Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

IET Generation, Transmission & Distribution

Research Article

Extreme learning machine-based alleviation ISSN 1751-8687


Received on 17th April 2019

for overloaded power system


Revised 1st August 2019
Accepted on 18th September 2019
E-First on 30th October 2019
doi: 10.1049/iet-gtd.2019.0531
www.ietdl.org

Imen Labed1 , Djamel Labed1


1Laboratory of Electrical Engineering, Department of Electrical Engineering, University of Constantine1, Constantine, Algeria
E-mail: imen.labed@lec-umc.org

Abstract: The present study proposes a corrective method of electric system overload provided that the wind farm is integrated
into the distribution system, taking into account the congestion cost. The authors attempted to mitigate the overload and to
monitor flow over transmission lines. Unified power flow controller device was the first suggestion utilised to solve this problem,
then due to its extremely fast training and the excellent generalisation performance, extreme learning machine algorithm is
employed. The fundamental point is the transmission line alleviation. In addition, other targets are realised: the load shedding
avoidance, minimisation of losses and congestion cost. This study is also designed to utilise PowerWorld Simulator and
MATLAB software to demonstrate methods for relieving transmission overloads. The accuracy of the proposed approach has
been tested for Algerian (Adrar) 22-bus system. Obtained results showed an improvement in power system behaviour.
Simulation results are exposed, discussed and compared at the end of this study.

1 Introduction problems, and the results showed that ELM has an outstanding
performance in relative terms [7].
Inspired by biology, artificial neural networks (NNs) are a common ELM does not need tuning of all parameters; the only free
technique in the field of machine learning. Used in classification, parameters need to be learnt are the connections (weights and bias)
pattern recognition, time-series prediction and other approximation between the hidden and the output layers, and this is what makes it
functions or regression. Many variations in the type of the network extremely fast to implement [6]. Huang combined the concept of
are in use, the activation functions and in the learning algorithm. local receptive fields (LRFs) with ELM and proposed an LRF-
Backpropagation (BP) is the most known learning algorithm for based ELM (LRF-ELM) to learn local correlations of input images
a feed-forward NN. The BP is essentially a first-order gradient [8].
method for parameter optimisation. However, it is clear that the This study endeavoured to solve economically an overloaded
gradient descent-based learning methods are generally very slow, power system due to renewable energy penetration. The
and suffering from some drawbacks [1]: contributions made in the present paper are two-fold. First, using a
device that belongs to the facts system, which is the unified power
• A low learning rate can lead to slow convergence. flow controller (UPFC) to monitor flows over overloaded
• A significant learning rate can lead to divergence. transmission lines. Second, for further optimisation in time and
• Overtraining the network reduces the generalisation accuracy, and based on the obtained results in the first section
performance. utilising UPFC we introduced ELM-based alleviation algorithm
• The algorithm can be very time-consuming, especially for the ELM as an innovative novel for power system mitigation under
large dataset. different several conditions. The substantial reduction in
computational time allows estimations closer to real time, relevant
Boser et al. introduced support vector machines (SVMs) that to short-term operational assessments.
analyse data efficiently [2, 3]; primarily SVMs were limited to The present paper is organised as follows: Section 2 talks about
classification problems only, and later on extended to solve power flow calculation and Newton–Raphson method. In Section
regression problems and named as support vector regression 3, UPFC is formulated. We explained the ELM algorithm in
(SVR). SVR creates separation surface using polynomial, Section 4. Section 5 presents the simulation results and discussion,
sigmoidal and Radial Basis Function (RBF) functions. The and finally the conclusion.
regression processes of these machines use only the data patterns
closest to the separation surface instead of all data patterns [4, 5]. 2 Load flow analysis – Newton–Raphson
In 2004, a new alternative to BP for learning feed-forward NNs
had been proposed by Huang; it is extreme learning machine The basic idea in power flow analysis is to determine the phase and
(ELM). This algorithm is easy to implement and does not suffer the voltage magnitude at any grid bus, according to the system of
from the drawbacks above. ELM is the idea of a random projection (1) which is composed of the real and the reactive powers [9, 10]
followed by linear regression. It has been proposed for training a
single hidden layer feed-forward neural networks (SLFNs). In n

ELM, the hidden nodes are randomly initiated and then fixed Pi = V i ∑ V j(Gi j cos δi j + Bi j sin δi j)
without iteratively tuning [6]. j=1
(1)
Huang demonstrated that ELM provides a much faster learning n

speed and other attractive properties compared with the alternative. Qi = V i ∑ V j(Gi j sin δi j Bi j cos δi j)
Different from traditional learning algorithms, ELM not only tends j=1

to reach the smallest training error but also the smallest norm of
weights and the excellent generalisation performance. The where Pi, Qi are real and reactive powers at PQ buses and Gij, Bij
performance of the ELM was compared with other algorithms are conductance and the susceptance of the shunt admittance of the
including SVMs and conventional BP, on some benchmark line [9, 10].

