Page 1 of 45

EEE 537 Power System Optimization (2 units)

Introduction to the formulation of Optimization problems: Problem variables, problem constraint, the
objective function. Calculus based techniques: Unconstrained Minimization: Powell’s method of
conjugate directions, Gradient methods, second order methods.

Constrained Minimization Problems: Indirect methods by unconstrained minimization, penalty function

approach. Direct methods for constrained optimization; Lagrange multipliers, Kuhn Tucker conditions;
methods of feasible directions. Linear (LP) and Non-linear programming. Application of LP to power
systems problems economic dispatch, automatic load shedding generation expansion studies. Non-
Calculus Based Methods: Guided random search techniques: Evolutionary algorithms, simulated
annealing, genetic algorithms;

Expected Learning Outcomes:

At the end of studying this course, students will be able to:
(i) Understand the basic techniques of performing derivative-based optimization of multi-variable
functions using any of the methods suitable for the problems. Identification of a suitable
methods will be an expected skill to be developed by each student.
(ii) Formulate and solve linear programming problems using a suitable technique as may be
demanded by the problem.
(iii) Learn how to apply other tools even not discussed here as may become relevant in the future.
(iv) Differentiated between constrained and unconstrained optimization problems.
(v) Understand the techniques involved in solving both indirect and direct methods for
unconstrained and constrained optimization respectively.
(vi) Apply Linear Programming in solving Economic dispatch problems of power systems.
(vii) Finally, appreciated the application of evolutionary algorithms in Power systems economic
Dispatch etc.
Introduction: Optimization is a best process of achieving:
(a) maximum utilization/minimum losses
(b) best possible solution
(c) feasible solution
gg (d) most economical solution/method.
Thus, optimization provides a logical method for the selection of the best choice from among all possible
design that are available such that all limitations and restrictions(constraints) placed on it are satisfied.
Simply put, it is a process or rationale to achieve an improved solution. It may apply to a physical design
or an operational procedure.
The approach to this study is numerical and in most cases practical if programmed for computer solutions.
The methods and/ algorithms developed will allow the developer to optimize the selected design concepts
against pre-selected design limitations.
The basic difference between the constrained and the unconstrained problems is the selection of admissible
points that are eligible for optimization. The so-called admissible points are those vectors that satisfy the
 Page 2 of 45

constraints of the problem. An optimum of the constrained problem is judged by the comparison of
admissible points.

In an optimization problem, the objective is to optimize (maximize or minimize) some function 𝑓. This
function 𝑓 is called the objective function. Formulation of an optimization problem involves taking
statements, defining general goals and requirements of a given activity, and transcribing them into a
series of well-defined mathematical statements. Single Objective Optimization is an effective approach
to achieve a “best” solution, where a single objective is maximized or minimized. In comparison,
Multiple Objective Optimization can derive a set of non-dominated optimal solutions that provide
understanding of the trade-offs between conflicting objectives.
For instance, an objective function to be maximized or minimized may be the revenue in a production of
TV sets, the yield per minute in a chemical process, the mileage per liter for a certain type of car, the
hourly number of customers served in a bank, the hardness of steel, or the tensile strength of a rope.
Conversely, one may like to minimize 𝑓 if 𝑓 is the cost per unit of producing certain Android phones, the
operating cost of some power plant, daily loss of heat in a heating system, the cost of gasoline used in
generating stations etc.
In most optimization problems the objective function 𝑓 depends on several variables
𝑥1 , . . . . 𝑥𝑛 (1)
In many problems the choice of values of Eqn. (1) is not entirely free but is subject to some constraints,
i.e., additional restrictions arising from the nature of the problem and the variables. For example, if 𝑥1 is
production cost, then 𝑥1 ≥ 0, and there are many other variables (time, weight, distance travelled by the
salesman, etc.) that can be take nonnegative values only.
Constraints can also have form of equations (instead of inequalities).
Decision Variables:
These are numerical quantities or which values are to be chosen in producing an optimum design. Such
quantities are length, mass, time, displacement, temperature, number etc. it is often represented as a
vector:
𝑋 = (𝑋1 , 𝑋2 , 𝑋3 , . . . 𝑋𝑛 ) 𝑓𝑜𝑟𝑎𝑛 𝑛 − 𝑑𝑖𝑚𝑒𝑛𝑠𝑖𝑜𝑛𝑎𝑙 𝑝𝑟𝑜𝑏𝑙𝑒𝑚.

Objective functions: This is an equation or expression associated with the system that the designer(s)
desire to optimize (maximize or minimize). In vector form, the objective function is represented by
𝐹 = 𝑓(𝑥) = 𝑥1 , 𝑥2 , 𝑥3 , . . . 𝑥𝑛 are the variable systems which varies from one variable system to n
variable system.
f(x)
f(x)

x x2 x1

1-variable system 2-variable system

The selection of the objective function is often the most important decision of the whole design process.
The nature of the objective function often influences the selection of the most efficient optimization
technique. An objective function is a single scalar value that is formulated from a set of design
responses. For example, if the design responses are defined from the strain energy of the nodes in a
region, the objective function could minimize the sum of the design responses; i.e., minimize the sum
of the strain energy, in effect maximizing the stiffness of the region. In optimization, we either
 Page 3 of 45

maximize a function or minimize a function. The minimum and maximum value of the variable are
referred to as the turning point of the system.

Min

Max

𝑀𝑎𝑥 (𝑓(𝑥)) ⟹ 𝑀𝑖𝑛(−𝐹(𝑥))

An optimization problem can be stated as:
𝑚𝑖𝑛(𝛷(𝑈(𝑥), 𝑥)), (2)
where Φ is the objective function that depends on the state variables, U, and the design variables, x.
The formula for the objective function that tries to minimize N design responses can be stated as:

𝜑𝑚𝑖𝑛 = 𝑚𝑖𝑛 (∑ 𝑁𝑊𝑖 (𝜑𝑖 − 𝜑𝑖𝑟𝑒𝑓 )) (3)

𝑖=1
where each design response, 𝜑𝑖 , is given a weight, W𝑖 , and a reference value, 𝜑𝑖𝑟𝑒𝑓 . The formula for
the objective function that tries to maximize N design responses can be stated as:

𝜑𝑚𝑎𝑥 = 𝑚𝑎𝑥 (∑ 𝑁𝑊𝑖 (𝜑𝑖 − 𝜑𝑖𝑟𝑒𝑓 )) (4)

𝑖=1
Constraints
Many problems have constraints. Each constraint limits the set of possible solutions, and together the
constraints define the feasible set X. Feasible design points do not violate any constraints. For example,
consider the following optimization problem:
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑓(𝑥1 , 𝑥2 )
Subject to
𝑥1 ≥ 0
𝑥2 ≥ 0 (5)
𝑥1 + 𝑥2 ≤ 1

The feasible set is plotted in Figure 3.

Constraints are typically written with ≤, ≥, or =. If constraints involve < or
> (i.e., strict inequalities), then the feasible set does not include the constraint Figure 3. The feasible set X
boundary. associated with equation (5).
In other words, the basic optimization problem is:
Minimize f (x)
subject to x ∈ X (6)
Here, x is a design point. A design point can be represented as a vector of values corresponding to
different design variables. An n-dimensional design point is written:
[x1, x2, · · · , xn] (7)
th
where the i design variable is denoted xi. The elements in this vector can be adjusted to minimize the
objective function f. Any value of x from among all points in the feasible set X that minimizes the
objective function is called a solution or minimizer. A particular solution is written x∗.
This formulation is general, meaning that any optimization problem can be
 Page 4 of 45

rewritten according to equation (6). In particular, a problem

𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝑓(𝑥) 𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 𝑥 ∈ 𝜒 (8)
can be replaced by
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 − 𝑓(𝑥) 𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜 𝑥 ∈ 𝜒 (9)
Constraints
As outlined Eqn. (2), an optimization problem can be stated as:
𝑚𝑖𝑛(𝛷(𝑈(𝑥), 𝑥)),
where Φ is the objective function that depends on the state variables, U, and the design variables, x.
Constraints, 𝚿 (𝒑𝒔𝒊), can be applied to the optimization problem, and constraints, 𝑘𝑖 , can be applied
to the design variables:
𝛹𝑖 (𝑈(𝑥), 𝑥) ≤ 0, 𝑘𝑖 (𝑥) ≤ 0,
where 𝛹𝑖(𝑈(𝑥), 𝑥) ≤ 𝛹𝑖∗ and Ψi is the design response that is constrained by the value 𝛹𝑖∗ . In
addition, 𝑘𝑖 (𝑥) ≤ 𝑘𝑖∗ , where 𝑘𝑖 is an expression for the layout of the design variables, such as
manufacturability, and 𝑘𝑖∗ is the constraint on the design variables.
The Optimization module can arrive at a solution that optimizes the objective function; however, if
the constraints are not satisfied, the result of the optimization may not be a feasible design. A
constraint is based on a design response and, similar to a design response, is formulated from a single
scalar value. Most optimizations have constraints that prevent the optimization from arriving at a
trivial solution. For example, if you are trying to maximize the stiffness of a structure,
the Optimization module will simply fill the entire design area if you do not apply any constraints.
However, if you apply a weight constraint that limits the weight to 50% of its initial value,
the Optimization module is forced to seek an optimum solution that both optimizes the stiffness
objective and satisfies the weight constraint. You can apply only volume constraints to both topology
optimization and to shape optimization; you cannot use volume as an objective function. You cannot
apply multiple constraints of the same type, such as volume, to the whole model or to a single region.
There are two types of constraint:
(a) Functional constraint. e.g. 𝐴𝑥 = 𝑏 is a function where
𝑥 = (𝑥1 , 𝑥2 , 𝑥3 , . . . 𝑥𝑛 ) 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒
𝑥≥0
𝑏≥0
𝐴 = 𝑚 × 𝑛 𝑚𝑎𝑡𝑟𝑖𝑥
𝑚 = 𝑛 𝑟𝑜𝑤 𝑚𝑎𝑡𝑟𝑖𝑥
𝑏 = 𝑚 𝑐𝑜𝑙𝑢𝑚𝑛 𝑚𝑎𝑡𝑟𝑖𝑥
𝑎11 𝑥1 + 𝑎12 𝑥2 . . . 𝑎1𝑚 𝑥𝑛 = 𝑏1
[ ; ; ; ]
𝑎𝑚1 𝑥1 𝑎𝑚2 𝑥2 . . . 𝑎𝑚𝑛 𝑥𝑛 = 𝑏𝑚
(b) Regional constraint.
𝑍1 ≤ 𝑟1 (𝑥) ≤ 𝑍𝑛
𝑍𝑘 ≤ 𝑟𝑘 (𝑥) ≤ 𝑍𝑚
The optimum values always lie in the extreme factors i.e., 𝑍𝑚 𝑎𝑛𝑑 𝑍𝑛
Local Optima: This is a point in the designed space that is higher than all other that is higher than all other point
within its immediate vicinity. A designed space may have several local optima.
Global Optima: This is the optimum design within the total allowable design space.
 Page 5 of 45

Figure 4: Strong and weak maxima and minima.

Saddle Point: This appears to be an optimal solution from a view point but is in fact inferior to some other points
in the solution or designed space.

Saddle
pt

Problem Variables: These are the individual elements in a group of independent variable parameters that
uniquely and completely define the design problem being considered. The variables are unknown values to be
solved in the optimization process (e.g., length, temperature, mass, time, current, voltage, resistance, etc.,).

Calculus based techniques

Unconstrained optimization problems consider the problem of minimizing an objective function that depends on
real variables with no restrictions on their values. Mathematically, let 𝑥 ∈ 𝑅 𝑛 be a real vector with n≥1 components
and let 𝑓: 𝑅 𝑛 → 𝑅 be a smooth function. Then, the unconstrained optimization problem is 𝒎𝒊𝒏𝒙 𝒇(𝒙). Where (∈
𝑖𝑠 𝑒𝑙𝑒𝑚𝑒𝑛𝑡 𝑜𝑓). Unconstrained optimization problems arise directly in some applications but they also arise
indirectly from reformulations of constrained optimization problems. Often it is practical to replace the constraints
of an optimization problem with penalized terms in the objective function and to solve the problem as an
unconstrained problem.
Therefore, unconstrained optimization involves finding the maximum or minimum of a differentiable function
of several variables over a nice set. To meet the complexity of the problems, computer algebra system can be used
to perform the necessary calculations. Most of the constrained optimization problems in power system operation
can be converted into unconstrained optimization problems. The major unconstrained optimization approaches that
are used in power system operation are gradient method, line search, Lagrange multiplier method, Newton –
Raphson optimization, trust – region optimization, quasi – Newton method, double dogleg optimization, and
conjugate gradient optimization, etc.

Powell’s Method in Practice

A general minimization procedure is designed so that it will minimize a quadratic function of m-variables in n-step
cycles. Such a method is good for quadratic functions but very slow for other types of function.
In Powell’s method, the function is minimized in the co-ordinate directions by changing one variable at a time. At
the end of the cycle of minimization, a new direction is found and a new cycle started. For multi-dimension system,
 Page 6 of 45

Powell’s method and others like it will find wide application. Methods like incremental search and internal halving
are very suited for one-dimensional system only.
Let the design variable be:

𝑋 = (𝑥1 , 𝑥2 , 𝑥, . . . 𝑥𝑛 ) . Define 𝑆𝑞 , where 𝑆𝑞 is a cyclic ordering of unit vectors:

That is:
𝑆1 = (100 . .0000)𝑡
𝑆2 = (010 . .0000)𝑡
𝑆3 = (001 . .0000)𝑡
.
.
𝑆𝑛 = (000 . .0001)𝑡
𝑆𝑛+1 = 𝑆1
If our starting point is 𝑆𝑞 and we take a step of length 𝛼𝑞 in the direction 𝑆𝑞 , then 𝑋𝑞+1 = 𝑋𝑞 + 𝛼𝑞 𝑆𝑞 ,

where 𝑋𝑞 is the old position and 𝑋𝑞+1 is the new position.

For minimization, 𝐹(𝑋𝑞+1 ) ≤ 𝐹(𝑋𝑞 )

For each variable,

(𝑖) (𝑖) (𝑖)
𝑋1𝑖+1 = 𝑋1 + 𝛼1 𝑆1
(𝑖) (𝑖) (𝑖)
𝑋2𝑖+1 = 𝑋2 + 𝛼2 𝑆2
(𝑖) (𝑖) (𝑖)
𝑋𝑛𝑖+1 = 𝑋𝑛 + 𝛼𝑛 𝑆𝑛
For each co-ordinate change of variable, determine 𝛼 by minimizing,

𝐹(𝑥 (𝑖+1) ) = 𝐹(𝑥𝑗𝑖 + 𝛼 (𝑖) 𝑆 (𝑖) )

Use the value of 𝛼 to determine the new co-ordinate point. At the end of the cycle, the new direction is
given by 𝑆 = 𝑋 − 𝑌, where X is the current (new) co-ordinate point and Y is the old co-ordinate point.
The problem is to actually minimize

𝐹(𝑋 (𝑖) + 𝛼 (𝑖) 𝑆𝑗 ) = 𝐹(𝛼), i.e., find the step size.

Example 1: Minimize 𝐹(𝑥) = 3𝑥12 + 𝑥22 − 12𝑥1 − 8𝑥2 , using Powel’s method. Start with F(0,0).
Solution:
Set Y = X = [0,0]
F(0, 0) = 0
Cycle 1:
Step1:
 Page 7 of 45

Change
(0) (0) (0) (0) (0)
𝑥𝑖1 = 𝑥𝑖 + 𝛼𝑖 𝑆𝑗 = 𝑥1 𝛼1

Since 𝑆1 is a unit vector of the cycle,

(0) (0) (0) (0) (0) 2 (0) 2 (0) (0) (0)
𝐹(𝑥1 + 𝛼1 , 𝑥2 ) = 3(𝑥1 + 𝛼1 ) + (𝑥2 ) − 12(𝑥1 + 𝛼1 ) − 8𝑥2

Substituting the penultimate value x = (0, 0), we have

(0) 2 (0)
𝐹(𝛼1 ) = 3(𝛼1 ) − 12(𝛼1 )
(0) (0)
Solving for 𝛼1∗ : 𝐹 ′ (𝛼1 ) = 6𝛼1 − 12 = 0, 𝑤ℎ𝑖𝑐ℎ 𝑦𝑖𝑒𝑙𝑑𝑠 𝛼1 = 2.
Hence, the new value of 𝑥1 𝑖𝑠 𝑥1 = 0 + 𝛼1∗ = 2
Test if x = (2, 0) minimizes F(x) : 𝐹(2,0) = 12 − (12 × 2) = −12
Therefore, 𝐹(𝑋𝑞+1 ) < 𝐹(𝑋𝑞 ), so we are moving in the right direction.

Step 2:
(1) (0) (0) (0) (0) (0)
Change 𝑥2 𝑡𝑜 𝑥2 + 𝛼2 𝑆2 = 𝑥2 + 𝛼2
(1) (1)
Substituting the new 𝑥1 and the present 𝑥2 in the objective function, we obtain:
(1) (0) (0) (1) 2 (0) (0) 2 (1) (0) (0)
𝐹(𝑥1 , 𝑥2 + 𝛼2 ) = 3(𝑥1 ) + (𝑥2 +𝛼2 ) − 12(𝑥1 ) − 8(𝑥2 + 𝛼2 )

(0) 2 (0) (0) (1) 2 (1)

𝐹(𝛼2 ) = 3(2)2 + (𝛼2 ) − 12(2) − 8(𝑥2 + 𝛼2 ) = (𝛼2 ) − 8𝛼2 − 12
(1)
Therefore, 𝐹 ′ (𝛼2 ) = 2𝛼2 − 8 = 0
(1)
Given 𝛼2 = 4
Test whether 𝑥 = (2,4) 𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒𝑠 𝐹(𝑥)
𝐹(2,4) = 3(2)2 + (4)2 − 12(2) − 8(4) = −28
Hence, 𝐹(𝑋𝑞+1 ) < 𝐹(𝑋𝑞 ), so we are moving in the right direction.

𝑆 = 𝑋 − 𝑌 = (2,4) − (0,0) = (2,4)

When do we stop this iteration?
Test when |𝐹(𝑥)(𝑖 + 1) − 𝐹(𝑥)𝑖| ≤ 𝜀 (𝑤ℎ𝑒𝑟𝑒 𝜀 is a specified tolerance value)
Cycle 2:
Step 1:
(2) (1) (1) (1) (1)
Change 𝑥𝑖 𝑡𝑜 𝑥1 + 𝛼1 = 𝑥1 + 2𝛼1

(1) 2 (1)
Hence, 𝐹(𝑥1 + 2𝛼1 , 𝑥2 ) = 3(𝑥1 + 2𝛼1 ) + 𝑥22 − 12(𝑥1 + 2𝛼1 ) − 8𝑥 2
 Page 8 of 45

(2)
Thus,𝐹 ′ (𝛼1 ) = 6(2 + 2𝛼1 )2 − 24 = 0 which yields 𝛼1 = 0 𝑖. 𝑒. , 𝑥1 = 2
There is no need to test for convergence since there is no change in 𝑥1 .
Step 2:
(2) (1) (1) (1) (1)
Change 𝑥2 𝑡𝑜 𝑥2 + 𝛼2 = 𝑥2 + 4𝛼1

𝐹(𝑥1 , 𝑥2 + 4𝛼2 ) = 3𝑥12 + ( 𝑥2 + 4𝛼2 )2 − 12𝑥1 − 8(𝑥2 + 𝛼2 )

where X = (2, 4).
𝐹(2,4) = 3(22 ) + (𝑥2 + 4𝛼2 )2 − 12(2) − 8(4 + 4𝛼2 )
Therefore, 𝐹 ′ (𝛼2 ) = 2(4 + 4𝛼2 )4 − 8(4) = 0
(2) (2)
This gave 𝛼2 = 0 while 𝑥2 = 4
Therefore, F(X) is minimum at X = (2, 4).
The optimal value of the objective function F(X) = - 28.
Thus, a quadratic function of two variables is minimized in two cycles. In conclusion, for a quadratic
equation or function, the number of variables is equal to the number of cycles needed to get convergence.
Powell showed that, for a quadratic form, k iterations of the above basic procedure produce a set of directions
µ𝑖 whose last k members are mutually conjugate. Therefore, N iterations of the basic procedure, amounting to N (N
+ 1) line minimizations in all, will exactly minimize a quadratic form.

Alternative Solution to the same question.

𝑭(𝒙) = 𝟑𝒙𝟐𝟏 + 𝒙𝟐𝟐 − 𝟏𝟐𝒙𝟏 − 𝟖𝒙𝟐
Given that 𝒙 = (𝟎, 𝟎) 𝒂𝒏𝒅 𝑺 = (𝟏, 𝟏)
Applying this 𝐹(𝑥1 , 𝑥1 ) ⟹ 𝑥1 + 𝑆𝛼 𝑖. 𝑒. , 𝑥1 𝑐ℎ𝑎𝑛𝑔𝑒𝑠 𝑤ℎ𝑖𝑙𝑒 𝑥2 𝑟𝑒𝑚𝑎𝑖𝑛𝑠 𝑐𝑜𝑛𝑠𝑡𝑎𝑛𝑡
Then substitute the values of 𝒙𝟏 into the given equation yields
𝑭(𝒙) = 𝟑(𝑥1 + 𝑆𝛼 )𝟐 + 𝒙𝟐𝟐 − 𝟏𝟐(𝑥1 + 𝑆𝛼 ) − 𝟖𝒙𝟐
Substituting the initial values of 𝑥1 𝑎𝑛𝑑 𝑆
𝑭(𝜶) = 𝟑(𝛼 )𝟐 − 𝟏𝟐𝜶
𝜕𝑓
At min. point 𝜕𝛼 = 0
𝜕𝑓
= 6𝛼 − 12
𝜕𝛼
∴𝜶=𝟐
Substituting, 𝑥1 = 2 𝑎𝑛𝑑 𝑥2 = 0 in the initial equation, 𝑓(0) = 0 𝑎𝑛𝑑 𝑓(2, 0) = 12 + 0 − 24 − 0 =
−12 . with in new value it is approaching minimum point.
Next step is to set 𝑥1 = 2 𝑎𝑛𝑑 𝑥2 = 𝑥2 + 𝑆𝛼, substitute into the main equation gives
𝑭(𝒙) = 𝟑(𝑥1 )𝟐 + (𝑥2 + 𝑆𝛼 )𝟐 − 𝟏𝟐𝑥1 − 𝟖(𝑥2 + 𝑆𝛼 )
𝑭(𝟐, 𝟎) = 𝒇(𝜶) = 𝟏𝟐 + 𝜶𝟐 − 𝟐𝟒 − 𝟖𝜶
𝜕𝑓
= 2𝛼 − 8
𝜕𝛼
∴ 𝜶 = 𝟒 hence the new value for 𝑥2 = 4
Checking the direction from the original equation
𝑭(𝟐, 𝟒) = 𝟏𝟐 + 𝟏𝟔 − 𝟐𝟒 − 𝟑𝟐 = −𝟐𝟖
This showed that we are still on course, going down towards the minimum value.
 Page 9 of 45

Having completed one cycle and a new value of S will be obtained via
𝑆𝑛𝑒𝑤 = 𝑋 − 𝑌
where X = new value for x and Y = old value for x.
𝑆𝑛𝑒𝑤 = (2 − 0; 4 − 0) = (2, 4)
Repeat the cycle with this new value of S
Now 𝑥1 → 𝑥1 + 𝑆𝛼 = 2 + 2𝛼
𝑭(𝒙) = 𝟑(2 + 2𝛼 )𝟐 + 𝟏𝟔 − 𝟏𝟐(2 + 2𝛼 ) − 𝟑𝟐
𝜕𝑓
= 2 × 3(2 + 2𝛼) × 2 − 24
𝜕𝛼
∴𝛼=0
Secondly 𝑥2 = 𝑥2 = 4
Since 𝛼 = 0 it means that 𝑥1 does not need any change and therefore 𝑥1 𝑟𝑒𝑚𝑎𝑖𝑛𝑠 2
Hence, 𝑥2 → 𝑥2 + 𝑆𝛼 = 4 + 4𝛼 𝑤ℎ𝑖𝑙𝑒 𝑥1 = 𝑥1 = 2
𝑭(𝑥) = 3𝑥12 + 𝑥22 − 12𝑥1 − 8𝑥2
𝐹(𝑥) = 12 + (4 + 4𝛼)2 − 24 − 8(4 + 4𝛼)
𝜕𝑓
= 2(4 + 4𝛼) × 4 − 32 = 0
𝜕𝛼
∴𝛼=0
Also, since 𝛼 = 0, then 𝒙𝟐 does not any further change remain at 𝑥2 = 4
∴ Optimum point =F(x) = -28 and the variable that produces this point are 𝑥 = 2, 4.
𝑂𝑝𝑡𝑖𝑚𝑢𝑚 𝑝𝑜𝑖𝑛𝑡 = −28
𝑉𝑎𝑟𝑖𝑎𝑏𝑙𝑒 𝑎𝑟𝑒 𝑥 = 2, 4.

GRADIENT METHODS
Unlike Powell’s method, all the variables are changed simultaneously for one cycle of the gradient
method. There are many types of gradient methods but we are going to treat to treat the steepest
decent/ascent method.
A large number of multi-dimensional optimization algorithms depend in some way in gradient
information.
In general, we have an n-dimensional vector 𝑦 = (𝑥1 , 𝑥2 ,. . . , 𝑥𝑛 ). In order to find the optimum of a
function, f, we consider a vector 𝑚𝑗 such that 𝑚1 is the direction of move or change of 𝑥1 𝑎𝑛𝑑 𝑚2 𝑖𝑠
the direction of change or move of 𝑥2 𝑎𝑛𝑑 𝑚𝑛 is the direction of change or move of 𝑥𝑛 .
𝑚1 is the direction of change or move of 𝑥𝑛
𝑚2 is the direction of change or move of 𝑥2
𝑚3 is the direction of change or move of 𝑥3
𝑚𝑛 is the direction of change or move of 𝑥𝑛
(𝑖)
That is, from position 𝑥𝑗 we move to a new position
(𝑖+𝑗) (𝑖) (𝑖)
𝑥𝑗 = 𝑥𝑗 + 𝑆𝑚𝑗 (1)
where S is the magnitude of the step size and for a 2-dimensional case.
For n-dimensional case (n-decision variables case), we have;
𝑛

𝑑𝑆 = √𝑑𝑥12 + 𝑑𝑥22 +. . . +𝑑𝑥𝑛2 𝑂𝑅 𝑑𝑆 = √∑ 𝑑𝑥𝑗2 (2)

𝑗=1

If we divide both sides by Ds, then it gives

 Page 10 of 45

𝑛 2 𝑛 2
𝑑𝑥𝑗 𝑑𝑥𝑗
1 = √∑ ( ) = ∑ ( ) (3)
𝑑𝑆 𝑑𝑆
𝑗 𝑗

This change of size will cause a change in the objective function f given by:
𝜕𝑓 𝜕𝑓 𝜕𝑓 𝜕𝑓
𝑑𝑓 = 𝑑𝑥1 + 𝑑𝑥2 + 𝑑𝑥3 + . . . + 𝑑𝑥
𝜕𝑥1 𝜕𝑥2 𝜕𝑥3 𝜕𝑥𝑛 𝑛
Dividing by dS throughout we have;
𝑛
𝑑𝑓 𝜕𝑓 𝑑𝑥1 𝜕𝑓 𝑑𝑥2 𝜕𝑓 𝑑𝑥3 𝜕𝑓 𝑑𝑥𝑛 𝜕𝑓 𝑑𝑥𝑗
= + + + . . . + =∑ (4)
𝑑𝑆 𝜕𝑥1 𝑑𝑆 𝜕𝑥2 𝑑𝑆 𝜕𝑥3 𝑑𝑆 𝜕𝑥𝑛 𝑑𝑆 𝜕𝑥𝑗 𝑑𝑆
𝑗=1
𝑑𝑓
A particular setoff displacement 𝑑𝑥𝑗 will make as large or as small as possible. This is the direction
𝜕𝑥𝑗
of the steepest ascent or descent. The problem may be stated as follows:
𝑑𝑓 𝜕𝑓 𝑑𝑥𝑗
Min/Max = ∑𝑛𝑗=1
𝑑𝑆 𝜕𝑥𝑗 𝑑𝑆
𝑑𝑥 2
s.t. ∑𝑛𝑗=1 ( 𝑑𝑆
𝑗
) =1
This is a constrained optimization problem, which must be converted to an unconstrained optimization
problem.

Using the Lagrangian function, we have:

Min/Max:
𝑛 𝑛 2
𝜕𝑓 𝑑𝑥𝑗 𝑑𝑥𝑗
𝐿=∑ − 𝜆 {∑ ( ) − 1}
𝜕𝑥𝑗 𝑑𝑆 𝑑𝑆
𝑗=1 𝑗=1
Solution Steps:
𝑑𝑥𝑗
Step 1: Differentiate L with respect to and set to zero.
𝑑𝑆
𝑑𝐿 𝑑𝑓 𝑑𝑥𝑗
𝑑𝑥𝑗 = − 2𝜆 =0 𝑓𝑜𝑟 𝑗 = 1, 2, 3, . . . , 𝑛 (5)
⁄ 𝑑𝑥𝑗 𝑑𝑆
𝑑𝑆
Step 2:
Differentiate L w.r.t 𝜆 and setto zero, then it gives;
𝑛 2
𝑑𝑥𝑗
∑( ) = 1 (6)
𝑑𝑆
𝑗=1
𝑑𝑥𝑗 1 𝑑𝑓
From Eqn. (5) = 2𝜆 . 𝑑𝑥 (7)
𝑑𝑆 𝑗
Substituting eqn. (7) into (6) gives;
𝑛 2
1 𝑑𝑓
2
∑( ) = 1 (8)
4𝜆 𝑑𝑥𝑗
𝑗=1
2
𝑑𝑓
and 2𝜆 ± √∑𝑛𝑗=1 (𝑑𝑥 ) =1 (9)
𝑗

Note that from eqn. (8) the change in xj as we make a move is given by:
𝑑𝑥𝑗 1 𝑑𝑓
= 2𝜆 . 𝑑𝑥 𝑗 = 1, 2, . . . 𝑛 and that our new position is:
𝑑𝑆 𝑗
𝑑𝑥𝑗 (𝑖)
𝑥𝑗𝑖+1 = 𝑥𝑗𝑖 , 𝑆 = 𝑥𝑗 + 𝑚𝑗
𝑑𝑆
 Page 11 of 45

𝑑𝑓
𝑑𝑥𝑗 1 𝑑𝑓 𝑑𝑥𝑗
where 𝑚𝑗 = = 2𝜆 . 𝑑𝑥 = ± (10)
𝑑𝑆 𝑗 2
𝑑𝑓
√∑𝑛
𝑗=1(𝑑𝑥 )
𝑗

The positive sign increases f and therefore it is ascent while the negative sign decreases f and is for descent.
The numerator is the gradient and the denominator is the normalizing factor. To find the step S, substitute
𝑥𝑗𝑖+1 in f and minimize w.r.t. S in order to find the optimum step size. We can test if the optimum obtained
is a minimum or maximum by checking the Hessian matrix. For minimum, H is positive definite at x.
test for convergence |𝑓(𝑥𝑗𝑖+1 ) − 𝑓(𝑥 𝑖 )| ≤ 𝜀

Steepest Descent/Ascent method

Given F(x), choose initial values of 𝑥 (0) = 𝑥10 , 𝑥20 , . . . , 𝑥 0𝑛 , i = with these values we calculate the
gradient given by;
𝜕𝑓
± ⁄𝜕𝑥
𝑗
𝐺𝑟𝑎𝑑𝑖𝑒𝑛𝑡, 𝑀𝑗 = 2
𝜕𝑓
√∑𝑛
𝑗=1( )
𝜕𝑥𝑗

Next, we look for value of 𝑥 𝑖. 𝑒. , 𝑥 ′ and given by 𝑥 ′ = 𝑥 0 + 𝑀0 𝑆 which is similar to 𝑥 𝑖+1 = 𝑥 2 + 𝛼 2 𝑆 𝑖 .

𝜕𝑓(𝑠)
We put the value of 𝑥 ′ 𝑖𝑛𝑡𝑜 the function to get the function dependent on S, we get 𝜕𝑆 = 0.
Summary of the solution steps
1. Let the decision variable be 𝑥 = 𝑥1 , 𝑥2 , . . . 𝑥𝑛
2. Starting point 𝑥 0 = 𝑥10 , 𝑥20 , . . . , 𝑥𝑛0
𝛿𝑓
3. Calculate the gradient 𝛿𝑥 ∀ 𝑗 = 1, 2, 3, . . . , 𝑛 𝑁𝐵: ∀ 𝑚𝑒𝑎𝑛𝑠 𝑓𝑜𝑟 𝑎𝑙𝑙
𝑗
𝜕𝑓
± ⁄𝜕𝑥
𝑗
4. Define 𝐺𝑟𝑎𝑑𝑖𝑒𝑛𝑡, 𝑀𝑗 = 2
𝜕𝑓
√∑𝑛
𝑗=1( )
𝜕𝑥𝑗

5. Set new values of x as 𝑥 𝑖+1 = 𝑥 𝑖 + 𝑚(𝑖) 𝑆, 𝑤ℎ𝑒𝑟𝑒 𝑚 = [𝑚1 , 𝑚2 , . . ., 𝑚𝑛 ]

6. Substitute 𝑥 𝑖+1 in the objective function, thus yielding a function F(s).
7. Determine the value of S that optimize F(s).
8. Calculate the new value of 𝑥 𝑖+1
9. Test for convergence |𝑓(𝑥 𝑖+1 ) − 𝑓(𝑥 𝑖 )| ≤ 𝜀

Example; 𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑭(𝑥) = 3𝑥12 + 𝑥22 − 12𝑥1 − 8𝑥2

Starting at 𝑥 0 = (0, 0), chose 𝜀 = 0.1.

Solution
𝜕𝑓
± ⁄𝜕𝑥
𝑗
Applying 𝑀𝑗 = 2
𝜕𝑓
√∑𝑛
𝑗=1( )
𝜕𝑥𝑗
𝜕𝑓 𝜕𝑓
= 6𝑥1 − 12 and = 2𝑥2 − 8
𝜕𝑥1 𝜕𝑥2
Substituting the initial conditions of 𝑥 0 = (0, 0)
𝜕𝑓 𝜕𝑓
| = −12 𝑎𝑛𝑑 | = −8
𝜕𝑥1 𝑥 0 𝜕𝑥2 𝑥 0
1 2
−12
𝑇ℎ𝑒𝑛, 𝑀10 = = −0.8321
√(−12)2 +(−8)2
 Page 12 of 45

−8
𝑀20 = = −0.5547
√(−12)2 +(−8)2
Step 5: New values are:
Thus, going to our equation
(1) (0)
𝑥1 = 𝑥1 + 𝑚1 𝑆 and
(1) (0)
𝑥2 = 𝑥2 + 𝑚2 𝑆
(1)
∴ 𝑥1 = 0 − 0.8321𝑆 = −0.8321𝑆
(1)
𝑥2 = 0 − 0.5547𝑆 = −0.5547𝑆
Step 6:
Substituting these values into the objective function to obtain;
Thus, 𝐹(𝑆) = 3(−0.8321𝑆)2 + (−0.5547𝑆)2 − 12(−0.8321𝑆) − 8(−0.5547𝑆)
𝜕𝑓(𝑠)
= 0;
𝜕𝑆
𝜕𝑓(𝑠)
= 6(−0.8321𝑆)(−0.8321) + 2(−0.5547𝑆)(−0.5547) + 12(0.8321) + 8(0.5547
𝜕𝑆
= 2 × 2.3849𝑆 + 14.4728 = 0
−14.4728
𝑆= = −3.0238
4.47698
Hence, substituting the value of S = -3.0238 into 𝑥1′ 𝑎𝑛𝑑 𝑥2′
(1)
𝑥1 = −0.8321 × −3.0238 = 2.5161
(1)
𝑥2 = −0.5547 × −3.0238 = 1.6773
⟹ 𝑓(2.5161, 1.6773) = −21.806
Thus, from our result we are minimizing because got a negative value which is far smaller than zero
which was the initial value we started with;
Iteration 2: Then find
𝜕𝑓
± ⁄𝜕𝑥 ⌋ ′
(1) 1 𝑥
1
𝑀1 = 2 2
𝜕𝑓 𝜕𝑓
√( ⌋ ) +( ⌋ )
𝜕𝑥1 𝑥′ 𝜕𝑥2 𝑥′
1 2

𝜕𝑓 𝜕𝑓
= 6𝑥1 − 12 and = 2𝑥2 − 8
𝜕𝑥1 𝜕𝑥2

𝜕𝑓
Hence, 𝜕𝑥 ⌋ (1)
= 6 × 2. 5161 − 12 = 3. 0966
1 𝑥1
𝜕𝑓
⌋ = 2 × 1.6773 − 8 = −4.6454
𝜕𝑥2 𝑥 (1)
2
(1) ±3.0966 3.0966
𝑀1 = = 5.5829 = 0.5547
√(3.0966)2 +(−4.6454)2
(1) −4.6454 −4.6454
𝑀2 = = = −0.8321
√(3.0966)2 +(−4.6454)2 5.5829
(1)
𝑥12 = 𝑥11 + 𝑀1 𝑆 = 2.5161 + 0.5547𝑆
(1)
𝑥22 = 𝑥21 + 𝑀2 𝑆 = 1.6773 − 0.8321𝑆
To find the value of S, plug back into the original equation and differentiate and equate to zero.
∴ 𝐹(𝑆) = 3(2.5161 + 0.5547𝑆)2 + (1.6773 − 0.8321𝑆)2 − 12(2.5161 + 0.5547𝑆)
− 8(1.6773 − 0.8321𝑆)
𝜕𝑓(𝑠)
= 0;
𝜕𝑆
 Page 13 of 45

6(2.5161 + 0.5547𝑆)(0.5547) + 2(1.6773 − 0.8321𝑆)(−0.8321) + 12(2.5161 +

0.5547𝑆)(0.5547) + 8(0.8321)=0

∴ 𝑆 = −1.728
𝑥12 = 𝑥11 + 𝑀1′ 𝑆 = 2.5161 + 0.5547 × −1.728 = 1.5576
𝑥2′ = 𝑥21 + 𝑀2′ 𝑆 = 1.6773 − 0.8321 × −1.728 = 3.115
𝐻𝑒𝑛𝑐𝑒, 𝑓(1.5576, 3.115) = −26.63
We are still going towards the minimum point since −26.63 < −21.806
Checking, −21.806 − (−26.63) ≤ 𝐸𝑟𝑟𝑜𝑟 𝑚𝑎𝑟𝑔𝑖𝑛
4.824 > 𝐸𝑟𝑟𝑜𝑟 𝑚𝑎𝑟𝑔𝑖𝑛 𝜀 ≤ 0.0001

Step 3: Iteration continues

𝜕𝑓 𝜕𝑓
= 6𝑥1 − 12 and = 2𝑥2 − 8
𝜕𝑥1 𝜕𝑥2

𝜕𝑓
Hence, 𝜕𝑥 ⌋ = 6 × 1.5576 − 12 = −2.6544
1 𝑥12
𝜕𝑓
⌋ = 2 × 3.115 − 8 = −1.77
𝜕𝑥2 𝑥 2
2
𝜕𝑓
± ⁄𝜕𝑥 ⌋ 3
1 𝑥
𝑀12 = 2
1
2
𝜕𝑓 𝜕𝑓
√( ⌋ ) +( ⌋ )
𝜕𝑥1 𝑥3 𝜕𝑥2 𝑥3
1 2
−2.6544 −2.6544
𝑀12 = √−2.65442 = = −0.832
+(−4.6454)2 3.1904
−1.77 −1.77
𝑀22 = = = −0.5548
√(3.0966)2 +(−1.77)2 3.1904
𝑥13 = 2 2
𝑥1 + 𝑀1 𝑆 = 1.5576 + 0.832𝑆
𝑥22 = 𝑥21 + 𝑀2′ 𝑆 = 3.115 − 0.5548𝑆
∴ 𝐹(𝑆) = 3(1.5576 + 0.832𝑆 )2 + (3.115 − 0.5548𝑆)2 − 12(1.5576 + 0.832𝑆 ) − 8(3.115
− 0.5548𝑆)
𝜕𝑓(𝑠)
= 0; 6(1.5576 − 0.832𝑆 )(−0.832) + 2(3.115 − 0.5548𝑆 )(−0.5548) − 12(−0.832)
𝜕𝑆
− 8(−0.5548) = 0
−3.1905
∴𝑆= = −0.669
4.7689
𝑥13 = 𝑥11 + 𝑀1′ 𝑆 = 1.5576 − 0.832(−0.669) = 2.1142
𝑥23 = 𝑥21 + 𝑀2′ 𝑆3.115 − 0.5548𝑆(−0.669) = 3.4862
𝐻𝑒𝑛𝑐𝑒, 𝑓(1.5576, 3.115) = −27.904
Continue the iteration till the error is less than 0.0001
Constrained Minimization Problems: Indirect methods by unconstrained minimization
Constrained Optimization problems are problems for which a function is to be minimized or maximized
subject to constrains. In other hand, it is a set of methods designed to identify efficiently and
systematically the best solution (the optimal solution) to a problem characterized by a number of
potential solutions in the presence of identified constraints.
 Page 14 of 45

In a real-life situation, every system is normally constrained. Constrain in itself is a limitation in

which any design procedures were based.
The general problem we considered here can be described as:
min{𝑓(𝑥)}
𝑥
Subject to
𝑔𝑖 (𝑥) ≤ 0, 𝑖 = 1, 2, . .. , 𝑚
ℎ𝑗 (𝑥) = 0, 𝑗 = 1, 2, . .. , 𝑝
Where X is the n dimensional vector, 𝑋 = (𝑥1 , 𝑥2 , . . . , 𝑥𝑛 ); 𝑓(𝑥) is the objective function; 𝑔𝑖 (𝑥) is
the inequality constraint; and ℎ𝑗 (𝑥) is the equality constraint.
Denote the whole search space as S and the feasible space as ℱ, ℱ ⊂ S. (⊂ 𝑖𝑠 𝑘𝑛𝑜𝑤𝑛 𝑎𝑠
𝑠𝑢𝑏𝑒𝑠𝑡) It is important to note that the global in ℱ might not be the same as that in S.

Linear constraints are relatively easy to deal with. Nonlinear equality constraints can be hard to handle.

 The penalty function approach converts a constrained problem into an unconstrained one by
introducing a penalty function into the objective function.
 The repair approach maps (repairs) an infeasible solution into a feasible one.
 The purist approach rejects all infeasible solutions in search.
 The separatist approach considers the objective function and constraints separately.
 The hybrid approach mixes two or more different constraint handling techniques.

Classification of the methods

Indirect methods: The constrained problem is converted into a sequence of unconstrained problems
whose solutions will approach to the solution of the constrained problem, the intermediate solutions
need not to be feasible

Direct methods: The constraints are taking into account explicitly; intermediate solutions are feasible

Transforming the Optimization problem

 Constraints of the problem can be transformed if need be.
𝑔𝑖 (𝑥) ≤ 0 ⇔ 𝑔𝑖 (𝑥) + 𝑦𝑖2 = 0
 Page 15 of 45

Where 𝑦𝑖 is a slack variable; constraint is active if 𝑦𝑖 = 0

 By adding 𝑦𝑖2 no need to add , 𝑦𝑖 ≥ 0
 If 𝑔𝑖 (𝑥) is linear, then linearity is preserved by 𝑔𝑖 (𝑥) + 𝑦𝑖 = 0, 𝑦𝑖 ≥ 0
 𝑔𝑖 (𝑥) ≥ 0 ⇔ − 𝑔𝑖 (𝑥) ≤ 0
 ℎ𝑗 (𝑥) = 0 ⇔ ℎ𝑗 (𝑥) ≤ 0 & − ℎ𝑗 (𝑥) ≤ 0

Examples of indirect methods

 Penalty function methods – for inequality constraints
 Lagrangian methods – for equality constraints

EQUALITY CONSTRAINED PROBLEMS

Lagrangian multipliers
This method converts equality constrained optimization problem to unconstrained problem.
Problem formulation: Minimizing or maximizing F(x) , where x = (x1, x2, . . . xn) subject to C equality
constraints of the form gi(x) = 0, for I – 1, 2, , . . . C. it is possible sometimes to give gi(x) = 𝛼𝑖 , where
𝛼𝑖 ≠ 0
𝐦𝐢𝐧 𝒇(𝒙) =
𝒙
Subject to
𝒈(𝒙) = 𝑪
In this case we rewrite the constraint as:
𝐿𝑖 (𝑥) = 𝑔𝑖 (𝑥) − 𝛼𝑖
Define a non-negative multiplier (constants) 𝜆𝑖 , such that the constrained problem gets converted to an
unconstrained problem
To handle this, one has to introduce a 𝜆 𝑖. 𝑒. , 𝑙𝑎𝑔𝑟𝑎𝑛𝑔𝑒 𝑚𝑢𝑙𝑡𝑖𝑝𝑙𝑖𝑒𝑟; 𝐿 = 𝑙𝑎𝑔𝑟𝑎𝑛𝑔𝑒 𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛
Which is an unconstrained function.
Hence,
𝑛

𝐿(𝑥, 𝜆) = 𝐹(𝑥) − ∑ 𝜆𝑖 (𝑔𝑥𝑖 )

𝑖=1

Once this Lagrange function is obtained you use the classical calculus to differentiate to get the answers
needed.

Exercise 1: Maximize 𝐹(𝑋) = 3𝑥12 + 𝑥22 + 2𝑥1 𝑥2 + 6𝑥1 + 2𝑥2

subject to 2𝑥1 − 𝑥2 = 4

Solution
Applying the Lagrange multiplier
𝐿 = 3𝑥12 + 𝑥22 + 2𝑥1 𝑥2 + 6𝑥1 + 2𝑥2 − 𝜆(2𝑥1 − 𝑥2 − 4)
Start the differentiation of L w.r.t various x
𝜕𝐿
= 6𝑥1 + 2𝑥2 + 6 − 2𝜆 = 0 (i)
𝜕𝑥 1
𝜕𝐿
= 2𝑥2 + 2𝑥1 + 2 + 𝜆 = 0 (ii)
𝜕𝑥2
𝜕𝐿
𝜕𝜆
= −2𝑥1 + 𝑥2 + 4 =0 (iii)

From Eqn.(iii)
𝑥2 = 2𝑥1 − 4 (iv)
Hence, substitute 𝑥2 into Eqns. (i) & (ii)
 Page 16 of 45

6𝑥1 + 2(2𝑥1 − 4) + 6 − 2𝜆 = 0
∴ 6𝑥1 + 4𝑥1 − 8 + 6 − 2𝜆 = 0
10𝑥1 − 2 − 2𝜆 = 0 (v)

2(2𝑥1 − 4) + 2𝑥1 + 2 + 𝜆 = 0
4𝑥1 − 8 + 2𝑥1 + 2 + 𝜆 =0
6𝑥1 − 6 + 𝜆 = 0 (vi)

From Eqn. (v) we observed that

𝜆 = 5𝑥1 − 1 (vii)
Substitute Eqn.(vii) into (vi)
6𝑥1 − 6 + 5𝑥1 − 1 = 0
11𝑥1 − 7 = 0

𝟕
𝒙𝟏 =
𝟏𝟏
Therefore, substituting 𝑥1 into Eqn. (iv)
𝑥2 = 2𝑥1 − 4
7
∴ 𝑥2 = 2( )−4
11
𝟑𝟎
𝒙𝟐 = −
𝟏𝟏
Solving for 𝜆 gives
𝜆 = 5𝑥1 − 1
7
=5× −1
11
𝟐𝟒
∴𝝀=
𝟏𝟏
𝟕 𝟑𝟎
Therefore, the minimum of 𝑭(𝒙) = 𝑭 (𝟏𝟏 , − 𝟏𝟏) = 𝟐𝟏. 𝟒

INEQUALITY CONSTRAINED PROBLEMS

max 𝐹(𝑥) 𝑜𝑟 min 𝐹(𝑥)
𝑥 𝑥
subject to
𝑔(𝑥) ≤ 0
≥ 0
This can be solved using transformation techniques which is based on trial and error. Also, we can use
penalty function which has many ways of approaching it but we may divide it into two processes
exterior and interior. In this case, we are to introduce some functions as shown,
max 𝐹(𝑥) 𝑜𝑟 min 𝐹(𝑥)
𝑥 𝑥
Subject to 𝐺(𝑥) ≥ 0
∞

𝑀(𝑥) = 𝐹(𝑥) + ∑ ∅[𝐺(𝑥)]

𝑗=1
What determine which function one has to use depends on ∅(𝐺; (𝑥)) i.e., either exterior or interior.
For interior penalty function problems
 Page 17 of 45

𝑁
1
∑ ∅(𝐺(𝑥)) = 𝑟𝑝 ∑
𝐺; (𝑥)
𝑗=1
For exterior penalty function problems
𝑁

∑ ∅(𝐺; (𝑥)) = 𝑟𝑝 ∑⟨𝐺;(𝑥)⟩2𝑗𝑗 ⟨𝐺⟩ = 𝐺 𝑓𝑜𝑟 𝐺 ≥ 0

𝑗=1
= 0 𝑓𝑜𝑟 𝐺 < 0 (𝑛𝑜𝑡 𝑢𝑠𝑒)
Where 𝑟𝑝 is a penalty function which is positive constant. NB: when G is negative one is not supposed to
use exterior method.
Penalty Function Algorithms
(i) Exterior penalty functions
Starting from any initial design x with a moderate value of 𝑟𝑝 , find a vector 𝑥𝑚 that minimizes
Step1: M(x) = F
Step 2: Investigate the constraints to determine whether the point 𝑥𝑚 is in the feasible domain
Step 3: If the result of step 2 is true then terminate, otherwise pick 𝑟𝑝+1 > 𝑟𝑝 , hence go back to step 1.

Example 2:
min 𝐹(𝑥) = 10𝑥12 + 𝑥22
𝑥
Subject to 𝑥1 + 𝑥2 ≥ 4 ,
let C= 0.1
start 𝑥𝑜 = (4,4).

Solution
Substitute the values of initial guess of 𝑥𝑜 (4,4) into the main function
𝐹(𝑥𝑜 ) = 10(4)2 + 42 = 176
𝑀(𝑥) = 10𝑥12 + 𝑥22 + 𝑟𝑝 (𝑥1 + 𝑥2 − 4)2 − for exterior case
To determine the initial value of 𝑟𝑝 , equate to the value of F(𝑥𝑜 ).
(0) (0)
∴ 𝐹(𝑥𝑜 ) = 𝑟𝑝 (𝑥1 + 𝑥2 − 4)2 = 176
𝑟𝑝 (4 + 4 − 4)2 = 176
𝟏𝟕𝟔
𝒓𝒑 = = 𝟏𝟏
𝟏𝟔

Direct method for Constrained Optimization Problems

(i) Lagrange multipliers
Given the objective function
𝑀𝑖𝑛 𝑓(𝑥1 , 𝑥2 ,. . . , 𝑥𝑛 )
Subject to equality constraint
𝑔𝑖 (𝑥1 , 𝑥2 ,. . . , 𝑥𝑛 ) = 0, 𝑖 = 1, 2, . . . , 𝑘
Use the Lagrange multiplier method to convert the constrained equality problem to unconstrained
problem by augmenting the cost function by introducing the k-vector 𝜆 of undetermined quantities.
The unconstrained function becomes
𝑘

𝐿 = 𝑓(𝑥) + ∑ 𝜆𝑖 𝑔𝑖
𝑖=1
The resulting conditions for constrained local minima of ℓ are the following
𝑘
𝜕𝐿 𝜕𝑓 𝜕𝑔𝑖
= + ∑ 𝜆𝑖 =0 (1)
𝜕𝑥𝑖 𝜕𝑥𝑖 𝜕𝑥𝑖
𝑖=1
 Page 18 of 45

𝜕𝐿
= 𝑔𝑖 = 0 (2)
𝜕𝑥𝑖
Then solving Eqns. (1) & (2) gives the parameter values for the minimum cost.

Example 3.
Apply the Lagrange method to solve constrained optimization to determine the minimum distance from
the origin of the 𝑥𝑦 plane to a circle described by
(𝑥 − 8)2 + (𝑦 − 6)2 = 25
The minimum distance is obtained by minimization of the distance square given by
𝑓(𝑥, 𝑦) = 𝑥 2 + 𝑦 2

Solution
The function can be rewritten in this form
𝑀𝑖𝑛 𝑓(𝑥, 𝑦)
s.t 𝑔𝑖 (𝑥) = (𝑥 − 8)2 + (𝑦 − 6)2 = 25
So,
ℓ = 𝑥 2 + 𝑦 2 + 𝜆{(𝑥 − 8)2 + (𝑦 − 6)2 − 25}
The necessary condition for extrema then becomes
𝜕𝐿
= 2𝑥 + 𝜆(2𝑥 − 16) = 0
𝜕𝑥
This equation gives
2𝑥(𝜆 + 1) = 16𝜆

𝜕𝐿
= 2𝑦 + 𝜆(2𝑦 − 12) = 0
𝜕𝑦
2𝑦(𝜆 + 1) = 12𝜆
Finally,
𝜕𝐿
= (𝑥 − 8)2 + (𝑦 − 6)2 − 25 = 0
𝜕𝜆
Solving the above three equations simultaneously gives
x = 4 and x = 12
y = 3 and y = 9
At 𝜆 = 1, (𝑥, 𝑦) = 4, 3
𝜆 = −3, (𝑥, 𝑦) = 12, 9
The minimum distance then becomes (𝑥, 𝑦) = 4, 3

For Inequality constraints of constrained optimization

Situations do arise when the optimization problem may contain both equality and inequality constraint
such as:
𝑀𝑖𝑛 𝑓(𝑥1 , 𝑥2 ,. . . , 𝑥𝑛 )
Subject to equality constraint
𝑔𝑖 (𝑥1 , 𝑥2 ,. . . , 𝑥𝑛 ) = 0, 𝑖 = 1, 2, . . . , 𝑘
and inequality constraint
𝑈𝑖 (𝑥1 , 𝑥2 ,. . . , 𝑥𝑛 ) ≤ 0, 𝑖 = 1, 2, . . . , 𝑚
To convert this problem to unconstrained one, we use extended Lagrange multiplier method or the Kuhn
Tucker conditions to include the inequality constraints by introducing M – vector µ of underdetermined
quantities. The unconstrained cost function becomes
𝑘 𝑚

𝐿 = 𝑓 + ∑ 𝜆𝑖 𝑔𝑖 + ∑ 𝜇𝑗 𝑈𝑗
𝑖=1 𝑗=1
 Page 19 of 45

the resulting necessary conditions becomes as usual, the necessary conditions for constrained local
minima of L are the following;
𝜕𝐿
= 0, 𝑖 = 1, . . ., 𝑛
𝜕𝑥𝑖
𝜕𝐿
= 𝑔𝑖 , 𝑖 = 1, . . ., 𝑘
𝜕𝜆𝑖
𝜕𝐿
= 𝑈𝑗 , 𝑗 = 1, . . ., 𝑚
𝜕𝜇𝑗

𝜇𝑗 𝑈𝑗 = 0 𝑎𝑛𝑑 𝜇𝑗 > 0 , 𝑗 = 1, . . . , 𝑚
Example 4:
Find the minimum value of the function 𝑓(𝑥, 𝑦) = 𝑥 2 + 𝑦 2
Subject to 𝑔(𝑥, 𝑦) = 𝑥 2 − 5𝑥 − 𝑦 2 + 20 = 0
𝑈(𝑥, 𝑦) = 2𝑥 + 𝑦 ≥ 6

Solution
𝑘 𝑚

ℓ = 𝑓 + ∑ 𝜆𝑖 𝑔𝑖 + ∑ 𝜇𝑗 𝑈𝑗
𝑖=1 𝑗=1
substituting in the above equation gives
ℓ = x 2 + y 2 + λ{(𝑥 2 − 5x − 𝑦 2 + 20) + μ(2x + y − 6)}
𝜕ℓ
= 2𝑥 + λ(2𝑥 − 5) + 2𝜇 = 0 (1)
𝜕𝑥
𝜕ℓ
= 2𝑦 + λ(−2𝑦) + 𝜇 = 0
𝜕𝑦
= 2𝑦 − 2𝑦λ + 𝜇 = 0 (2)
𝜕ℓ
= 𝑥 2 − 5x − 𝑦 2 + 20 = 0 (3)
𝜕λ
𝜕ℓ
= 2x + y − 6 = 0 (4)
𝜕μ
Eliminating μ from Eqns. (1) & (2)
2𝑥 + λ(2𝑥 − 5) + 2𝜇 = 0 × 1
2𝑦 − 2𝑦λ + 𝜇 =0 × 2
2𝑥 + λ(2𝑥 − 5) + 2𝜇 = 0
4𝑦 − 4𝑦λ + 2𝜇 = 0
-
2𝑥 − 4𝑦 + λ(2𝑥 − 5) − −4𝑦λ = 0
(2𝑥 − 4𝑦) + λ(2𝑥 + 4𝑦) − 5λ = 0 (5)
From eqn. (4)
𝑦 = 6 − 2𝑥 (6)
Substituting the value of y in Eqn. (5)
[2𝑥 − 4(6 − 2𝑥 )] + λ[2𝑥 + 4(6 − 2𝑥)] − 5λ = 0
Expanding the above equation gives
10𝑥 − 6𝑥λ + 19λ − 24 = 0
(10 − 6λ)𝑥 = 24 − 19λ
24 − 19λ
∴𝑥=
10 − 6λ
Substituting x into Eqn. (6)
2(24−19λ) 60−36λ−48+38λ 𝟏𝟐+𝟐𝛌
𝑦 = 6 − 10−6λ = =
10−6λ 𝟏𝟎−𝟔𝛌
Substituting x and y into (3)
24−19λ 2 24−19λ 12+2λ 2
( 10−6λ ) − 5 ( 10−6λ ) − (10−6λ) + 20 = 0
 Page 20 of 45

[576−9+2λ+361λ2 ] [120−95λ] [144+48λ+4λ2 ]

− + + 20 = 0
(10−6λ)2 10−6λ (10−6λ)2
{(361λ2 − 912λ + 576)} − [(120 − 95λ)(10 − 6λ)] + [4λ2 + 48λ + 144] + 20[(10 − 6λ)(10 − 6λ)]
=0
(10 − 6λ)2
507λ2 −1690λ+1232
(10−6λ)2
= 0
Cross multiplying, we obtain
507λ2 − 1690λ + 1232 = 0
−𝑏±√𝑏 2 −4𝑎𝑐
Applying this almighty formular of λ = 2𝑎
1690±598 2288 1092
λ = 1014 = 1014 𝑜𝑟 = 𝟐. 𝟐𝟔 𝒐𝒓 𝟏. 𝟎𝟖
1014
∴ 𝑾𝒉𝒆𝒏 λ = 2.26
12+2(2.26)
𝑦 = 10−6(2.26) = −4.64
24−19(2.26)
𝑥= = 5.32
10−6(2.26)
When 𝛌 = 𝟏. 𝟎𝟖
12+2(1.08)
𝑦 = 10−6(1.08) = 4.02
24−19(1.08)
𝑥= = 0.99
10−6(1.08)
From Eqn. (2)
When 𝑥 = 5.32 & 𝑦 = −4.64, 𝑡ℎ𝑒𝑛 λ = 2.26
𝜇 = 2𝑦λ − 2y = (2 × −4.64 × 2.26) − 2(−4.64)
𝜇 = −20.97 + 9.28
∴ 𝝁 = −𝟏𝟏. 𝟔𝟗
𝑩𝒖𝒕 𝑾𝒉𝒆𝒏 𝒙 = 𝟎. 𝟗𝟗; 𝒚 = 𝟒. 𝟎𝟐 & λ = 1.08
𝜇 = (2 × 4.02 × 1.08) − 2(−4.64)
∴ 𝝁 = 𝟎. 𝟔𝟒.
Therefore, solutions are as follows
(𝒙, 𝒚) = (𝟓. 𝟑𝟐, −𝟒. 𝟔𝟒)
For λ = 2.26 , μ = −11.69
(𝒙, 𝒚) = (𝟎. 𝟗𝟗, 𝟒. 𝟎𝟐)
For λ = 1.08 , μ = 0.64

Exercises
No 1:
Find the optimal solution to the problem
𝑀𝑖𝑛 𝑓(𝑥) = 𝑥12 + 2𝑥22 + 10𝑥32
Subject to
𝑔1 (𝑥) = 𝑥1 + 𝑥22 + 𝑥3 − 5 = 0
𝑔2 (𝑥) = 𝑥1 + 5𝑥2 + 𝑥3 − 7 = 0
Suppose that 𝑔1 (𝑥) = 0.01 & 𝑔2 (𝑥) = 0.02. Find the corresponding change in the optimal value of f(x)

No2: An electrical company manufactures two pole models, each on separate production line. The daily
capacity of the first line is 60 poles and that of the second is 75 poles. Each unit of the first model uses
10 pieces of materials, whereas each unit of the second model requires 8 pieces of the same materials.
The maximum daily availability of the special material is 800 pieces. The profit per unit of models 1 and
2 is N30 and N20, respectively. Determine the optimum daily production of each model.

No 3: Find the minimum value of the function

min 𝑓(𝑥, 𝑦) = 𝑥 2 + 𝑦 2
 Page 21 of 45

Subject to
𝑔(𝑥, 𝑦) = (𝑥 − 8)2 + (𝑦 − 6)2 = 25
𝜇(𝑥, 𝑦) = 2𝑥 + 𝑦 ≥ 12

Linear programming (LP) and Non-linear programming (NLP).

There are two basic ways in which the LP problems are handled:
(a) The graphical method (when the problem involves only two variables).
(b) The algebraic or simplex method (when there are many variables).

The Standard form of the LP model

In order to develop a general solution method, the LP problem must be in a common format, called the
standard form. The properties of the standard LP form are:
 All the constraints are equations with nonnegative in right hand side
 All variables are nonnegative
 The objective function may be maximization or minimization
The followings are the ways to convert to standard form
CONSTRAINTS
1. A constraint of the type ≤ 𝑜𝑟 ≥ can be converted to an equation by adding a slack variable to
(subtracting a surplus variable from) the left-hand side of the constraint.
For example, if the constraint is 𝑥1 + 2𝑥2 ≤ 6
Add a slack variable 𝑆1 ≥ 0 to the left side to obtain the equation
𝑥1 + 2𝑥2 + 𝑆1 = 6, 𝑆1 > 0
Note that if the constraint represents a limit on the usage of a resources, S represents the slack or unused
amount of the resources.
Consider also the constraint
3𝑥1 + 2𝑥2 − 3𝑥3 ≥ 5
Since the LHS is not smaller than the RHS, we subtract. That should give
3𝑥1 + 2𝑥2 − 3𝑥3 − 𝑆2 = 5, 𝑆2 ≥ 0
2. The RHS of an equation can always be made nonnegative by multiplying both sides by -1.
Example: 2𝑥1 + 3𝑥2 − 7𝑥3 = −5 is mathematically equivalent to −2𝑥1 − 3𝑥2 + 7𝑥3 = 5
3. The direction of an inequality is reversed when both sides were multiplied by -1.
Example: 2 < 4 𝑏𝑢𝑡 − 2 > −4, Similarly, 2𝑥1 − 𝑥2 ≤ −5 𝑖𝑠 𝑡ℎ𝑒 𝑠𝑎𝑚𝑒 𝑎𝑠 −2𝑥1 + 𝑥2 ≥ 5

VARIABLES
An unrestricted variable 𝑦1 can be expressed in terms of two nonnegative variables by using
substitution;
𝑦1 = 𝑦1′ − 𝑦1′′ , 𝑦1′ 𝑦1′′ ≥ 0.
The substitution must be effected in both the constraints and the objective function.

OBJECTIVE FUNCTION
The objective function is usually maximization or minimization of functions but one can always convert
a maximization problem to minimization and vice versa.
Example: 𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝑧 = 5𝑥1 + 2𝑥2 + 3𝑥3 , is mathematically equivalent to 𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 (−𝑧) =
−5𝑥1 − 2𝑥2 − 3𝑥3 .
Equivalent meaning that for the same set of constraints, that the optimum values of 𝑥1 , 𝑥2, & 𝑥3 𝑎𝑟𝑒 𝑡ℎ𝑒
𝑠𝑎𝑚𝑒 𝑖𝑛 𝑏𝑜𝑡ℎ 𝑝𝑟𝑜𝑏𝑙𝑒𝑚𝑠.
To buttress it
Write the following LP model in the standard form
𝑀𝑖𝑛 𝑍 = 2𝑥1 + 3𝑥2
 Page 22 of 45

Subject to
𝑥1 + 𝑥2 = 10
−2𝑥1 + 3𝑥2 ≤ −5
7𝑥1 − 4𝑥2 ≤ 6
𝑥1 unrestricted
𝑥2 ≥ 0

Do the following changes

1. Multiply the second constraint by -1 and subtract a surplus variable 𝑆2 ≥ 0 from the LHS.
2. Add a slack variable 𝑆3 ≥ 0 to the LHS of the third constraint
3. Substitute 𝑥1 = 𝑥1′ − 𝑥1′′ , 𝑤ℎ𝑒𝑟𝑒 𝑥1,′ 𝑥1′′ ≥ 0, in the objective function and all the constraints.
The standard form is
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑍 = 2𝑥1′ − 2𝑥1′′ + 3𝑥2
Subject to
𝑥1′ − 𝑥1′′ + 𝑥2 = 10
2𝑥1′ − 2𝑥1′′ − 3𝑥2 − 𝑆2 = 5
7𝑥1′ − 7𝑥1′′ − 4𝑥2 + 𝑆3 = 6
𝑥1′ , 𝑥1′′ , 𝑥2 , 𝑥3 , 𝑆2 , 𝑆3 ≥ 0

Graphical Solution of LP models

Methods of feasible directions.
As said earlier, it is only model with at most two variables that can be solved graphically. The first step
in the graphical method is to plot the feasible solution space that satisfies all the constraints
simultaneously. The Figure 1 depicts the required solution space.
For instance, in Example 1: 𝑀𝑎𝑥 𝑍 = 3𝑥𝐸 + 2𝑥1
Subject to
𝑥𝐸 + 2𝑥1 ≤ 6 (1) 𝑥1 ≥ 0 (6)
2𝑥𝐸 + 𝑥1 ≤ 8 (2)
−𝑥𝐸 + 𝑥1 ≤ 1 (3)
𝑥1 ≤ 2 (4)
𝑥𝐸 ≥ 0 (5)
Solving the above example graphically, the non-negativity restrictions 𝑥𝐸 ≥ 0 and 𝑥1 ≥ 0 confine all the
feasible values to the first quadrant (which is defined by the space above or on the 𝑥𝐸 − 𝑎𝑥𝑖𝑠 and to the
right or on the 𝑥1 − 𝑎𝑥𝑖𝑠). The space enclosed by the remaining constraints is determined by first
replacing (≤) by (−) for each constraint, thus yielding a straight-line equation. Each straight line is then
plotted on the (𝑥𝐸 , 𝑥1 ) plane, and the region in which each constraint holds when the inequality is activated
is indicated by the direction of the arrow on the associated straight-line. An easy way to determine the
direction of the arrow is to use the origin (0,0) as a reference point. If (0, 0) satisfies the inequality, the
feasible direction should include the origin; otherwise, it should be on the opposite side. For instance, (0,
0) satisfies the inequality −𝑥𝐸 + 𝑥1 ≤ 1, meaning that inequality is feasible in the (half) space that
includes the origin. If the constraint happens to pass through the origin, a reference point that does not lie
on the associated straight-line must be selected. Applying this procedure to example 1. To find the optimum
solution, we move the revenue line “uphill” to the point where any further increase in revenue would render an
infeasible solution. That solution occurs at point C. since C is the intersection of lines 1 and 2, the values of
𝑥𝐸 𝑎𝑛𝑑 𝑥1 are determined by solving the following two equations simultaneously
𝑥𝐸 + 2𝑥1 = 6
2𝑥𝐸 + 𝑥1 = 8
 Page 23 of 45

X1
5
Using Eqn. (1), 𝒙𝑬 = 𝟎; 𝒙𝟏 = 𝟑 𝒘𝒉𝒆𝒏 𝒙𝟏 = 𝟎; 𝒙𝑬 = 𝟔
8
7 Using Eqn. (2), 𝒙𝑬 = 𝟎; 𝒙𝟏 = 𝟖 𝒘𝒉𝒆𝒏 𝒙𝟏 = 𝟎; 𝒙𝑬 = 𝟒
6 2
3
5 Using Eqn. (3), 𝒙𝑬 = 𝟎; 𝒙𝟏 = 𝟏 𝒘𝒉𝒆𝒏 𝒙𝟏 = 𝟎; 𝒙𝑬 = −𝟏
1 4 Using Eqn. (4); 𝒙𝟏 = 𝟐
3 G
K 4
2 H E D Using Eqn. (5); 𝒙𝑬 = 𝟎
Solution C
1 F
Space J
A B XE Using Eqn. (6); 𝒙𝟏 = 𝟎
0 1 2 3 4 5 6 6

Figure 1 Graphical solution to example 1

Each point within or on the boundary of the solution ABCDEF satisfies all the constraints and hence
represents a feasible point.
𝑀𝑎𝑥 𝑍 = 3𝑥𝐸 + 2𝑥1 At Pt B: solve Eqn. (2) & (6)
At point A of Figure 1; 𝑍 = 0; 𝑥𝐸 = 0; 𝑥1 = 0
,, Point B ; 𝑍 = 12; 𝑥𝐸 = 4; 𝑥1 = 0 At Pt C; solve Eqn. (1) & (2) simultaneously
2 10 4
At Point C : 𝑍 = 12 3 ; 𝑥𝐸 = 3 ; 𝑥1 = 3 At Pt D; solve Eqn. (1) & (4)
,, Point D ; 𝑍 = 10; 𝑥𝐸 = 2; 𝑥1 = 2
At Pt E; solve Eqn. (3) & (4)
,, Point E ; 𝑍 = 7; 𝑥𝐸 = 1; 𝑥1 = 2
,, Point F ; 𝑍 = 2; 𝑥𝐸 = 0; 𝑥1 = 1 At Pt F; solve Eqn. (3) & (5)

The algebraic or Simplex method

The simplex method involves the use of algebraic solution (example the Gauss-Jordan method) in finding
the values of the variables to provide the optimal value of the objective function. The following are the
basic definitions usually associated with the simplex method.
Let there be M-constraint equations and n variables
1. Basic solution: This is the unique solution resulting from setting n-m variables equal to zero
2. Feasible basic solution: If the basic solution satisfies all the nonnegativity constant, it is called a
feasible basic solution.
3. Non-basic variables: These are the variables set to zero. The remaining are said to be basic variables
4. Entering variable: This is a current non-basic variable that will enter the set of basic variables at the
next iteration
5. The leaving variable: This is a current basic variable that will leave the basic solution in the next
solution.
6. Optimality condition: This states that in the case of maximization, if all the non-basic variables have
nonnegative coefficients in the z-equation of the current tableau, the variable with the most negative
coefficient is selected as the entering variable.
7. Entering column: This is the column that contains the entering variable.
8. Pivot equation: This is the row associated with the leaving variable.
9. The Pivot element: This is the element at the intersection of the entering column and pivot equation.
The entering variable in the maximization (minimization) is the non-basic variable in the most negative
(positive) coefficient in the z-equation. A tie is broken arbitrary. When all the non-basic coefficients in
the z- equation are nonnegative (non-positive), the optimum is reached. This is the optimality condition.
For both maximization (minimization) problems, the leaving variable is the basic variable having the
smallest ratio (with positive denominator). A tie is broken arbitrary. This is the feasibility condition.
 Page 24 of 45

Computational Details of the Simplex Algorithm

The following are the steps of the simplex algorithm
Step 0: Using the standard form, determine a starting basic feasible solution by setting n-m appropriate
(non-basic) variables at zero level.
Step 1: Select an entering variable from among the current (zero) non-basic variables which, when
increased above zero, can improve the value of the objective function. If none exits, stop; the current basic
solution is optimal. Otherwise go to step 2.
Step 2: Select a leaving variable from among the current basic variables that must be set to zero (become
non-basic) when the entering variable became basic.
Step 3: Determine the new basic solution by making the entering variable basic and the leaving variable
non-basic; go to step 1. Note that the entering and leaving variables are obtained by using optimality and
feasibility conditions. The next iteration is carried out by using the Guass-Jordan method. This involves
Type 1: (Pivot equation)
New pivot equation = old pivot equation + pivot element
Type 2 (all equations including Z)
𝑁𝑒𝑤 𝑒𝑞𝑢𝑎𝑡𝑖𝑜𝑛 = 𝑜𝑙𝑑 𝑒𝑞𝑢𝑎𝑡𝑖𝑜𝑛𝑠 – 𝑖𝑡𝑠 𝑒𝑛𝑡𝑒𝑟𝑖𝑛𝑔 𝑐𝑜𝑙𝑢𝑚𝑛 𝑐𝑜𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡𝑠) 𝑥 (𝑛𝑒𝑤 𝑝𝑖𝑣𝑜𝑡 𝑒𝑞𝑢𝑎𝑡𝑖𝑜𝑛).

Worked Example 1:
𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝑍 − 2𝑥1 − 𝑥2 + 3𝑥3 − 5𝑥4 = 0
Subject to
𝑥1 + 7𝑥2 + 3𝑥3 + 7𝑥4 ≤ 46
3𝑥1 − 𝑥2 + 𝑥3 + 2𝑥4 ≤ 8
2𝑥1 + 3𝑥2 − 𝑥3 + 𝑥4 ≤ 10
𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ≥ 0
Using simplex method

Solution
The LP problem should be in standard format.
𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝑍 = 2𝑥1 + 𝑥2 − 3𝑥3 + 5𝑥4 + 0𝑆1 + 0𝑆2 + 0𝑆3
Subject to
𝑥1 + 7𝑥2 + 3𝑥3 + 7𝑥4 + 𝑆1 = 46
3𝑥1 − 𝑥2 + 𝑥3 + 2𝑥4 + 𝑆2 = 8
2𝑥1 + 3𝑥2 − 𝑥3 + 𝑥4 + 𝑆3 = 10
𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑆1 , 𝑆2 , 𝑆3 ≥ 0
n = 7
m= 3
Set n-m variable to zero, therefore, 𝑆1, 𝑆2 , 𝑎𝑛𝑑 𝑆3 becomes basic variables
Leaving Entering Variable Key column
Initial Simplex table variable
𝐶𝐵i 𝐶𝑗 2 1 -3 5 0 0 0 Solu Ratio
Basic 𝑋1 𝑋2 𝑋3 𝑋4 𝑆1 𝑆2 𝑆3 tion
Variables
0 𝑆1 1 7 3 7 1 0 0 46 46 4
=6
7 7
0 𝑆2 3 -1 1 2 0 1 0 8 8/2=4
0 𝑆3 2 3 -1 1 0 0 1 10 10/1=10
𝑍𝑗 (0x1)+(0x3 0 0 0 0 0 0
)+(0x2)=0
𝐶𝑗 − 𝑍𝑗 2 1 -3 5 0 0 0
Where 𝐶𝑗 is the coefficient of the basic variable Key row
 Page 25 of 45

To find 𝑍𝑗 variables apply this formular

3

𝑍𝑗 = ∑(𝐶𝐵1 ) (𝑎𝑖𝑗 ) ≤ 0
𝑖=1

𝐶𝑗 − 𝑍𝑗 for column X1= 2-0 =2

After solving for 𝐶𝑗 − 𝑍𝑗 , then calculate for Optimality Conditions

For Max:
All 𝐶𝑗 − 𝑍𝑗 ≤ 0 𝑖. 𝑒 0 𝑜𝑟 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑣𝑎𝑙𝑢𝑒
For Min:
All 𝐶𝑗 − 𝑍𝑗 ≥ 0 𝑖. 𝑒 0 𝑜𝑟 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑣𝑎𝑙𝑢𝑒
Hence select the max value in 𝐶𝑗 − 𝑍𝑗 to give you key column. Similarly, to select the key row;
select the least value in the ratio column.
Pivot or key element in this case is 2. From the above table, 𝑋4 is the entering variable while 𝑆2 is the
leaving variable
First iteration
No. 𝐶𝐵i 𝐶𝑗 2 1 -3 5 0 0 0 Solu Ratio
iteration
B.V. 𝑋1 𝑋2 𝑋3 𝑋4 𝑆1 𝑆2 𝑆3
Ist 0 𝑆1 19 21 1 0 1 7 0 18 1.7
iteration − − −
X4 2 2 2 2
enter 5 𝑋4 3 -1/2 1/2 2/2 =1 0/2=0 0/2=0 0/2= 0 8/2=4
S2
leaves
2
0 𝑆3 1 7 3 0 0 1 1 6 1.7
− −
2 2 2 2
𝑍𝑗 15 5 5 5 0 0 0
−
2 2 2
𝐶𝑗 − 𝑍𝑗 13 7 11 0 0 0 0
− −
2 2 2
To find the new values of entering row, divide all row values by the pivot number or key
element i.e., 2
To get all other new rows apply this formular = (𝑖)𝑁𝑒𝑤 𝑣𝑎𝑙𝑢𝑒 = 𝑜𝑙𝑑 −
𝑐𝑜𝑟𝑟𝑒𝑠𝑝𝑜𝑛𝑑𝑖𝑛𝑔 𝑐𝑜𝑙𝑢𝑚𝑛 𝑣𝑎𝑙𝑢𝑒 × 𝑐𝑜𝑟𝑟 𝑟𝑜𝑤 𝑣𝑎𝑙𝑢𝑒
𝑃𝑖𝑣𝑜𝑡 (𝑘𝑒𝑦 𝑒𝑙𝑒𝑚𝑒𝑛𝑡)
(ii) 𝑂𝑙𝑑 𝑣𝑎𝑙𝑢𝑒 – 𝑁𝑒𝑤 𝑣𝑎𝑙𝑢𝑒 𝑥 𝑘𝑒𝑦 𝑐𝑜𝑙𝑢𝑚𝑛
𝑣𝑎𝑙𝑢𝑒𝑠 𝑜𝑓 𝑡ℎ𝑒 𝑅𝑜𝑤
(iii) 𝐹𝑜𝑟 𝑅𝑜𝑤 𝑡ℎ𝑎𝑡 ℎ𝑎𝑣𝑒 𝑝𝑖𝑣𝑜𝑡 =
𝑃𝑖𝑣𝑜𝑡 (𝑘𝑒𝑦 𝑒𝑙𝑒𝑚𝑒𝑛𝑡)
For 𝑆1 1
− ×−
7
3 19 2 2 7
Under 𝑋1 : 𝑁𝑒𝑤 = 1 − (7 × ) = − Second Iteration for X4: S2: 𝑁𝑒𝑤 = 0 − ( 21⁄ ) =−
2 2
2 2
1 21
𝑋2 : 𝑁𝑒𝑤 = 1 − (7 × − ) = 3
1
− ×−
19
22
2 2 Under X1: 𝑁𝑒𝑤 = −( 2
21⁄ )
2
=
1 1 2 2 21
𝑋3 : 𝑁𝑒𝑤 = 3 − (7 × ) = −
2 2 1 − ×
1 21
− ×0
1
𝑋4 : 𝑁𝑒𝑤 = 7 − (7 × 1) = 0 X2: 𝑁𝑒𝑤 = −2 2 2
− ( 21⁄2
) = 0 S3: 𝑁𝑒𝑤 = 0 − ( 212⁄ ) =0
2
𝑢𝑛𝑑𝑒𝑟 𝑆1 : 𝑁𝑒𝑤 = 1 − (7 × 0) = 1
1 7 1 1 1
− ×8
𝑆2 : 𝑁𝑒𝑤 = 0 − (7 × ) = − X3: 𝑁𝑒𝑤 =
1
−
− ×−
2
( 21 2
) =
10
Solu: 𝑁𝑒𝑤
2
= 4 − ( 21 ) =
30
2 2 2 ⁄2 21 ⁄2 7

1
− ×0
X4: 𝑁𝑒𝑤 = 1 − ( 212⁄ ) =1
2

1
− ×1 1
S1: 𝑁𝑒𝑤 = 0 − ( 212⁄ ) = 21
2
 Page 26 of 45

0
𝑆3 : 𝑁𝑒𝑤 = 0 − (7 × ) = 0
2
7×8
Solu: = 46: 𝑁𝑒𝑤 = 46 − ( 2 ) = 18
8×1
Solu: = 10: 𝑁𝑒𝑤 = 10 − ( )=6
2

SECOND ITERATION
𝐶𝐵𝑗 𝐶𝑗 2 1 -3 5 0 0 0 Solu Ratio
B.V. 𝑋1 𝑋2 𝑋3 𝑋4 𝑆1 𝑆2 𝑆3
1 𝑋2 19 1 1 0 2 7 0 12
− − −
21 21 21 21 7
5 𝑋4 22 0 10 1 1 7 0 30
21 21 21 42 7
0 𝑆3 22 0 35 0 7 7 1 0
− −
6 21 21 6
𝑍𝑗 91 1 49 5 3 21 0
21 21 21 42
49 112 3 21
𝐶𝑗 − 𝑍𝑗 0 − 21 0
− − −
21 21 42
nd 21
For 2 iteration use the pivot ( 2 ) to divide all through the row.

Second Iteration for S3:

7 19 7 7
1 ×− 22 ×− 7
2 2 2 2
Under X1: 𝑁𝑒𝑤 = −( 21⁄ )= S2: 𝑁𝑒𝑤 = 0 − ( 21 ) =
2 2 6 6 ⁄2
7 21 7
7 × ×0
2 2 2
X2: 𝑁𝑒𝑤 = − ( 21 ) = 0 S3: 𝑁𝑒𝑤 = 1 − (21 ) = 1
2 ⁄2 ⁄2
7 1 7
3 − ×− 35 ×18
X3: 𝑁𝑒𝑤 = − − 2
( 21 2
) =− Solu: 𝑁𝑒𝑤 = 6 − ( 21
2
)=0
2 ⁄2 21 ⁄2
7
×0
2
X4: 𝑁𝑒𝑤 = 0 − (21 ) = 0
⁄2
7
×1 7
2
S1: 𝑁𝑒𝑤 = 0 − (21 ) = −
21 ⁄2
After the 2nd iteration the values in 𝑍𝑗 were all positive hence the optimum reach

Artificial Starting Solution

In the simplex problem treated earlier, the slack variables were used as the starting basic solution. In the
cases where the original constraints were an equation or of the type (≥) we no longer have a ready starting
basic feasible solution. When one encounters this kind of problem, one has to use the method usually
referred to as the M- technique (method of penalty).

Worked Example: 2
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑍 = 4𝑥1 + 𝑥2
Subject to:
3𝑥1 + 𝑥2 = 3
4𝑥1 + 3𝑥2 ≥ 6
 Page 27 of 45

𝑥1 + 2𝑥2 ≤ 4
𝑥1 , 𝑥2 ≥ 0
Solution:
Put in the standard form to get
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑍 = 4𝑥1 + 𝑥2
Subject to:
3𝑥1 + 𝑥2 =3
4𝑥1 + 3𝑥2 − 𝑥3 =6
𝑥1 + 2𝑥2 + 𝑥4 =4
𝑥1 , 𝑥2 ≥ 0

The first and second equations do not have variables that play the role of a slack. Hence, we augment the
two with artificial variables 𝑅1 𝑎𝑛𝑑 𝑅2 in these two equations follows:
3𝑥1 + 𝑥2 + 𝑅1 =3
4𝑥1 + 3𝑥2 − 𝑥3 + 𝑅2 = 6
One can now penalize 𝑅1 𝑎𝑛𝑑 𝑅2 in the objective function by assigning them very large positive
coefficients in the objective function. Let M > 0 be a very large constant; then the LP with its artificial
variables becomes.
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑍 = 4𝑥1 + 𝑥2 + 𝑀𝑅1 + 𝑀𝑅2
Subject to:
3𝑥1 + 𝑥2 + 𝑅1 =3
4𝑥1 + 3𝑥2 − 𝑥3 + 𝑅2 =6
𝑥1 + 2𝑥2 + 𝑥4 = 4
𝑥1 , 𝑥2 , 𝑥3, 𝑅1 , 𝑅2 , 𝑥4 ≥ 0
Notice the reason behind the use artificial variables. There were 3 equations and 6 unknowns. