Electric Power Systems Research 143 (2017) 415430

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Application of nonlinear model predictive control based on swarm

optimization in power systems optimal operation with wind
resources
Sahar S. Kaddah, Khaled M. Abo-Al-Ez, Tamer F. Megahed
Electrical Engineering Department, Faculty of Engineering, Mansoura University, Mansoura 35516, Egypt

a r t i c l e i n f o a b s t r a c t

Article history: With the increased share of power generation based on wind energy, the complexity of the unit commit-
Received 19 December 2015 ment (UC) and economic dispatch (ED) problems increases due to the stochastic nature of wind power.
Received in revised form 30 August 2016 Therefore, an accurate and fast optimization method is needed when the generation process involves
Accepted 16 September 2016
large quantities of wind sources to effectively manage generation mix and load requirements. In this
paper, model predictive control (MPC) is used to solve power system UC/ED problems with the presence
Keywords:
of wind energies. Because the UC/ED problem is nonlinear, non-convex and mixed integer problem, the
Nonlinear
MPC must be used as nonlinear. To produce a nonlinear MPC (NMPC), MPC must be integrated with a fast
Model predictive control
Swarm optimization optimization methodology. This paper presents a generic mathematical formula for a NMPC, integrated
Controlled autoregressive integration with swarm optimization technique to describe the nonlinear behavior in the mathematical formula-
moving average tion. This new formulation will be called swarm model predictive control (SMPC) optimization. The
Optimal operation control model will be able to address the effect of the system disturbances and uctuations using a con-
Wind power trolled autoregressive integrated moving average (CARIMA). A general form of future predictions can be
expressed as a function of input and output past data, and a future control sequence, and the degree
of freedom in the SMPC problem. Also, the prediction part improves the swarm technique, because it
identies the size of search space in a better way. In this paper, UC schedule is designed using the swarm
technique ofine, while ED is solved using the proposed SMPC optimization method on real-time basis
and fed into the automatic generation control system. There is no load decit in real time ED results
in less spinning reserve requirements compared with swarm optimization and dynamic program tech-
niques used discretized load. The system under study is a standard IEEE 30 bus modelled in MATLAB
environment, with all data from the city of Florida, USA.
2016 Elsevier B.V. All rights reserved.

1. Introduction

A main function of the smart electrical power system is to coordinate and control power generation from both conventional and
renewable sources including wind power. When solving unit commitment (UC) and economic dispatch (ED) problem for large power
system, a high computational capacity is required to the system. Single optimization approaches cannot achieve this goal, so combination
techniques were introduced to integrate two or more optimization techniques in order to benet from their strengths and overcome the
weakness in solving the optimization problems. There is a rich literature on hybrid methods as in [14]. Although these combinations are
better than single approaches, some combination methods computational effort represents a burden. Also, none of the currently available
optimization techniques are able to avoid shortage or surplus of wind power by solving the dispatch problem with a high frequency, e.g.,
every 24 min. UC/ED problems quickly become very complex and extremely difcult to solve within a limited time.
Thus, to account for the variations in power supply from the wind energy sources, and solve UC/ED problems within limited times, this
paper introduces an improved model predictive control (MPC) method. MPC has been popular since the 70s of the past century [5]; and
in the 90s, it has become a standard advanced control technique in many petrochemical processes [6]. Recently, the application scope for

Corresponding author.
E-mail addresses: mtamer fa@mans.edu.eg, mtamer fa@yahoo.com (T.F. Megahed).

http://dx.doi.org/10.1016/j.epsr.2016.09.013
0378-7796/ 2016 Elsevier B.V. All rights reserved.
416 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430

MPC has expanded and covers not only the petrochemicals and rening elds, but also food processing, automotive, metallurgy, aerospace
and defense industries [7]. The main advantages of MPC are its ability to predict performance over the future horizon, its quick processing
capacity that enables it to tackle changing patterns within very short timelines, and its being suited for multivariable control operations.
In addition, the superior performance of MPC in systematically handling constraints, through incorporating the constraints directly into
the objective functions, makes it theoretically a perfect real-time optimal control paradigm equipped with process integration ability.
Real-time is mean that MPC can solve the dispatch problem with a high frequency.
Even though MPC has been used for a long time as a standard control in industry, its utilization is limited in electrical systems and is
still the subject of many researches. Electricity is a vital economic sector that requires an accurate control model such as MPC to reach
optimal operation. Therefore, numerous researches are being carried out with the aim of improving the performance of MPC algorithms
or addressing the weakness of MPC such as the stability and robustness issues in electrical systems. Authors in [8] solved the dynamic ED
problem using MPC in both regulated and deregulated systems. Authors in [9] presented potential benets of applying MPC to solve the
multi-objective economic dispatch problem in electric power systems with intermittent resources in order to minimize generation costs.
A predictive economic dispatch algorithm is utilized as an optimization strategy that computes the dispatch schedules by minimizing
marginal dispatch costs. Authors in [10] incorporated stochastic problem with MPC scheme to further compensate the uncertainty though
the feedback mechanism. Authors in [11] investigated the German power system for high penetration levels of wind and solar energy and
used MPC to manage storage units charging from renewable source. Authors in [12] combined optimization of the UC problem and the
economic MPC problem, for optimal operation of power systems with the increased share of intermittent renewable energy sources in the
power supply.
In this paper, MPC is used to solve power system UC/ED problems on the presence of high wind energy penetration. The uctuating
behavior of wind power generation is addressed by developing a forecasting model using autoregressive integrated moving average
(ARIMA) [13]. Most of the MPC software available in the market uses linear models, whereas most processes are nonlinear. Because the
UC/ED problem is nonlinear, non-convex and mixed integer problem, the MPC must be used as nonlinear. Papers [812] solved UC/ED
problem using standard MPC by linearization of the problem, while the problem must be solved as nonlinear to reach the most accurate
analysis.
To produce a nonlinear MPC (NMPC), MPC must be integrated with an optimization methodology as in [14,15]. In this paper, a NMPC is
developed to resolve the problem using simultaneous methods [5], where the system is reduced to algebraic equations through a weighted
residual method. To avoid defects of simultaneous methods, where the valid iteration for the system takes a considerably shorter time
than the optimization, the MPC is integrated with fast optimization techniques, e.g. particle swarm. In addition, NMPC prediction horizon
is made longer than optimization time. To sum up, this paper presents an MPC model that is modied to be a swarm algorithm based
model predictive control. This new formulation is generic for several industrial purposes. In electrical generation problems, this paper
introduces the NMPC to minimize operation cost by calculating the generation capacity of the generating units to satisfy all constraints.
So, the new formula is used in solving the optimal dispatch problem. Swarm model predictive dispatch (SMPD) decisions are based on
UC schedules that record the daily ordering of the units on/off [16] in the system, to match the anticipated load with the forecasted wind
power using swarm technique. Also, SMPC is enhanced by inputting disturbance rates to achieve best optimization in limited timelines,
where the disturbance effect is calculated by using the prediction model. Disturbance in electrical systems can be due to factors including
uctuation of renewable sources, variation of the load demand, and response of the automatic generation control (AGC), etc. Controlled
autoregressive integrated moving average (CARIMA) is used to predict the disturbance, and enhance the swarm performance by better
identifying search areas. The main advantage of SMPD is its mathematical formulation as a real-time optimization problem solver that
computes the dispatch actions. The real-time optimization decisions are subject to AGC to manage the machines generation level.
The rest of the paper is organized as follows: Section 2 introduces the multi-objective UC formulation; Section 3 introduces the swarm
optimization methodology; Section 4 introduces the new generic formulation of NMPC; Section 5 illustrates NMPC application in electrical
eld; Section 6 introduces the framework models implemented in Matlab; nally, Section 7 illustrates numerical case studies. The system
under study in this paper is a standard IEEE 30 bus, with wind power data of the city of Florida, USA. The values of total power generation
from wind farms are representing a penetration of 35%.

2. Unit commitment formulation

The objective of this paper is to minimize the total operation cost which contains fuel cost (FC ti ), operation and maintenance cost (OMC ti ),
and start-up cost (SUC ti ) as shown in Eq. (1). The objective function is subject to multi-constraints. These constraints are: the power balance
constraint, the generation limits constraint, and the ramp rate constraint as shown in Eq. (2) [17].
T    N
Minimize FC ti + OMC ti + SUC ti 1 Uit1 .Uit (1)
t=1 i=1

FC ti = ai P 2i + bi Pi + ci (2a)

OMC ti = m.P ti .t + Rt .rp (2b)

tc /
SUC ti = (Cb .tb .Fi ) + Cc (1 e ) .Fi + Cf (2c)
NL
Subject < to Lt = gl [Vx2 + Vy2 2Vx Vy cos(x y )] (2d)
l=1
N  
Pit + PW
t
Dt + L t + R t (2e)
i=1

Pi,min Pit Pi,max (2f)

Pit1 Pit i (2g)

 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430 417

Pit Pit1 i (2h)

Where: T is a number of hours in the planning horizon; N is the number of all units in the system; Pit is a real power generation for ith unit
at time t [MW]; ai , bi , ci is the cost coefcient for ith unit; Fi is the fuel price for ith unit [$/MBtu]; tb is number of hours the unit was banked
[Hour]; Cc is cold-start constant [MBtu]; Cb is banking-start constant [MBtu/Hour]; tc is number of hours the unit was cooled [Hour];  is
thermal time constant for the generating unit; Cf is xed start-up cost [$]; m is maintenance cost coefcient of ith unit at time t per MWh
[$/MWh]; Rt is a real power reserve at time t [MW]; rp is the reserve cost [$];PW t is the wind power generated at time t [MW];Dt is real

power demand at time t [MW]; Lt is a real power loss at time t [MW]; NL is the number of transmission lines; gl is the conductance of
the Lth line that connects buses x to bus y; Vxt , Vyt are the bus voltage at time t [V]; tx , ty are the bus voltage angle at time t [deg]; Pi,min is
minimum real power capacity for ith unit [MW]; Pi,max is the maximum real power capacity for ith unit [MW]; i is the ramp down rate
for ith unit [MW/minute]; and i is the ramp up rate for ith unit [MW/minute].

3. Particle swarm technique

In this paper, particle swarm technique is used to nd UC decision in a day schedule. Particle swarm optimization is a commonly used
optimization technique, which simulates the behavior of bird ocking or sh schooling [18]. Swarm consists of pi (t) particles; and each
particles position denotes xi (t); the position pi (t) changes by adding a velocity vi (t) to the current position. The velocity and position update
equations are given by Eqs. (3a), and (3b) respectively [19].
   
vi (t) = j vi (t 1) + c1 .r1 xpbest xi (t) + c2 .r2 xgbest xi (t) (3a)

xi (t) = xi (t 1) +vi (t) (3b)

Where: c1 , c2 are the positive numbers represent cognitive and social components respectively; r1 , r2 are the uniform distribution numbers
in the range [0,1]; j is an inertia weights.
To improve swarm algorithm to speed up the process of optimization, inertia weight is inserted. Inertia weight as shown in Eq. (4) is a
proportional agent that is related to the speed of last iteration [20].

max min
j = max j (4)
jmax

Where: j is the number of current iteration; jmax is the maximum number of iteration; max is the initial weight = 0.9; min is the nal
weight = 0.4.
Swarm algorithm can be applied through the following steps:

1. Initialize the swarm, pi (t), xi (t) of each particle. pi (t) is random within the hyperspace, with t = 0.
2. Evaluate the performance of each particle, using its current position xi (t).
3. Evaluate the tness function for each particle and nd the xpbest .
4. For each individual particle, compare the particles tness value with its xpbest . If the current value is better than the xpbest value, then
this value updates the current particles position, xi , as pi .
5. Identify the particle that has the best tness value. The value of its tness function is identied as xgbest and its position as pg .
6. Update the xi (t); vi (t) of all the particles
7. Repeat steps 25 until a stopping criterion is met (e.g., maximum number of iterations or a sufciently good tness value).