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070 5058
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
It is clear that in the non-linear power equations mentioned Transmitted uncompensated real (Po) and reactive (Qo) power
above, there is no equality between the number of equations and flows are given by expressions below [16, 18, 19]:
the variables number; therefore, resort to some other methods that
are applicable for solving these equations [11–13]. The classical V rV s
resolution is not able here to solve the problem; this may explain P0(δ) = sin(δ) (6)
x
the reason for which numerical methods are used as well as the
Newton–Raphson approach [13, 14]. On the basis of the Jacobian V r(V s cos(δ) − V r)
matrix, whose elements are partial derivative values of either P or Q0(δ) = (7)
x
Q for either V or δ, we can write the equation as follows: (see (2)) .
ΔPi, ΔQi are real and reactive power mismatch at bus bar (i), Δδij is The real and reactive powers with UPFC are
the voltage angle between bus bar (i) and bus bar (j) and Δδi is the
voltage angle mismatch at bus bar (i). ΔVV r sin(φ)
To get the variation of the voltage magnitude and the phase, we P(δ, φ) = P0(δ) + (8)
x
have to inverse the Jacobian matrix as it is presented in the
equation below: V r(V s cos(δ) + ΔV cos(φ) − V r)
Q(δ, φ) = (9)
(k )
x
[∂δ ] [∂P ](k )

¯ (k)
= [J]−1 (3) The magnitude and the phase angle of the injected voltage in series
∂V [∂Q(k)]
with the transmission line to provide the required real and reactive
powers are given by the following equations [20, 21]:
• By consequence, the new values of the mentioned targets [15,
16] X
ΔV =
Vr
(Q(δ, φ) − Q0)2 + (P(δ, φ) − P0)2 (10)
V ik + 1 = V ik + ΔV ik
(4)
δik + 1
= + Δδik δik P(δ, φ) − P0
φ = atan (11)
Q(δ, φ) − Q0
• The iterative system stops when the convergence criterion is
satisfied [9]
where Vr and Vs are voltage magnitudes at receiving and sending
(k ) end lines, X is the transmission line impedance, Po, Qo are
ΔPi <ε
(5) uncompensated real and reactive line power flows, and ΔV is the
ΔQi (k )
<ε magnitude of the injected voltage in series with the line [22, 23].

Once Vik + 1 and δik + 1 are known, the real and reactive powers 4 Extreme learning machine
can always be calculated using the equation system number (1) [15, Huang et al. proposed an ELM algorithm, as a new learning
16]. scheme working for single hidden layer feed-forward neural
networks (SLFNs) [6, 24].
3 Unified power flow controller ELM is entirely different from traditional iterative learning
algorithms as it randomly selects the input weights (W) and biases
The UPFC is the most gifted version of flexible AC transmission (b) for hidden nodes and analytically calculates the output weights
system devices. It serves to control all the three parameters (β) by finding the least-squared solution.
(voltage, impedance and phase angle) simultaneously [16]. The In doing so, it is proven that the training error can still be
UPFC is a combination of a static compensator (STATCOM) shunt minimised with even better generalisation performance [6, 24].
controller STATCOM and a static synchronous series compensator Notably, the learning speed of ELM can be thousands of times
controller (SSSC). Coupled with a common DC capacitor allowing faster than traditional feed-forward network learning algorithms
the bidirectional transit of the real and the reactive powers between such as the BP algorithm. Compared to conventional learning
the series SSSC output and the shunt STATCOM output [16, 17]. algorithms, ELM not only tends to reach the smallest training error
The main reasons for using UPFC are as follows: but also to obtain the smallest norm of weights [25, 26].
For N arbitrary distinct samples (xi, ti), where
1. Passing reactive power flow bidirectionally.
2. Maintaining DC voltage regulation.
xi = (xi1 , xi2 , …, xin )T ∈ Rn (12)
3. Workability in a wide range of operating conditions.

∂P2 ∂P2 ∂P2 ∂P2 ∂P2 ∂P2


⋯ ⋯
∂δ2 ∂δ3 ∂δn ∂ V¯ m + 1 ∂ V¯ m + 2 ∂ V¯ n
∂P3 ∂P3 ∂P3 ∂P3 ∂P3 ∂P3
⋯ ⋯
ΔP2 ∂δ2 ∂δ3 ∂δn ∂ V¯ m + 1 ∂ V¯ m + 2 ∂ V¯ n Δδ2
ΔP3 ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ Δδ3
⋮ ∂Pn ∂Pn ∂Pn ∂Pn ∂Pn ∂Pn ⋮
⋯ ⋯
ΔPn ∂δ2 ∂δ3 ∂δn ∂ V¯ m + 1 ∂ V¯ m + 2 ∂ V¯ n Δδn
= (2)
ΔQm + 1 ∂Qm + 1 ∂Qm + 1 ∂Qm + 1 ∂Qm + 1 ∂Qm + 1 ∂Qm + 1 Δ V¯ m + 1
⋯ ⋯
ΔQm + 2 ∂δ2 ∂δ3 ∂δn ∂ V¯ m + 1 ∂ V¯ m + 2 ∂ V¯ n Δ V¯ m + 2
⋮ ∂Qm + 2 ∂Qm + 2 ∂Qm + 2 ∂Qm + 2 ∂Qm + 2 ∂Qm + 2 ⋮
⋯ ⋯
ΔQn ∂δ2 ∂δ3 ∂δn ∂ V¯ m + 1 ∂ V¯ m + 2 ∂ V¯ n Δ V¯ n
⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮ ⋮
∂Qn ∂Qn ∂Qn ∂Qn ∂Qn ∂Qn
⋯ ⋯
∂δ2 ∂δ3 ∂δn ∂ V¯ m + 1 ∂ V¯ m + 2 V¯ n
IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070 5059
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ti = (ti1 , ti2 , …, tim) TRm (13) H† = (HTH)−1H (20)

Standard SLFNs with N ≈ nodes and activation function g(x) are This solution is accurate as long as the square matrix (HTH) is
mathematically modelled as invertible. However, it is singular in some applications. In SLFN, it
is singular in most of the cases since there is a tendency to select
N≈ m ≪ n. In ELM. Huang et al. have solved this problem using
∑ βig(wi . x j + bi) = o j Moore–Penrose pseudo inverse of H matrix to overcome this
i=1 (14) problem [27, 30, 31].
j = 1, … , N
5 Simulation and results analysis
where wi = (wi1, wi2, …, win) T is the weight vector connection of 5.1 Grid and problem description
the ith hidden node and input nodes, βi = (βi1, βi2, …, βim)T is the
weight vector connecting the ith hidden node and output nodes, bi Keeping in mind the end goal to check the execution of the above
thoughts referred to earlier, we considered in the simulation section
is the threshold of the ith hidden node and wi and x j denote the an Algerian (Adrar) 22-bus system using power word simulator
inner products of wi and x j. software.
That standard SLFN with N ≈ hidden nodes and g(x) activation The software contains a highly effective power flow analysis
function can approximate these N samples with zero package capable of efficiently solving systems with up to 100,000
buses. Simulation section is divided into four phases:
N≈
∑ ∥ oj − t j ∥ = 0 (15) (i) First step: We performed load flow analysis using Newton–
j=1 Raphson method for Algerian (Adrar) 22-bus system, under normal
conditions.
which means there exist βi, wi and bi such that (ii) Second step: We inserted a wind turbine (20 MW) into the
system; we studied the grid behaviour, identified the weak
N≈ N≈ elements and determined the number of the overloaded lines.
∑ βigi(x j) = ∑ βig(wi . x j + bi) = t j (16) (iii) Third step: We inserted UPFC in the violated lines to monitor
i=1 i=1 flows, to eliminate the overload, to minimise losses and so the
congestion cost.
Equation (16) can be written compactly as
(iv) Fourth step: On the basis of the third step results, we
introduced ELM-based alleviation algorithm to manage the
Hβ = T (17)
congestion and optimise the simulation time, losses and congestion
cost. ELM built predictor is then applied in power system for
where
different congestion conditions to verify its generalisation
performance in real time. ELM-based approach is performed
utilising MATLAB environment.
H(w1…wN , x1…xN , b1…bN ≈)
The Algerian (Adrar) 22-bus system exposed in Fig. 1 is a hybrid
g(w1 . x1 + b1) … g(wN ≈ . x1 + bN ≈) (18) grid, where photovoltaic energy reaches 46 MW, and the wind farm
has a capacity of 20 MW which is disconnected in the beginning.
= ⋮ … ⋮ Bus number 3 is considered as a slack bus. The two generators
g(w1 . xN + b1) … g(wN ≈ . xN + bN ≈) N ∗ N≈
connected at bus 11 are on automatic generation control (AGC)
mode to allow the power output of generators changing
T T automatically. Generators connected at bus 11 and also at bus 6 are
β1 t1
on automatic voltage regulator (AVR) mode to regulate voltage
β= ⋮ and T = ⋮ buses automatically. Load flow is solved based on Newton–
βN ≈ N ≈ ∗ m tN N ∗m
Raphson method, which is an available option in PowerWorld
Simulator software.
Thus, the ELM algorithm can be summarised as follows: Table 1 exposes the load demand variation during winter days.
Voltage profile and losses for the steady state (before renewable
Step 1: Randomly assign input weight W i and bias bi energy integration) are exposed in Figs. 2a and b.

(i = 1, 2, …, N) . 5.2 Wind farm insertion


Step 2: Calculate the hidden layer output matrix H. In this stage, 20 MW of wind source is connected to the studied
Step 3: Calculate the output weight β by β = H†T. electric system at bus 15. It is observed in Fig. 3 the violations
occurring in lines (7–21), (21–10), (10–22), (22–18) and (18–5)
If the hidden layer output matrix is non-square (that means due to the wind energy insertion. It is also noted the variation in
N ≈ < N), the determinant of H (det H = 0) and (rank H < N ≈), in load flow, the generators output and the voltage profile.
linear algebra matrix, is assessed as a singular matrix that cannot The proposed case (while integrating the wind farm) is studied
be directly inverted [27–29] during winter days, where the load average variation is within: the
evening peak 145 MW and the night dip, which reaches 81 MW.
β = (HTH)−1HTT (19) The two critical scenarios that could happen to the net and must be
studied while integrating a wind farm are either:
The inverse of H cannot be determined if H is not a full-rank
• A high wind energy penetration rate with maximum demand.
matrix. In this case, the generalised inverse matrix, known as
pseudo inverse, is used. To calculate the pseudo inverse, various • A high penetration rate with minimal demand.
approaches can be applied. The most widely known type is the
Moore–Penrose pseudo inverse [27, 30]. The pseudo inverse of the Since it is more severe and to guarantee system security, we
discussed in this paper the electric net behaviour during evening
matrix H, namely H†, can be computed via least-square solution
peak: 145 MW, with a high wind energy penetration (20 MW).

5060 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 1 Algerian (Adrar) 22-bus system in power word simulator

Table 1 Algerian electric demand during a winter day


Day peak, MW Evening peak, MW Night dip, MW
117 145 81

Fig. 2 Power system profile at steady state


(a) Voltage profile,
(b) Total losses

Fig. 3 Wind farm insertion in Algerian (Adrar) 22-bus system

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070 5061
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 4 Grid profile while wind farm integration
(a) Voltage profile,
(b) Total losses

Table 2a Line flow and limit monitoring


From num To num MW from Mvar from MVA from Percentage of MVA
5 18 −13.4612 26.6405 29.8483 96.7373
7 21 28.5823 −13.6572 31.6775 103.2067
21 10 23.2835 −19.052 30.0849 98.4114
22 10 −18.8158 26.4295 32.4431 101.3847
22 18 18.8158 −26.4295 32.4431 101.3847

Table 2b Load shedding


Number of buses Current MW Current Mvar Base MW Base Mvar
9 9.9924 2.0000 10.0000 2.0000
13 9.9791 1.9958 10.0000 2.0000
14 9.7806 1.9958 10.0000 2.0000
15 4.9870 0.9974 5.0000 1.0000
16 4.6657 0.9331 5.0000 1.0000
16 4.6657 1.8663 5.0000 2.0000
19 4.6068 0.9214 5.0000 1.0000
19 4.6068 1.8427 5.0000 2.0000
20 9.9436 2.9831 10.0000 3.0000

Voltage profile and losses in this case (wind energy connection) respectively. Furthermore, the training, testing time and the
are exposed in Figs. 4a and b. accuracy values for ELM proposed approach are exposed in
Table 2a exposes line flow in violated lines precising the Table 4a as well for SVR and NNs algorithms in Table 4b (Fig 8).
overload percentage, whereas Table 2b shows the unsatisfactory
load connected to different buses. 5.4.2 Critical scenarios using ELM algorithm: In this phase, we
are going to vary the load at bus 15:(8, 10, 13) MW as it is shown
5.3 UPFC insertion-system relief in Figs. 9a–12a, the main aim is to examine the ELM approach
generalisation performance. The system relief results are exposed
To reduce the flow over the violated lines and to enhance the in Figs. 9b–11b and grid parameters are shown in Tables 5a, b–7a,
system security, we connected UPFC devices in overloaded lines b: first case, load at bus 15 = 8 MW see (Fig. 9, Table 5a, b);
(7–21), (21–10), (10–22) (22–18) and (18-5) as it is shown in second case, load at bus 15 = 10 MW, see (Fig. 10, Table 6a, b);
Fig. 5. Power flow is always solved based on Newton–Raphson and third case, load at bus 15 = 13 MW; see (Fig. 11, Table 7a, b).
technique.
The grid voltage profile and total system losses are shown in
Figs. 6a and b, respectively. Line flow obtained results are 5.5 Results analysis
presented in Table 3a and load consumption, which is satisfied in 5.5.1 Electrical grid steady-state operation: In the considered
Table 3b. power system [Algerian (Adrar) 22-bus system Fig. 1], the net is
Significant enhancement of the system performance is observed operating under normal conditions (steady state). The load is
in Figs. 6a and b. By reducing the grid losses and maintaining the satisfied, no violations, weak elements or overloaded transmission
voltage profile of the various buses within acceptable values. lines. It maintains requirements of reliability and safety. Power
flow analysis is solved using Newton–Raphson approach. Levels of
5.4 Extreme learning machine the voltage at any bus lie within acceptable values 0.99–1.06 pu
(Fig. 2a). The total system loss rate is acceptable 0.6 MW what
5.4.1 ELM algorithm and implementation: On the basis of means (0.4%) of the total consumption (Fig. 2b). AGC mode is on
UPFC simulation section (Fig. 5), obtained results are used as for the two generators connected at bus 11 to enable the generation
inputs of ELM-based algorithm (ELM). The off-line training variation automatically (generators rescheduling). Generators
validation and test process are the first step to build the predictor connected at bus 11 and also at bus 6 are on AVR mode to regulate
model (adjust the weights of the hidden layers), then the model can voltage buses automatically within generators limits. The total
be utilised in real power system time. The performance, training– hourly cost reaches the value €3990.90/h to reduce this rate; wind
testing time and regression are illustrated in Figs. 7a–c, source is inserted in the electrical grid.

5062 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 5 System relief using UPFC

Fig. 6 Grid profile for system relief using UPFC


(a) Voltage profile,
(b) Total losses using UPFC

Table 3a Line flow and limit monitoring


From num To num MW from Mvar from % of MVA
5 18 −12.9390 −7.5041 46.7423
7 21 27.2428 2.8996 85.9665
21 10 22.1400 2.4917 69.624
22 10 −18.0231 −2.4331 56.8329
22 18 18.0233 2.4400 60.0993

Table 3b Load shedding


Number of buses Current MW Current Mvar Base MW Base Mvar
9 10.0000 2.0000 10.0000 2.0000
13 10.0000 2.0000 10.0000 2.0000
14 10.0000 2.0000 10.0000 2.0000
15 5.0000 1.0000 5.0000 1.0000
16 5.0000 1.0000 5.0000 1.0000
16 5.0000 2.0000 5.0000 2.0000
19 5.0000 1.0000 5.0000 1.0000
19 5.0000 2.0000 5.0000 2.0000
20 10.0000 3.0000 10.0000 3.0000
Total losses = 0.7834 MW and − 15.7190 Mvar.

5.5.2 Wind farm penetration: Fig. 3 shows the wind farm (20 • Significant transits on lines (7–21)103%, (21–10)98%, (10–
MW) insertion at bus 15, which caused: 22)101%, (22–18)101% and (18–5)96%.
• Voltage drops (0.57 pu) at bus 19 (Fig. 4a).
• Generators output and load flow variation. • Load shedding Table 2b.

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070 5063
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
• Significant reduction in the operation hourly cost €2886.72/h. low-voltage led to overload and overheating of transmission lines
and risk of equipment failures. It is not allowed to transport
The increased reactive power generated by the plant connected at reactive energy. In general, it is more cost-effective at the system
the slack bus (32 Mvar), supplied by transmitting lines, caused the level to generate reactive power at the location, where it is needed
voltage to fall drastically at the receiving-end side (buses). This to avoid higher losses and large voltage deviations. Furthermore,

Fig. 7 ELM proposed algorithm training process


(a) Performance,
(b) Training and testing times,
(c) Regression

Table 4a Training, testing time and accuracy for ELM


Training time, s Testing time, s Training accuracy Test accuracy
0.0313 0.0023 4.256 × 10−21 1.946 × 10−21

5064 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table 4b Training, testing time and accuracy for SVR and NN algorithm
Algorithm NN SVR
training time, s 21 34
accuracy 2.410 × 10−16 3.666 × 10−14
Total losses = 0.7816 MW and −15,7507 Mvar.

Fig. 8 System relief using ELM algorithm

the transport of reactive power is possible only over limited while others increase their output (generation values at bus 11:9.8
distances. MW, slack generator: 22 MW and −3 Mvar) as results:
We have a threat to the system security, which must be solved
immediately. System losses reached 2.39 MW (Fig. 4b). The • Congestion problem well-handled in optimal manner,
critical limits are mentioned in red colour. overloaded lines relief (7–21) 86%, (21–10)70%, (10–22)57%,
(22–18)60% and (18–5)47% (Fig. 5).
5.5.3 Response of power system to UPFC connection: A • Voltage profile improvement (bus16 from 0.58 to 1.01 pu), (bus
corrective control action must be taken to eliminate such overloads. 19 from 0.57 to 1 pu) (Fig. 6a). This enhancement in voltage
The most practised technique for overload alleviation is the facts values is caused by the UPFC connexion as well the AVR in
devices insertion. In our case, we tried to improve the system generators connected at buses 6 and 11. To allow for
behaviour and to enhance the grid parameters utilising UPFC automatically maintaining generators output terminal voltage at
device. a set value under varying load, by modifying the excitation of
To change the power flow way in the electrical system, we need the generator within the machine's limit.
to be able to control all the three parameters simultaneously: • Avoidance of: elements outages, cascade failure and load
voltage, impedance and phase angle. Four UPFC devices are shedding which is satisfied.
connected in congested lines (7–21), (21–10), (22–18) and (18–5) • Remarkable reduction in total losses, attaining 0.78 MW.
to control the transferred powers at sending and receiving sides and • Minimum rescheduling cost reaching €2865.69/h.
to monitor flows through transmission lines.
The initial real and reactive power flows in the transmission 5.5.4 Extreme learning machine: For further optimisation in
lines are 28.54 MW and − 13.66 Mvar (7–21), 23.25 MW and time and accuracy, we introduced ELM-based alleviation algorithm
−19.02 Mvar (21–10), 18.78 MW and −26.36 Mvar (10–22), 18.78 ELM.
MW and −26.36 Mvar (22–18) and 13.43 MW and −26.54 Mvar The proposed ELM approach has been implemented in
(18–5). After the UPFC connection, the line flow is changed to MATLAB software.
27.24 MW and 2.90 Mvar (7–21), 22.14 MW and 2.49 Mvar (21– Results obtained in the precedent section (UPFC results) are
10), 18.02 MW and 2.43 Mvar (10–22), 18.02 MW and 2.44 Mvar used as input matrix; containing power system operating variables,
(22–18) and 12.94 MW and 7.50 Mvar (18–5) as it is shown in which are: bus number (n), bus voltage values (Vi), phase angle
(Fig. 5). The series converter of UPFC controls the transmission
(δi), load demand (PDi real and QDi reactive powers). The target
line real/reactive power flows by injecting a series voltage,
whereas the shunt converter controls the UPFC bus voltage/shunt matrix is line flow parameters: (Pij real and Qij reactive powers),
reactive power. The interaction between the series injected voltage, transmission line capacity (C) and generators output (Pgi, Qgi)
and the transmission line current leads to a real and reactive power (Fig. 12).
exchange between the series converter and the power system. That The database is randomly partitioned into two non-overlapped
means UPFC can generate or absorb the needed reactive power sets. One consists of 75% of the total input matrix for training and
locally by the switching operation of its converters. Thus, the validation. The other set which is about 25% is reserved as the test
decrease of the reactive generated power by the plant connected at set. The optimal ELM structure can be determined as the one
the slack bus. which results in the lowest validation error. The generalisation
We can see clearly that the line flow, bus voltage and the (ELM approach used in real power system time) is based on the
generator's output are changed (AGC mode on and AVR mode on). increment of the power demand among the load bus (load at bus
A variation in the generation schedule is noted due to the decrease 15).
in transmission line reactive power), some generators back down The training process has been repeated for 50 times, Tables 4a
and b show a comparison between the proposed ELM-based
IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070 5065
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 9 ELM-based alleviation for Algerian grid overload (load at bus 15 = 8 MW)
(a) Algerian grid with wind farm and load at bus 15 = 8 MW,
(b) System relief based on ELM (load at bus 15 = 8 MW)

alleviation and other techniques which are the NN and SVR. It is The new parameters offered by ELM-based alleviation
clear that ELM approach is the fastest algorithm and the most algorithm are sent instantly to AVR and AGC to manage the
accurate. The performance of ELM-based alleviation algorithm is congestion.
evaluated including the mean squared error. The best training and The generators are rescheduled using AGC. AGC optimises the
test accuracy are about 4.2564 × 10−21 and 1.9465 × 10−21, operation of generation units using real-time data such as
respectively, reached at 58 iterations as it is mentioned in Table 4a frequency, actual generation, tie-line load flows and plant units’
and also Figs. 7a and b. The training time reaches 0.03125 s, controller status to provide generation changes (Figs. 9b–11b). The
whereas the test time is 0.0023 s. The exactitude can be interpreted AVR allows for automatically maintaining generators output
in the plot (Fig. 7c), the outputs and the targets are entirely terminal voltage at a set value under varying load, taking into
identical. account the limit of excitation.
The model is built; bias and weights are fixed. The reduction of the transported reactive power participated in
Fig. 8 shows the system relief utilising ELM trained model. The system mitigation. This reduction is due to the minimisation of the
overload is eliminated successfully. A significant reduction in reactive power generation at bus 3 −3.52, −3.55, −3.66 Mvar,
system loss is obtained: 0.7816 MW and −15.7507 Mvar, close to Tables 5b–7b for the three cases, respectively.
results found in UPFC section: 0.7834 MW and −15.7190 Mvar. It is evident to note from the different results of the studied
Finally, to improve the generalisation performance of the ELM- cases: the overload mitigation, load satisfaction, enhancement of
based alleviation approach, the trained model can be applied online the voltage profile, minimisation of the total system losses and cost
for real power system time, by varying the load in bus 15 to reach congestion. The benefit of the utilised technique is the easy, high
the different values 8, 10, 13 MW. The overload results are shown and global control and coordination. Corrections can be achieved
in Figs. 9a–11a. automatically since AGC and AVR regulator signals, and the tie-
line protections are masterminded automatically and instantly. The

5066 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
new corrected planning schedule issue of the proposed novel can 6 Conclusion
be sent instantly to the AGC and AVR to proceed the congestion
management without any need to the UPFC devices insertion since As a novel and promising learning technology, ELM has solved
the facts technology is very expensive, and can amplify the many real-world engineering problems including electric power
congestion cost more. systems complications. The ability to regulate power flow through
The timing is the most crucial factor in alleviating the grid transmission lines and to alleviate system overload is of particular
overload, and this is the advantage of the proposed approach. The importance. Therefore, we employed ELM in this paper to treat
time reduction is a minimisation of the threat and risk of outages this huge and severe problem caused by the wind energy
and cascade failure; furthermore, a blackout. Additionally, the connection. In this paper, we have investigated its remarkable
diminution of the flow in the congested lines and also the efficiency, simplicity and impressive generalisation performance.
congestion management time leads to a fuel optimisation; this can The ELM proposed novel eliminated complete overload, without
be seen clearly by making a simple comparison between losses and the need of inserting UPFC devices, and by consequence
congestion cost in the different cases. minimising the congestion cost. The case study showed that the
value of the weight of its hidden layer need not be tuned, and this is
• First case: total losses: 0.7725 MW and − 15.8885 Mvar; what makes it extremely fast. Obtained simulation results showed a
congestion cost: €2878.68/h real-time learning about seconds, milliseconds and even
microseconds, whereas other techniques such as NNs and SVMs
• Second case: total losses: 0.7707 MW and − 15.8836 Mvar;
are suffering from slow learning speed and the poor learning
congestion cost: €2888.48/h
scalability. The observed reduction in computational time allows
• Third case: total losses 0.7647 MW and − 15.6352 Mvar; estimations closer to the real time of power system, relevant to
congestion cost: €2891.78/h

Fig. 10 ELM-based alleviation for Algerian grid overload (load at bus 15 = 10 MW)
(a) Algerian grid with wind farm integration and load at bus 15 = 10 MW,
(b) System relief based on ELM (load at bus 15 = 10 MW)

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070 5067
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 11 ELM-based alleviation for Algerian grid overload (load at bus 15 = 13 MW)
(a) Algerian grid with wind farm integration and load at bus15 = 13 MW,
(b) System relief based on ELM (load at bus 15 = 13 MW)

Table 5a Line flow and limit monitoring


From num To num MW from Mvar from Percentage of MVA
5 18 −12.5374 −7.6219 45.8413
7 21 26.8301 2.8569 84.7097
21 10 21.7305 2.5029 68.3568
22 10 −17.6182 −2.5086 55.6121
22 18 17.6182 2.5086 59.0451

Table 5b Generator output


Number of bus Gen MW Gen Mvar
3 22.82 −3.47
11 13.66 5.00
11 13.30 5.00
Total losses = 0.7725 MW and − 15.8885 Mvar.

5068 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Table 6a Line flow and limit monitoring
From num To num MW from Mvar from Percentage of MVA
5 18 −12.3771 −7.6297 45.4369
7 21 26.6655 2.8059 84.1893
21 10 21.5671 2.4702 67.8379
22 10 −17.4566 −2.4975 55.1072
22 18 17.4566 2.4975 58.5805

Table 6b Generator output


Number of buses Gen MW Gen Mvar
3 22.47 −3.52
11 11.33 5.00
11 10.97 5.00
Total losses = 0.7707 MW and −15.8838 Mvar.

Table 7a Line flow and limit monitoring


From num To num MW from Mvar from Percentage of MVA
5 18 −12.1898 −7.5711 44.8428
7 21 26.4731 2.6912 83.5495
21 10 21.3760 2.3717 67.2099
22 10 −17.2675 −2.4181 54.4874
22 18 17.2675 2.4181 57.9587

Table 7b Generator output


Number of buses Gen MW Gen Mvar
3 22.68 −3.55
11 12.25 5.00
11 11.89 5.00
Total losses = 0.7647 MW and −15.6352 MV.

short-term operational assessments, which is the major gain in this


paper.
The effectiveness of the ELM proposed novel could add a
reasonable economic and technique improvement in real power
system operation, especially the overload problem.

Fig. 12 Flowchart of ELM algorithm and implementation

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070 5069
© The Institution of Engineering and Technology 2019
17518695, 2019, 22, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2019.0531 by Cochrane Poland, Wiley Online Library on [07/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
7 References [17] Mehrdad, T.H., Mehdi, A.J.: ‘Robust method for state estimation of power
system with UPFC’, Turk. J. Electr. Eng. Commun., 2010, 18, (4), pp. 571–
[1] Guang-Bin, H., Qin-Yu, Z., Chee-Kheong, S.: ‘Extreme learning machine: 582
theory and applications’, Neurocomputing, 2006, 70, pp. 489–501 [18] Maza, O.J.M., Acha, E., Garcia, S., et al.: ‘Overview of power electronics
[2] Shafiullah, M., Abido, M.A., Al-Hamouz, Z.: ‘Wavelet-based extreme technology and applications in power generation transmission and
learning machine for distribution grid fault location’, IET Gener. Transm. distribution’, J. Mod. Power Syst. Clean, 2017, 5, pp. 499–514
Distrib., 2017, 11, (17), pp. 4256–4263, doi:10.1049/iet-gtd.0656 [19] Vikram Rajan, J.: ‘Unified power flow controller used power system stability
[3] Boser, B.E., Guyon, I.M., Vapnik, V.N.: ‘A training algorithm for optimal enhancement under three phase fault’, Int. J. Res. Eng. Technol., 2014, 3, (7),
margin classifiers’. Proc. COLT ‘92 Proc. Fifth Annual Workshop on pp. 131–139
Computational Learning Theory, Pittsburgh, PA, USA, 27–29 July 1992, pp. [20] Ghahremani, E, Innocent, K.: ‘Optimal placement of multiple-type FACTS
144–152 devices to maximize power system loadability using a generic graphical user
[4] Vapnik, V., Golowich, S.E., Smola, A.: ‘Support vector method for function interface’, IEEE Trans. Power Syst., 2013, 28, (2), pp. 764–778
approximation, regression estimation, and signal processing’, Adv. Neural Inf. [21] Sebaa, M., Allaoui, T., Denai, M., et al.: ‘A robust adaptive fuzzy control of a
Process. Syst., 1997, 9, pp. 281–287 unified power flow controller’, Turk. J. Electr. Eng. Commun., 2012, 20, (1),
[5] Canu, S., Grandvalet, Y., Guigue, V., et al.: ‘SVM and kernel methods pp. 87–98
MATLAB toolbox’, Percept. Syst. Inf., 2003, 69, p. 70 [22] Georgilakis, P., Hatziargyriou, N.: ‘Unified power flow controllers in smart
[6] Huang, G.A., Guang-Bin, H., Shiji, S., et al.: ‘Trends in extreme learning power systems: models, methods, and future research’, IET Smart Grid, 2018,
machines: a review’, Neural Netw., 2015, 61, pp. 32–48 2, (1), pp. 2–10, doi:10.1049/iet-stg.2018.0065
[7] Yan, X., Yuanyu, D., Zhao, Y.D., et al.: ‘Extreme learning machine-based [23] Aree, P: ‘Precise dynamic initialization of fixed-speed wind turbines under
predictor for real-time frequency stability assessment of electric power active-stall and active-pitch controls from their aerodynamic power
systems’, Neural Comput. Appl., 2013, 22, pp. 501–508 coefficients using unified Newton–Raphson power-flow approach’, IET
[8] Shan, P., Xinyi, Y.: ‘Deep convolutional extreme learning machine and its Gener. Transm. Distrib., 2018, 12, (1), pp. 9–19, doi:10.1049/iet-
application in handwritten digit classification’, Comput. Intell. Neurosci., gtd.2016.1225
2016, 3, pp. 1–10, doi: 10.1155/2016/3049632 [24] Yang Gu, A.B.N., Junfa Liu, A., Yiqiang Chen, A., et al.: ‘TOSELM:
[9] Gopiya, N.S., Khatod, D.K., Sharma, M.P.: ‘Dispersed generation impact on timeliness online sequential extreme learning machine’, Neurocomputing,
distribution systems’, Int. J. Electr. Electron. Eng., 2012, 2, (1), pp. 2231– 2014, 128, (27), pp. 119–127
5284 [25] Xu, Y., Dong, Z.Y., Meng, K., et al.: ‘Real-time transient stability assessment
[10] Saadat, H.: ‘Power system analysis’, in Stephen, W. (Eds.): ‘McGraw Hill model using extreme learning machine’, IET Gener. Transm. Distrib., 2011, 5,
series in electrical and computer engineering’ (McGraw Hill, New York, NY, (3), pp. 314–322
USA, 1999, 3rd edn.), 697 pages [26] Zhixiang, X.C., Houying, Y.Z., Yuguang, G., et al.: ‘Modified extreme
[11] Labed, I., Labed, D., Mansour, Z., et al.: ‘Influence of the wind farm learning machine with sigmoidal activation functions’, Neural Comput. Appl.,
integration on load flow, voltage in electrical power system’, Int. J. Hydrog. 2013, 22, (3), pp. 541–550
Energy, 2016, 41, (29), pp. 12603–12617 [27] Yakup, K., Apdullah, Y., Esen, Y., et al.: ‘LU triangularization extreme
[12] Jankovıc, S., Ivanovıc, B.: ‘Application of combined Newton–Raphson learning machine in EEG cognitive task classification’, Neural Comput.
method to large load flow models’, Electr. Power Syst. Res., 2015, 40, pp. Appl., 2017, 1, pp. 1–10, doi: 10.1007/s 00521-017-3142-
127–134 [28] Tawfek, M., Zhao, Y.D., Jin, M.: ‘An advanced approach for optimal wind
[13] Houria, S., Rezak, A., Abdellah, M.: ‘Impact of large scale power plant power generation prediction intervals by using self-adaptive evolutionary
connection on congestion in the Algerian electricity transmission system’, extreme learning machine’, Renew. Energy 2018, 126, pp. 254–269
Energy, 2018, 159, pp. 115–120 [29] Yuan, A., Hong Chen, B.N., Yicong, Z.C., et al.: ‘Generalization ability of
[14] Houria, S., Rezak, A., Abdellah, M.: ‘Optimal design of the electric extreme learning machine with uniformly ergodic Markov chains Peipei’,
connection of a wind farm’, Energy, 2018, 165, pp. 972–983 Neurocomputing, 2015, 167, pp. 528–534
[15] Gholami-Khesht, H., Monfared, M.: ‘Novel grid voltage estimation by means [30] Huang, G.B., Siew, C.K.: ‘Extreme learning machine with randomly assigned
of the Newton–Raphson optimization for three-phase grid connected voltage RBF kernels’, Int. J. Inf. Technol., 2005, 11, (1), pp. 16–24
source converters’, IET Power Electron., 2014, 7, (12), pp. 2945–2953, doi: [31] Mahmoud, T., Dong, Z.Y., Ma, J: ‘Advanced method for short-term wind
10. 1049 /iet-pel.2014.0061 power prediction with multiple observation points using extreme learning
[16] Kamel, S., Abdel-Akher, M., Chen, Z., et al.: ‘Developed generalized unified machines’, J. Eng., 2018, 1, pp. 29–38, doi:10.1049/joe.2017.0338
power flow controller model in the Newton–Raphson power-flow analysis
using combined mismatches method’, IET Gener. Transm. Distrib., 2016, 10,
(9), pp. 2177–2184, doi:10. 1049/iet-gtd.2015.1247

5070 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 22, pp. 5058-5070
© The Institution of Engineering and Technology 2019

You might also like