Levenberg–Marquardt algorithm for faster convergence.
Accepted 1 August 2011
This neural identifier is then utilized during predictive control of the UPFC. The damped Gauss–Newton
Available online 11 August 2011
method employing ‘backtracking’ as the line search method for step selection is used by the predictive
controller to predict the future control inputs. The 4- machine 2-area power system which is a bench-
Keywords:
mark power system is used to demonstrate the performance of the proposed controller. The system
Neural networks
Predictive control
under consideration is simulated for different transients over a range of operating conditions using Mat-
Identification lab/Simulink. The proposed neural network predictive controller exhibits superior damping performance
Power system transient stability in comparison to the conventional PI controller. The simulation results also establish convergence of the
Unified power flow controller (UPFC) minimization algorithm to an acceptable solution within single iteration.
© 2011 Elsevier B.V. All rights reserved.
1. Introduction Numerous control strategies have been reported for using the
UPFC effectively. The most commonly employed controllers for
With the increase in electrical power demand, power systems the UPFC have been of the PID (Proportional + Integral + Derivative)
are increasingly becoming complex to operate and at the same time type because of their simplicity and ease in design. However, these
less secure. Such a stressed system is continuously under threat of controllers suffer from a serious drawback in the form of deteri-
losing stability following a disturbance. Though power system sta- oration in the performance when the system is made to operate
bilizers (PSSs) have been commonly employed for damping rotor under widely ranging operating conditions and subjected to tran-
oscillations but the damping provided by the PSS has been found to sients. To overcome this drawback, controllers based on robust
be inadequate under many operating conditions. This inadequacy control techniques [2–5] and direct methods [6–8] have been used.
in damping the rotor oscillations is especially prominent in inter- A model-free approach providing non-linear control is offered by
area oscillations, thereby resulting into a requirement of some the fuzzy logic approach and is reported in the literature [9–12].
additional control. In recent times, the availability of high power Neural networks offer alternatives to the conventional linear and
semiconductor devices for power system applications have led to non-linear control methods. They have an inherent capability to
technologies such as Flexible AC Transmission Systems (FACTS) for learn and store information regarding the non-linearities of the
secure loading, power flow control and damping of power system system and provide this information whenever required. This ren-
oscillations. The FACTS devices have emerged as effective alter- ders the neural networks suitable for system identification and
natives of the additional control required for damping the power control applications [8,13–16]. Two separate continually online-
system oscillations. Of all the FACTS devices, the unified power flow trained neurocontrollers employing neuroidentifiers are reported
controller (UPFC) is the most versatile and capable of providing [14] where the parameters of both the neuroidentifiers and neuro-
stability to the system subjected to transient disturbances due to controllers were updated continuously based on the error only one
its ability to control, simultaneously or selectively, all the param- time step ahead. However, this adaptation is based on a short term
eters affecting power flow in the transmission line, i.e. voltage, goal and may not ensure a long term satisfactory performance. A
impedance and phase angle [1]. single neuron radial basis function neural network (RBFNN) using
H∞-learning method has been reported [15] to improve transient
stability performance of power system. However, this work uses a
∗ Corresponding author. Tel.: +91 9779090688. genetic optimization scheme to optimize a set of coefficients in the
E-mail address: tiwaris@nitj.ac.in (S. Tiwari). described method which may increase the computation time.
1568-4946/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2011.08.003
4582 S. Tiwari et al. / Applied Soft Computing 11 (2011) 4581–4590
Nomenclature
(ii) Use the previous calculated control input vector and the neural oscillations satisfactorily can further be enhanced whenever
identifier to predict the performance of the plant. required by using a UPFC, the work undertaken here employs
(iii) Calculate a new control input vector that minimizes the cost a UPFC, with its series and shunt converters operating in direct
function. voltage injection mode and automatic voltage control mode respec-
(iv) Repeat steps (ii) and (iii) till desired minimization is attained. tively, for this purpose. The problem for the undertaken work can
(v) Provide the first control input from the vector to the plant. be stated as given below.
(vi) Repeat steps (ii)–(v) for each time step. Problem: Use of neural network predictive control for the series
branch of the UPFC to improve the transient stability performance
of a PSS equipped multimachine power system over a wide range
4. Problem statement of operating conditions.
Fig. 4. (a) Block diagram of the test system; and (b) Simulink diagram of the test system.
The output of the tansigmoidal neuron i of layer 1 will be where, N = size of the training dataset, Poq = desired value of the out-
put of the neural network when qth input is presented, Pmq = actual
a1 (i) = tansig(n1 (i)) (10) output of the neural network when the qth input is presented,
In matrix form, the output of layer 1 (consisting of 15 tan- eq = error for the qth input.
sigmoidal neurons) is given by Since the performance index in (14) is sum of squares of non lin-
ear functions, the numerical optimization techniques for non linear
a115×1 = tansig([W]115×4 a04×1 + b115×1 ) (11) least squares can be used to minimize this objective function [26].