4. Nonlinear model predictive control

NMPC is a control methodology for optimal operation and control of dynamic system behavior. The optimization yields an optimal input
sequence for the entire horizon [5]. The idea of using NMPC is to predict the output in the prediction horizon. The prediction horizon and
the size of control action are adjusted according to the driving process output to target, besides some operating constraints. First move
trajectory of computed control action sequence is implemented while the other trajectory moves are optimized using nonlinear behavior
through incorporating the swarm technique to better identify the objective rather than tracking a predened trajectory as in the traditional
MPC framework.
Fig. 1 illustrates NMPC system that consists of an estimator and a dynamic optimizer. Inputs to the NMPC are the target trajectory r,
and the estimated state x. Then, NMPC returns manipulated inputs u, to the physical plant, such that the output z, ts the target as well as
possible.
NMPC scheme can be summarized as follows:

1. Obtain measurements/estimates of the states of the system

2. Compute an optimal input signal by minimizing a given objective function over a certain prediction horizon in the future using a model
of the system.
3. Implement the rst part of the optimal input signal until new measurements/estimates of the state are available.
4. Return to step 1
418 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430

Fig. 1. NMPC block diagram.

Fig. 2. A generic stochastic input-output model.

4.1. Mathematical formulation of SMPC

A generic multi-input multi-output model relation is formulated for the process model of the system shown in Fig. 2. Swarm technique
is used to develop nonlinear performance in MPC to produce SMPC.
d
x (t) = f (x (t) , u (t) , d (t) , (t)) (5a)
dx
y (t) = g(x (t) , n (t)) (5b)

z (t) = h(x (t)) (5c)

Where: x(t) is the state vector, u(t) is the manipulated input variables, d(t) is the disturbance, n(t) is the stochastic measurement noise and
w(t) is the stochastic process noise. The measured outputs y(t) which satises the objective function and the controlled outputs z(t).
From the swarm Eq. (3), one can represent state vector trajectory as:

xi (t + 1) = A.xi (t) + B.vi (t) + C.xpbest + D.xgbest (6)

C C
Where: A = 1+C1+C ; B = 1+C j+C ; C = 1+C 1+C ; D = 1+C 2+C
1 2 1 2 1 2 1 2
Through the MPC theory applied in state-space mode [7], the input/output system can be represented using swarm optimization
technique based on MPC as follows:

xk+1 = Axk + Bvk + Cxpbest + Dxgbest + EUk + Fk (7a)

yk = Gxk + dk + nk (7b)

Zk = GZ xk + dk (7c)

Where k = 0, 1. . .. Tp ; Tp is the predictive horizon.

Assume that the model and the true system are identical. Then, process noise and measurement noise k , nk are neglected. Also, assume
that the model best guesses are dk+n = . . . = dk+1 = dk so the model equations are:

xk+1 = Axk + Bvk + Cxpbest + Dxgbest + EUk (8a)

yk = Gxk + dk (8b)

Zk = GZ xk + dk (8c)

The rst step of prediction can be used recursively to nd an n-steps of prediction as follows:

xk+1 = Axk + Bvk + Cxpbest + Dxgbest + EUk (9a)

 
xk+2 = A Axk + Bvk + Cxpbest + Dxgbest + EUk + Bvk+1 + Cxpbest + Dxgbest + EUk+1 (9b)

xk+2 = A2 xk + ABvk + Bvk+1 + AEUk + EUk+1 + (AC + C) xpbest + (AD + D)xgbest

 
xk+3 = A A2 xk + ABvk + Bvk+1 + AEUk + EUk+1 + (AC + C) xpbest + (AD + D)xgbest + Bvk+2 + Cxpbest + Dxgbest + EUk+2 (9c)
 
xk+3 = A3 xk + A2 Bvk + ABvk+1 + Bvk+2 + A2 EUk + AEUk+1 + EUk+2 + A2 C + AC + C xpbest + (A2 D + AD + D)xgbest (9e)
 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430 419

A general expression of the n-step ahead of prediction is given as:

xk+n = An xk + An1 Bvk + An1 EUk + An2 Bvk+1 + An2 EUk+1 + + ABvk+n2 + Bvk+n1
n  
+AEUk+n2 + EUk+n1 + Ai1 Cxpbest + Dxgbest (10)
i=1

So, the system output can be similarly as follow:

yk+n = GAn xk + GAn1 Bvk + GAn1 EUk + GAn2 Bvk+1 + GAn2 EUk+1 + + GABvk+n2 + GBvk+n1 + GAEUk+n2
n  
+ GEUk+n1 + G Ai1 Cxpbest + Dxgbest + dk (11)
i=1

Zk+n = GZ An xk + GZ An1 Bvk + GZ An1 EUk + GZ An2 Bvk+1 + GZ An2 EUk+1 + + GZ ABvk+n2 + GZ Bvk+n1 + GZ AEUk+n2
n  
+ GZ EUk+n1 + GZ Ai1 Cxpbest + Dxgbest + dk (12)
i=1

This prediction combines past and future data, so it is advisable to be more careful with notation and construction of the predictions.
A common notation in the literature uses a double subscript where the rst term determines the sample of the prediction (how many
step ahead) and the second denotes the sample of which the prediction was made. Therefore, the general expression of the n-step ahead
prediction is:

xk+n|k = An xk + An1 Bvk|k + An1 EUk|k +

An2 Bvk+1|k + An2 EUk+1|k + +

+ABvk+n2|k + Bvk+n1|k +

AEUk+n2|k + EUk+n1|k +
n  
Ai1 Cxpbest + Dxgbest (13)
i=1

yk+n|k = GAn xk + GAn1 Bvk|k +

GAn1 EUk|k + GAn2 Bvk+1|k +

GAn2 EUk+1|k + + GABvk+n2|k +

GBvk+n1|k + GAEUk+n2|k +
n  
GEUk+n1|k + G Ai1 Cxpbest + Dxgbest + dk (14)
i=1

Zk+n|k = GZ An xk + GZ An1 Bvk|k +

GZ An1 EUk|k + GZ An2 Bvk+1|k +

GZ An2 EUk+1|k + + GZ ABvk+n2|k +

GZ Bvk+n1|k + GZ AEUk+n2|k +
n  
GZ EUk+n1|k + GZ Ai1 Cxpbest + Dxgbest + dk (15)
i=1