The Levenberg–Marquardt algorithm is said to be more efficient
where, [W ]115×4 = weight matrix between input and hidden layers as compared to other methods for convergence of the Backpropa-
and, b115×1 = bias to the hidden layer neurons. gation algorithm for training a moderate-sized feedforward neural
Similarly, the net input to neuron i of layer 2 (output layer) is network [26,27]. This method is an approximation to the Newton’s
method. The performance index V in (14) is a function of weights
15
and biases, x and can be given by
n2 (i) = w(i, j)a1 (j) + b2 (i) (12)
j=1
N
V (x) = eq2 (x) (15)
Therefore, the output of layer 2 (consisting of only one linear q=1
neuron) is given in matrix form by
To improve the performance of the neural network, x needs to
Pm = purelin(W21×15 a115×1 + b2 ) (13) be modified during training till the desired performance level is
achieved. For this, the function V(x) is minimized with respect to x
where, W 21×15 = weight vector between hidden and output layers and the update in x as per the Newton’s method [22] would be
and, b2 = bias to the output layer neuron.
−1
The neural network is trained offline using the Backpropaga- x = −[∇ 2 V (x)] ∇ V (x) (16)
tion algorithm through simulation to learn the forward dynamics
where ∇ 2 V(x)
is the Hessian matrix and ∇ V(x) is the gradient of
of the plant. The performance index for the neural network under
V(x).
consideration is
From (13), it can be shown that
1 1
N N
V= (Poq − Pmq )2 = (eq )2 (14) ∇ V (x) = J T (x)e(x) (17)
N N
q=1 q=1 ∇ V (x) = J (x)J(x) + S(x)
2 T
(18)
4586 S. Tiwari et al. / Applied Soft Computing 11 (2011) 4581–4590
and
N
S(x) = eq (x) ∇ 2 eq (x) (20)
q=1
Fig. 7. (a) Interarea mode of oscillations for different control schemes; (b) local mode of oscillations for different control schemes; (c) DC capacitor voltage characteristics
for different control schemes; and (d) voltage profile of Bus11 in the test system.
and its steady state value over some specified future time horizon. 6.2. The cost function minimization algorithm
The optimization block also minimizes the deviation in the control
action making it smooth and ensuring its steady state behavior. A cost function minimization algorithm is used with
The actual value of the active power at future time instants corre- an objective of minimizing C in (23) with respect to
sponding to the tentative control inputs are predicted by the neural [Vq (n + 1) Vq (n + 2) . . . Vq (n + Nu )]T denoted by V q where,
identifier. A cost function employing the integral square error (ISE) (n+1) is the next immediate time instant in future. This is accom-
criterion and consisting of squared deviations between the refer- plished by setting the Jacobian of (23) to zero and solving for V q .
ence and predicted active power values and the weighted square Since the cost function C is a non linear least squares problem, it is
of the change in control input over successive future time instants minimized using the damped Gauss–Newton method [28], which
is formulated as given below: is an optimization technique meant for the non linear squares
problems. This algorithm is actually the Gauss–Newton method
employing a line search. Since the damped Gauss–Newton method
N2
2 always takes descent steps that satisfy the line search criteria, it
C= (Poref (t + j) − Pm (t + j))
is always locally and usually globally convergent on almost all
j=N1
non-linear least squares problems, including large residual or very
Nu non-linear problems [28]. Using this minimization algorithm C
(Vq (t + j − 1) − Vq (t + j − 2))
2 is minimized iteratively, denoted by C(k), to determine the best
+ (23)
V q . An intermediate control input vector is generated for each
j=1
iteration of C(k) and is denoted by
Fig. 8. (a) Interarea mode of oscillations for different control schemes during prefault conditions; (b) interarea mode of oscillations for different control schemes; (c) local
mode of oscillations for different control schemes; and (d) DC capacitor voltage characteristics for different control schemes during prefault conditions.
where k is selected using Backtracking [28], Jacobian is denoted Sk = Vq (k + 1) − Vq (k) (29)
by
T The updating of Vq as per (25) continues till the desired min-
∂C(k) ∂C ∂C ∂C imization is attained. Then, the first control input, Vq (n + 1) from
= ... (26)
∂Vq ∂Vq (n + 1) ∂Vq (n + 2) ∂Vq (n + Nu ) the best determined Vq is provided as the control input Vq (n + 1) to
the plant.