The double subscript makes it clear that where a value is a future or a predicted value as opposed to known. It is convenient to separate
predictions into the part which is known, and the part which is yet to be determined.

yk+n|k = yknown + yunknown (16a)

n  
yknown = GAn xk + G A i1
Cxpbest + Dxgbest + dk (16b)
i=1

yunknown = GAn1 EUk|k + GAn2 Bvk+1|k +

GAn2 EUk+1|k + . . . + GABvk+n2|k +

GBvk+n1|k + GAEUk+n2|k +

GEUk+n1|k (16c)
420 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430

Zk+n|k = Zknown + Zunknown (17a)

n  
Zknown = GZ An xk + GZ i1
A Cxpbest + Dxgbest + dk (17b)
i=1

Zunknown = GZ An1 EUk|k +

GZ An2 Bvk+1|k + GZ An2 EUk+1|k +

. . . + GZ ABvk+n2|k + GZ Bvk+n1|k +

GZ AEUk+n2|k + GZ EUk+n1|k (17c)

The aim will be to choose the unknown inputs to ensure that the overall prediction is satisfactory.

4.2. Vector notation

A simple arrow notation captures the impression of a vector representing a set of predictions. Like its components, this vector of vectors
may be a vector in itself.

xk+1|k

x
k+2|k
=
.
..

x
k+n|k

Axk + Bvk|k + Cxpbest + Dxgbest + EUk|k

2
A xk + ABvk|k + Bvk+1|k + AEUk|k + EUk+1|k + (A + 1) (Cxpbest + Dxgbest )

. (18)
.
.
n
 
A xk + An1 Bvk|k + An1 EUk|k + . . . + Bvk+n1|k + EUk+n1|k + ni=1 Ai1 Cxpbest + Dxgbest
Separating into past and decision variable gives:

xk+1|k Axk

x A2 x
k+2|k k
= +
. .
.. ..

x An x
k+n|k k

Bvk|k

ABv + Bv
k|k k+1|k
+
.
..

An1 Bv + . . . + Bv
k|k k+n1|k

EUk|k

AEU + EU
k|k k+1|k
+
.
..

An1 EU + . . . + EU
k|k k+n1|k

Cxpbest + Dxgbest


(A + 1) )(Cxpbest + Dxgbest )

. (19a)
.
.

n i1  
i=1 A Cxpbest + Dxgbest
 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430 421

xk+1|k A

x A2
k+2|k
= xk +
. ..
.. .

x An
k+n|k

vk|k

B 0 0

AB B 0 vk+1|k
+
. .. .. ..
.. ..
. . . .
An1 B B
v
k+n1|k

Uk|k

E 0 0

AE E 0 Uk+1|k
+
. .. .. ..
.. ..
. . . .
An1 E E
k+n1|k U

C 0 0

AC C 0

. .. .. .. xpbest +
.. .
. .
An1 C C

D 0 0

AD D 0

. .. .. .. xgbest
..
. . .
An1 D D
(19b)

Where: vector and matrix can be represented as follow:

Uk|k Uk|k Uk|k

U U U
k+1|k k+1|k k+1|k
= , U k; = < U k; = U ;; ;
. . . k
.. .. ..

U U U
k+n1|k k+n1|k k+n1|k

Therefore, the general expression for the future state using swarm optimization based on MPC feedback strategy, and including past
and decision variables can be expressed as follow:
v U
X
k+1 = PX xk +JX xpbest +LX xgbest + IX k +HX k (20)
Depend on Past Depend upon Decision Variables
     
Similarly; output expression for measured output y k+1 and control output Z k+1 can be obtained as follow:

y k+1 = Pxk + Jxpbest + Lxgbest + I v U k + dk (21)

Z k+1 = Pz xk + Jz xpbest + Lz xgbest + Iz v k + Hz U k + dk (22)

Where: P = GPX ; J = GJX ; L = GLX ; I = GIX ; H = GHX ; Pz = Gz PX ; Jz = Gz JX ; Lz = Gz LX ; I z = Gz IX ; Hz = Gz HX .

4.3. Prediction concept of SMPC

All other optimizing methods are searched through different points, and then they nd the best trajectory between these points, which
represents the best solution. However, optimization may never reach the desired control trajectory if an unmeasured nonzero disturbance
that occurs to the system or if the model error is present.
In the proposed formulation, to achieve offset free tracking performance for the SMPC to reach the desired control trajectory, in the
presence of unmeasured nonzero disturbance, disturbance states are incorporated into the process model as described in previous section.
The most important feature of the SMPC, that is not available in other methods, is that it can reduce disturbance and uctuation [11].
Disturbance can be reduced by using the prediction concept [7]. This can be done by anticipating the output performance, and based on
this anticipation the control trajectory will drive the performance to reach the desired output in a smooth manner and without uctuations.
422 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430

Start

Build System

Section 1
Build Scenario UC Optimizing

SMPD

SMPD loop SMPD Update

Gen. Compute

Section 2
Feed System To AGC

Sim. Step No
Reached?
Plot Result
Yes

End Return history

Fig. 3. Flowchart of the SMPD developed using Matlab.

Also the prediction model suggests the right place to search in, and predicts the swarm parameter to get the best optimization in less
time. CARIMA is used to represent the prediction model in SMPC. The reason for using CARIMA is that it incorporates a disturbance estimate,
and therefore can give unbiased prediction in the steady state, irrespective of some parameter uncertainty.
This section displays the formulation of the prediction of the future target, where disturbance can be estimated using CARIMA. For
convenience, the output is regularly measured, and the rst-step ahead prediction model uses variables of the output and input increment.

a (z) yk = b(z)Uk (23)

Where: a (z), b(z) [7] represent open-loop pole polynomial, open-loop zero polynomial, Uk is process incremental input Uk = Uk
Uk1 .

LetA (z) = a (z)  (24a)

A (z) yk = b(z)Uk (24b)

A (z) = 1 + A1 z 1 + . . . + An z n (24c)

b (z) = b1 z 1 + . . . + bm z m (24d)

Where: n, m are vector of coefcients of polynomial.

Discrete models are one-step-ahead prediction models, that is, given data at sample k, can determine data at sample k + 1.

yk+1 + A1 yk + . . . + An ykn+1 = b1 Uk + b2 U k1 + . . . + bm Ukm+1 (25)

The one-step-ahead prediction can be used recursively to nd an n-steps ahead of predictions as follows:

yk+1 + A1 yk + . . . + An ykn+1 = b1 Uk + b2 U k1 + . . . + bm Ukm+1 (26a)

yk+2 + A1 yk+1 + . . . + An ykn+2 = b1 Uk+1 + b2 U k + . . . + bm Ukm+2 (26b)

yk+3 + A1 yk+2 + . . . + An ykn+3 = b1 Uk+2 + b2 U k+1 + . . . + bm Ukm+3 (26c)

yk+n + A1 yk+n1 + . . . + An yk = b1 Uk+m + b2 U k+m1 + . . . + bm Uk (26d)

 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430 423

Simultaneous equations can be represented using a matrix/vector format, after which, the solution is obvious. An illustration and some
suitable notation are given here, which clearly separate past and future variables.

yk+1 Uk
yk
y
k+2 Uk+1
CA + HA ..
.
= Cb
.
+

... .
ykn+1 .
y U
k+n k+m

U k1 1 0 ... 0

A1 1 ... 0
Hb .. CA = .
(27a)
. .. .. .. ..
. . .
Ukm+1 An1 An2 ... 1

CA y k+1 + HA y k = Cb U k + Hb U k1 (27b)

y u U y
k+1 = CA 1 Cb k + CA 1 Hb k1 CA 1 HA k (27c)
     
Futureinput Pastdata

k = y k+1 yk (28)

A1 A2 . . . An b1 0 ... 0 b2 b3 . . . bm

A2 A3 ... 0 b2 b1 ... 0 b3 b4 ... 0
Where: HA = . .. .. .. C =
b .. .. .. .. , H =
b ..
.. .. .. , k is a disturbance.

.. . . . . . . . . . . .
An An1 ... 0 b bm2 ... b bm bm1 ... 0
m1 1

5. Economic dispatch formulation using SMPC

In this paper, SMPC is used as a dispatch solution in electrical power application. SMPD computes an optimal action based on a dynamic
system, for which an empirical model is obtained by system identication, and its predicted future evolution. The objective is related to
force the system to follow a predened trajectory, or related to an objective function. The optimal trajectory of the cost is obtained by
solving the following equations:
 
Minimize  U k, y k (29)

Subject to X k+1 = PX xk + JX xpbest + LX xgbest + IX v k + HX U k (30a)

y k+1 = Pxk + Jxpbest + Lxgbest + I v k +H U k + dk (30b)

Z k+1 = Pz xk + Jz xpbest + Lz xgbest + Iz v k + Hz U k + dk (30c)

Umin < Uk < Umax (30d)

Umin < Uk < Umax (30e)

Zmin < Zk < Zmax (30f)

Zk + k Zmin (30g)

Zk k Zmax (30h)

k 0 (30i)

The symbols used in electrical formulation are: xk is the control trajectory (power production); yk is the generation cost; Uk is the
amount of power generation from each unit; Umin is the minimum unit power generation; Umax is the maximum unit power generation;
U min , Umax are the ramp down and up rate respectively; Zk is the load demand.
The power system in this paper contains two types of sources, controllable power generators from conventional source, and non-
controllable predictable power generators representing power production from the wind energy sources. ARIMA is used to predict wind
power generation. Each generator is an independent system and modelled separately, and does not directly affect the others. However,
they are interconnected in order to collaborate to satisfy the overall load demand. The controllable and non-controllable power generators
are represented by the following linear systems:

Zi (s) = Gi (s) .Ui (s), Zw (s) = Hw (s) .Uw (s) (31)

Where Zi is the load demand feed from conventional sources; Zw is the load demand feed from wind sources; Gi (s) and Uw (s) are the transfer
functions for conventional sources and wind sources respectively.
424 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430

Table 1
Generators data.

Coefcients Thermal unit number

Unit No. 1 2 3 4 5
BUS 1 2 13 23 27

Cost ai [$/MW2 h] 1.0E-4 1.0E-4 5.0E-4 5.0E-4 1.0E-4

bi [$/MWh] 0.06 0.10 0.05 0.05 0.06
ci [$/h] 27.0 29.0 29.0 35.0 27.0
Pmin [MW] 5 5 5 5 5
Pmax [MW] 150 150 150 150 150
i [MW/h] 50 50 50 50 50
i [MW/h] 75 75 75 75 75

Wind power generation cost is almost zero (with a negligible maintenance cost). Therefore, all energy generated from wind farms must
be utilized, and the traditional units generate the difference amount between the load demand and wind energy. So, the SMPD response to
deal with wind units must be made faster than other conventional units, and this is done by using the system dynamics transfer functions.
1 1
Gi (s) = , Hw (s) = (32)
i s + 1 w s + 1
Wind sources time constant w is less than 5 s to enable a fast dynamic response, and react very quickly to set-point changes. Meanwhile,
power plant dynamics time constant i will be more than 10 s, and thereby have no inuence in our simulations. Also, the time constant for
conventional plants is not equal, but it depends on the response time of automatic generation control of each station. Hence, this sampling
time or lower will be used throughout the SMPD simulations.

6. SMPD implementation

The following part describes the framework and the purpose of the developed algorithm as shown in Fig 3. The framework contains
two sections: the rst is the processing system building, and the second is accounting the economic operation of the system. Because the
most important properties of SMPD is the ability to work online, the rst section of the program will include set up of the power system
conguration and simulation parameters, which is done ofine. The second section computes the economic operation representing the
online SMPD. The framework models were implemented in Matlab 2010b on a PC with Intel CoreTM 2 Duo 3.00-GHz CPU and 4.00 GB of
memory.

Build System Builds the system model connection of transmission line and generator data.
Build Scenario Includes built scenarios of generation. These scenarios depend on UC output, and swarm optimization is used to solve the
UC problem.
SMPD Build the SMPD matrix parameter.
SMPD loop Perform the loop simulation. The function does the following: simulate the system (output measurements from sensor and
state simulation); execute the SMPD algorithm; and store applicable output values from the loop simulation.
SMPD Update Update the current SMPD loop matrices. The following are updated: current power production range (trajectory) for
each power plant; current forecasted power production from wind sources using ARIMA; input and input rate constraints; and
estimates of the disturbance using CARIMA.
Gen. Compute Solve the SMPD loop to compute generation of each unit which satises the objective function. It builds the non-constant
vectors and solves the soft constraint. The obtained solution is returned and fed into the system for each simulation step.
Return History Compare control loop action with the previous historical data.

7. Numerical case study

This section introduces a numerical study combining the UC optimization problem and the SMPD problem in the application of the
optimal operation cost of power systems. In this example, control of the generation strategy is developed with increased wind power
generation. UC is used to develop a day ahead schedule; while EMPD is used to determine the optimal output of each generator while
satisfying the system constraints. Then, online control to the generator is developed through AGC.
The case study is conducted on an IEEE 30 bus as shown in Fig. 4. The system has 5 conventional generators in addition to wind farms.
The wind station capacity is 300 MW. The values of total power generation from wind farms are representing a penetration of 35%. All data
of generators are described in Table 1 [21].
Load curve [22] is conducted on a real power system operated by a local utility company in Florida, in 2012. Winter is used as the load
prole in this simulation, and the percentage of spinning reserve is set at 10% of the load demand as shown in Fig. 5.

7.1. Wind power forecasting using ARIMA

Wind farms generate power at negligible operation cost, thus they should be used as must run units. With high penetration of wind
power which has an intermittent nature, calculating the output power depends on the stochastic behavior of the wind sources. For this,
ARIMA is used to forecast next day wind power generation to achieve maximum benet from wind power in UC/ED planning. ARIMA
algorithm is developed in [13]. However, the weather is likely to change during the day of operations and the power supply from wind
energy sources is unlikely constant. Using shorter term forecasting in the range of 03 h will help in wind power prediction. Wind power
is forecasted at 10-min intervals. The actual power for the Florida wind energy and the predicted power are shown in Fig. 6.
 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430 425

Fig. 4. IEEE 30 bus systems.

Fig. 5. Load demand curve.

Fig. 6. Predicted and actual wind power for Florida output at 10-min interval.
426 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430

Fig. 7. Changes of the best schedule decision according to the number of iteration at hour 3 a.m.

Table 2
Hourly schedule of the conventional units.

unit Day Hour

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 1 1 1 1 1 1 1
3 0 0 0 0 0 0 0 1 1 1 1 1 1 1 0 0 0 1 1 1 1 1 1 1
4 0 0 0 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0
5 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

7.2. Day-ahead planning

The rst step in control strategy is planning the committed and decommitted (binary decisions) of the generator units for the next
day. UC problem derives the production plan for the next 24 h. Table 1 lists the operational parameters for the UC problem. Mathematical
formulation for UC problem is presented in Section 2. Based on the available forecasts of next day demand load and power supply from
wind energy sources, UC decision is obtained for conventional units hour by hour. Swarm optimization technique introduced in Section
3- is used to solve the UC problem once every 24 h. Optimal swarm parameters obtained by the model training are: swarm size ns = 50,
dimension size = 5 acceleration coefcients c1 = 2, c2 = 2, inertia weight max = 0.9, min = 0.4 and maximum number of iteration jmax = 100.
Best schedule is obtained after 30 iterations as shown in Fig. 7. Also, the time taken to get the best schedule is 0.6330 s.
The hourly schedule of conventional units is displayed in Table 2. As shown in Table 2, all units are committed. Cheapest power units
no. 1 and no. 5 remain online continuously during the whole period to produce the majority of the load, whereas the more expensive and
fast power units, no. 2, no. 3 and no. 4, operate whenever it is necessary. Unit no. 2 remains ofine in order to satisfy the minimum down
time requirement (two h). Then, this unit initiates the start-up process at 2:00 AM (in accordance with the start-up ramp rate limit). Unit
no. 2 is decommitted again at 2:00 PM and the unit initiates the start-up process at 5:00 p.m. Unit no. 3 initial start-up is at 7:00 AM and
remain online for 7 h. Then, unit no 3 is start-up again at 5:00 PM Unit no. 4 (as the most expensive unit) is committed at 2:00 AM and
6:00 PM for three hours in each interval, and the unit cannot be shut-down before it satises the requirement for the minimum up time
limit (three hours).

7.3. Real-Time planning

The second step in the control strategy is to prepare an even shorter-term economic planning (real-time control) within the 24-h
planning horizon. This step can be achieved using SMPD that will anticipate future events and take action accordingly rather than the UC
problem. The SMPD described in Section 5 does not include binary decisions. Consequently, the controller only re-optimizes the production
plan and does not decide whether a unit shall be committed or decommitted in a day schedule. Committed and decommitted decisions of
power units for each time step are solely determined by the UC problem.
The variable cost for the UC is [$/MWh], while the variable cost for the SMPD is determined by converting the data into [$/MW ]. Sample
time of the system generator units is up to  = 20 s depending on the unit response, and the simulation predictive horizon is Tp = 100 time
steps. Therefore, the operation parameters of the SMPD can be expressed in Table 3.

7.4. Simulation layout

SMPD simulation was run on the Simulink software package. The control loop as shown in Fig. 1 is realized using Simulink as shown in
Fig. 8. S-Functions are used to implement constrained and system model of the SMPD loop.
Simulink layout in Fig. 8 is described as follows:-
Step 1: Construct the system conguration of line data, bus data and generator data. Step 1 is constructing only once because it represents
the network conguration. This step in the above owchart is represented under Build System.
Step 2: This step consists of three stages. First stage: construct forecasting model of the next 24-h for wind power using ARIMA. Second
stage: based on the load prole distribution in the studied area, construct 24-h demand prole. Final stage: based on the previous two
 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430 427

Table 3
Operation parameter for SMPD.

Unit 1 2 3 4 5

 [s] 20 15 15 20 10
ai [$/MW2 ] 5.5E-7 2.8E-7 1.4E-6 2.0E-6 5.5E-7
bi [$/MW ] 3.3E-4 2.8E-4 1.4E-4 2.1E-4 3.3E-4
ci [$/] 0.15 0.08 0.08 0.146 0.15
Umin [MW] 5 5 5 5 5
Umax [MW] 150 150 150 150 150
U min [MW/] 0.28 0.14 0.14 0.21 0.28
U max [MW/] 0.417 0.2 0.2 0.31 0.417

Fig. 8. SMPD simulink layout.

stages and generator data, obtain the UC decision through using swarm optimization technique. Step 2 is initiated once every 24 h. This
step in the above owchart is represented under Build Scenario.
Step 3: Identify the prediction horizon (Tp ) of SMPD. Then, construct the matrix parameters of SMPD PX , JX , LX , IX andHX . Matrix param-
eters change only when the prediction horizon value changes. Therefore, step 3 is initiated once when the value of Tp is changed. This step
in the above owchart is represented under SMPD.
Step 4: This step contains the online dispatching to compute the production of each generation units. A digital clock initiates obtaining
the forecasted wind power generation, UC decision, and the corresponding value of load demand, at  = 20 s intervals. This data is fed into
the SMPD controller. From SMPD controller and the system model, power production u(t) is computed. This step in the above owchart is
represented under SMPD loop.
Step 5: The values of the power generation u(t) of each unit are fed to AGC through the dynamic response simulation. This step in the
above owchart is represented under Gen. Compute and Feed System.
Step 6: The value of the power generation from conventional units are variable according to the load demand prole and wind power
generation. Therefore, the most important feature of SMPD is model updating. The SMPD Update step, as shown in the above owchart,
feeds the SMPD with the previous power generation and disturbance data. Disturbance is estimated using CARIMA to achieve offset free
tracking performance.
428 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430

Fig. 9. Actual and discretized load curve.

Table 4
Results of a daily operation by three dispatch methods.

Dispatch Type SMPD Swarm technique Dynamic program

Total Demand [MW] 36256 35878 35634

Decit of Load [MW] 378 622
Deviation [%] 1.04 1.71
Total Generation Cost [$] 9591.76 9838.51 9697.329
Runtime [s] 0.04 0.605 2.612
Speedup 15 66

7.5. SMPD simulation result

To ensure the accuracy and speed of the proposed optimization technique, the dispatch result from the SMPD is compared with the
results of other optimization techniques e.g. swarm optimization [23] and dynamic program [20]. SMPD computes the solution on a real-
time optimization, while swarm optimization is computed each 20 min, and dynamic optimization is computed for each hour. Demand
curve is discretized at hourly intervals to simulate the dynamic program operation, and is also discretized at 20-min intervals to simulate
the operation of swarm optimization technique as illustrated in Fig. 9. Load discretizations have an impact on the solution because the
total value of discretized load is less than the real demand load by 378 MW with the 20 min discretization and by 622 MW with the hourly
discretization. Therefore, in order to overcome the negative impact of discretization in the simulations, this paper recommends using the
SMPD which receives the real demand curve.
Fig. 10 illustrates the power generation output of the conventional units using SMPD, swarm optimization technique and the dynamic
program. Swarm technique and dynamic program results include the binary decisions of the conventional units. Therefore, the UC decision
for these techniques is different from the UC decision for the SMPD as described in Table 2. Units no. 1 and no. 2 contribute the most load
feeding. On the other hand, unit no. 4 contributes less in all three types of dispatch solutions. Although the load demand in SMPD is greater
than the demand in the other two techniques, the power generation curves for all units using SMPD are lower than the generation curves
of the other techniques.
Fig. 11 illustrates the total generation cost for all units using SMPD, swarm optimization and dynamic program. It is noticed that the
generation cost using SMPD is less than the other techniques.
Table 4 illustrates the simulation results numerically comparing the three methods of solving the operation problem. With the swarm
optimization and dynamic programming, the power generated is less than the real load due to load discretization. In the swarm technique
the shortage is 1.04%, and in the dynamic program it is 1.71%, while SMPD is free of any shortage. Therefore, the swarm technique and
dynamic program have an increased need for spinning reserve to overcome the shortage drawback. On the other hand, SMPD needs less
spinning reserve. This is desirable, as the spinning reserve is costly and implies unutilized capacity. Also, Table 4 illustrates the runtime
for solving the three methods. It is noticed that the runtime for SMPD is signicantly less than the other two methods. SMPD gives a 15
speedup compared to swarm technique and 66 speedup compared to dynamic program. Consequently, SMPD can indeed be solved in
real-time to obtain an optimal solution.

8. Conclusion

This paper focuses on optimizing the operation of power system by solving the economic dispatch (ED) problem using swarm model
predictive dispatch (SMPD), which is employed for the rst time in this industrial application. The growing uncertainty associated with the
increasing share of intermittent renewable energy sources in the power supply, such as wind has presented new challenges for optimal
operation of power systems. Motivated by these challenges, a novel control strategy is used, which shows capability of managing uncertainty
 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430 429

Fig. 10. Optimal power generation of conventional units.

with exibility. Also, the control strategy is combined with prediction analysis to reduce the model disturbance. Controlled autoregressive
integrated moving average (CARIMA) is used to measure the residual disturbance.
A two-level control strategy is presented: At the high-level, the day-ahead production plan is developed by solving the unit commitment
(UC) problem. At the low-level, the minutes-ahead production plan is set by applying swarm technique and dynamic program, and the
real-time control is performed by applying the SMPD. The proposed solution model is applied on standard IEEE 30 bus system and the
available data of the city of Florida in USA.
The developed control strategy is tested on a power system consisting of a conventional power plant and non-controllable farms of
wind turbines. The results of the simulations successfully show that the innovated control strategy appears to provide a feasible and a
promising solution to overcome some of the important challenges. Furthermore, it shows that the SMPC method plays an important role
in the control of optimal power system operations.
430 S.S. Kaddah et al. / Electric Power Systems Research 143 (2017) 415430

Fig. 11. Total generation cost.

Additionally, solving the UC problem with SMPD yields the optimal production plan. This optimal production plan has fully satised
the load demand. Meanwhile ED solution applying load discretization resulted in 1.04% power shortage when using swarm technique and
1.71% power shortage when using dynamic program. Moreover, the runtime for the SMPD is 66 faster than solving the ED with dynamic
program and 15 faster than solving using swarm technique.

References

[1] M. Sudhakaran, P. Raj, Integrating genetic algorithms and tabu search for unit commitment problem, Int. J. Eng. Sci. Technol. 2 (1) (2010) 5769.
[2] C.C. Columbus, S.P. Simon, Prot based unit commitment for GENCOs using parallel PSO in a distributed cluster, ACEEE Int. J. Electr. Power Eng. 2 (3) (2011).
[3] C.-C.F. Chen, P. Schonfeld, A hybrid heuristic technique for optimizing intermodal logistics timed transfer systems, Procedia Soc. Behav. Sci. 48 (2012) 25662576.
[4] A. Heidari, M.R.A. Pahlavani, H. Dehghani, Reliability and security constrained unit commitment with hybrid optimization method, Majlesi J. Electr. Eng. 9 (1) (2014)
919.
[5] F. Khani, M. Haeri, Robust model predictive control of nonlinear processes represented by Wiener or Hammerstein models, Chem. Eng. Sci. 129 (2015) 223231.
[6] M. Sarailoo, Z. Rahmani, B. Rezaie, A novel model predictive control scheme based on bees algorithm in a class of nonlinear systems: application to a three tank system,
Neurocomputing 152 (2015) 294304.
[7] J.A. Rossiter, Model-Based Predictive Control: A Practical Approach, CRC Press, 2013.
[8] A. Elaiw, X. Xia, A. Shehata, Application of model predictive control to optimal dynamic dispatch of generation with emission limitations, Electr. Power Syst. Res. 84 (1)
(2012) 3144.
[9] M.D. Xie, Model predictive economic/environmental dispatch of power systems with intermittent resources, in: Power & Energy Society General Meeting, 2009.
PES09. IEEE, IEEE, 2009.
[10] A. Parisio, L. Glielmo, Stochastic model predictive control for economic/environmental operation management of microgrids, in: Control Conference (ECC), European.
2013. IEEE, 2013.
[11] P.A. Jonas, Predictive Power Dispatch for 100% Renewable Electricity Scenarios Using Power Nodes Modeling Framework, MS Thesis, Swiss Federal Institute of
Technology, Zurich, Switzerland, 2011.
[12] Dinesen, P.J., Unit Commitment and Economic Model Predictive Control for Optimal Operation of Power Systems. 2015.
[13] S.S. Kaddah, K.M. Abo-Al-Ez, Tamer F. Megahed, M.G. Osman, Impact of renewable resources forecasting on unit commitment solution of Egyptian electric grid, Int. J.
Sci. Eng. Res. 6 (8) (2015) 510518.
[14] D. Schlipf, D.J. Schlipf, M. Khn, Nonlinear model predictive control of wind turbines using LIDAR, Wind Energy 16 (7) (2013) 11071129.
[15] R. Bindlish, Nonlinear model predictive control of an industrial polymerization process, Comput. Chem. Eng. 73 (2015) 4348.
[16] T. Logenthiran, W.L. Woo, Lagrangian relaxation hybrid with evolutionary algorithm for short-term generation scheduling, Int. J. Electr. Power Energy Syst. 64 (2015)
356364.
[17] S. Marrouchi, S. Chebbi, Unit commitment optimization using gradient-genetic algorithm and fuzzy logic approaches, in: Complex System Modelling and Control
Through Intelligent Soft Computations, Springer, 2015, pp. 687710.
[18] C.-S. Lee, H.V.H. Ayala, L. dos Santos Coelho, Capacitor placement of distribution systems using particle swarm optimization approaches, Int. J. Electr. Power Energy
Syst. 64 (2015) 839851.
[19] A. Uriarte, P. Melin, F. Valdez, An improved particle swarm optimization algorithm to optimize modular neural network architectures, in: Design of Intelligent Systems
Based on Fuzzy Logic, Neural Networks and Nature-Inspired Optimization, Springer, 2015, pp. 155162.
[20] A. Ahmad, A. ul Asar, Unit Commitment Using Hybrid Approaches. PhD Thesis, Department of Electrical Engineering, University of Engineering and Technology, Taxila,
Pakistan, 2010.
[21] S. zyn, et al., Charged system search algorithm for emission constrained economic power dispatch problem, Energy 46 (1) (2012) 420430.
[22] Commission, F.P.S., Review of the 2012 Ten-Year Site Plans for Floridas Electric Utilities. 2012.
[23] C. Shang, D. Srinivasan, T. Reindl, An improved particle swarm optimisation algorithm applied to battery sizing for stand-alone hybrid power systems, Int. J. Electr.
Power Energy Syst. 74 (2016) 104117.