and Hessian is denoted by
⎡ ⎤
∂2 C ∂2 C
··· 7. Simulation results
⎢ ∂Vq (n + 1)2 ∂Vq (n + 1)∂Vq (n + Nu ) ⎥
⎢ ⎥
∂2 C(k) ⎢ .. .. .. ⎥
=⎢ . ⎥ To investigate the performance of the proposed controller, the
∂V 2q ⎢ . . ⎥ multi-machine power system shown in Fig. 4(a) is simulated for
⎣ ∂2 C ∂2 C ⎦
··· three different operating conditions corresponding to three differ-
∂Vq (n + Nu )∂Vq (n + 1) ∂Vq (n + Nu )2 ent levels of real power flowing from Area 1 to Area 2. One of the
(27) transmission line of the double circuit transmission line between
buses 9 and 11 is subjected to a three phase short circuit fault at
6.2.1. Hessian update point A, which is very near to bus 9, for a duration of 200 ms (at
The initial Hessian approximation, H0 in this work is taken as t = 80 s). The fault is cleared by opening the circuit breakers at the
the Identity matrix, i.e. H0 = I. This initial choice of the Hessian two ends of this line, followed by automatic line reclosing while
matrix ensures that it is certainly positive definite and results in working at all the three operating conditions resulting into Case 1,
the first step being in the steepest-descent direction [28]. The sub- Case 2 and Case 3. In Case 4, the system is operated at the same oper-
sequent updates of the Hessian matrix have been made using the ating condition as in Case 3 but is subjected to two three phase short
BFGS update [28] as follows: circuit faults at short interval. These case studies are undertaken to
T
assess the performance of the proposed controller working in coor-
(Sk − (yk /Hk ))STk + Sk (Sk − (yk /Hk )) dination with the PSS as compared to the situation where the same
Hk+1 −1 = Hk −1 +
yTk Sk system is (i) equipped with PSS alone and (ii) employing a PI con-
trolled UPFC in addition to the PSS where the PI controller provides
T
(Sk − (yk /Hk )) yk Sk STk the reference for the quadrature component Vq of the series injected
− (28)
2 voltage. It is tuned manually to reduce the overshoot observed dur-
(yTk Sk )
ing transient in case of the only PSS system and also to minimize
where the steady state error.
S. Tiwari et al. / Applied Soft Computing 11 (2011) 4581–4590 4589
8. Conclusion [6] V. Azbe, R. Mihalic, Application of the direct Lyapunov method for optimum
control of a UPFC, in: Proceedings of IEEE Russia Power Tech, June 27–30, 2005,
pp. 1–6.
Neural network predictive control for the series converter of [7] V. Azbe, U. Gabrijel, D. Povh, R. Mihalic, The energy function of a general multi-
UPFC is proposed in this work. The performance of the proposed machine system with a unified power flow controller, IEEE Trans. Power Syst.
controller is then investigated in a benchmark multimachine sys- 20 (2005) 1478–1485 (August (3)).
[8] C.C. Chu, H.C. Tsai, Application of Lyapunov-based adaptive neural network
tem subjected to transients at different operating points. This UPFC damping controllers for transient stability enhancement, in: Proceedings
controller damps the interarea and local modes of oscillation in of IEEE Power and Energy Society General Meeting—Conversion and Delivery
the system very effectively in all the cases under consideration of Electrical Energy in the 21st century, 2008, pp. 1–6.
[9] S. Mishra, P.K. Dash, G. Panda, TS-fuzzy controller for UPFC in a multimachine
as compared to the conventional PI controller and performs sat-
power system, IEE Proc. Gener. Transm. Distrib. 147 (1) (2000) 15–22.
isfactorily even at those operating points where the PI controller [10] P.K. Dash, S. Mishra, Damping of multimodal power system oscillations by
fails to stabilize the system. The proposed neural network pre- FACTS devices using non-linear Takagi–Sugeno fuzzy controller, Int. J. Electr.
Power Energy Syst. 25 (2003) 481–490 (July (6)).
dictive controller hence, provides a significant improvement in
[11] P.K. Dash, S. Morris, S. Mishra, Design of a non-linear variable-gain fuzzy con-
the transient stability performance of the system over a wide troller for FACTS devices, IEEE Trans. Control. Syst. Technol. 12 (2004) 428–438
range of operating conditions. It is also established that the use (May (3)).
of UPFC supports the PSS equipped system further, and makes the [12] D.Z. Fang, X. Yang, S. Bao, Design of UPFC fuzzy-logic damping controller using
the strategy of oscillation energy descent, Int. J. Emerg. Electr. Power Syst. 6 (2)
operation of the system possible even at those operating points (2006) (Article 3).
which are otherwise not feasible in an only PSS system. The [13] P.K. Dash, S. Mishra, G. Panda, A radial basis function neural network controller
use of the off-line trained neural identifier to identify the plant for UPFC, IEEE Trans. Power Syst. 15 (2000) 1293–1299 (November (4)).
[14] G.K. Venyagamoorthy, R.P. Kalyani, Two separate continually online-trained
over a range of operating conditions, thereby saving the other- neurocontrollers for a unified power flow controller, IEEE Trans. Ind. Appl. 41
wise required computational time for continuous updating of the (2005) 906–916 (July/August (4)).
parameters of the neural identifier, along with an effective mini- [15] S. Mishra, Neural network based adaptive UPFC for improving transient stability
performance of power system, IEEE Trans, Neural Netw. 17 (2006) 461–470
mization algorithm employed by the predictive controller, imparts (March (2)).
the proposed controller a potential for use in intelligent control of [16] V.K. Chandrakar, A.G. Kothari, Comparison of RBFN based STATCOM, SSSC and
future grids. UPFC controllers for transient stability improvement, in: Proceedings of IEEE
Power Engineering Society Power Systems Conference and Exposition (PSCE)
Atlanta, USA, October 29 to November 1, 2006, pp. 784–791.
Appendix A. Appendix [17] L.Y. Dong, L. Zhang, M.L. Crow, A new control strategy for the unified power
flow controller, in: Proceedings of IEEE PES Winter Meeting, vol. 1, New York,
PSS data Sensor time Gain = 30 USA, 2002, pp. 562–566.
[18] D.W. Clarke, C. Mohtadi, P.C. Tuffs, Generalized predictive control—Part I: the
constant = 15 ms
basic algorithm, Automatica 23 (1987) 137–148 (March (2)).
UPFC ratings Series Shunt
[19] D.W. Clarke, C. Mohtadi, P.C. Tuffs, Generalized predictive control—Part II:
converter = 160 MVA converter = 160 MVA extensions and interpretations, Automatica 23 (1987) 149–160 (March (2)).
Vdcbase = 40 kV Cdc = 750 F [20] D.W. Clarke, C. Mohtadi, Properties of generalized predictive control, Automat-
Neural identifier Initial = 0.001 = 10 ica 25 (1989) 859–875 (November (6)).
Neural network predictive N1 = 1 N2 = 5 Nu = 2 [21] D. Soloway, P.J. Haley, Neural generalized predictive control—a Newton–
controller data Raphson implementation, in: Proceedings of IEEE International Symposium on
Control weighting factor, Intelligent Control, Michigan, USA, September 15–18, 1996, pp. 277–282.
= 0.01 [22] X. Lei, E.N. Lerch, D. Povh, Optimization and coordination of damping controls
for improving system dynamic performance, IEEE Trans. Power Syst. 16 (2001)
473–480 (August (3)).
References [23] M. Klein, G.J. Rogers, S. Moorty, P. Kundur, Analytical investigation of factors
influencing power system stabilizer performance, IEEE Trans. Energy Convers.
[1] N.G. Hingorani, L. Gyugyi, Understanding FACTS: Concepts and Technology of 7 (1992) 382–390 (September (3)).
Flexible AC Transmission Systems, first ed., IEEE Press, Delhi, 2000. [24] P. Kundur, Power System Stability and Control, Tata McGraw-Hill, New Delhi,
[2] M. Vilathgamuwa, X. Zhu, S.S. Choi, A robust control method to improve the 1994.
performance of a unified power flow controller, Electr. Power Syst. Res. 55 [25] M.T. Hagan, H.B. Demuth, Neural networks for control, in: Proceedings of Amer-
(2000) 103–111. ican Control Conference, CA, USA, June 2–4, 1999, pp. 1642–1656.
[3] B.C. Pal, Robust damping of interarea oscillations with unified power flow con- [26] S. Kollias, D. Anastassiou, An adaptive least squares algorithm for the effi-
troller, IEE Proc. Gen. Transm. Distrib. 149 (6) (2002) 733–738. cient training of artificial neural networks, IEEE Trans. Circuits Syst. 36 (1989)
[4] A.H.M.A. Rahim, S.A. Al-Baiyat, A robust damping controller design for a uni- 1092–1101 (August (8)).
fied power flow controller, in: Proceedings of 39th IEEE Universities Power [27] M.T. Hagan, M.B. Menhaj, Training feedforward networks with the Marquardt
Engineering Conference (UPEC), vol. 1, Bristol, UK, September 6–8, 2004, pp. algorithm, IEEE Trans. Neural Netw. 5 (1994) 989–993 (November (6)).
265–269. [28] J.E. Dennis, R.B. Schnabel, Numerical Methods for Unconstrained Optimization
[5] A.H.M.A. Rahim, J.M. Bakhashwain, S.A. Al-Baiyat, Robust damping controls for and Nonlinear Equations, SIAM, Philadelphia, 1996.
a unified power flow controller, Int. J. Emerg. Electr. Power Syst. 6 (2) (2006) [29] D.W. Marquardt, An algorithm for least-squares estimation of nonlinear param-
1–21. eters, J. Soc. Ind. Appl. Math. 11 (1963) 431–441 (June (2)).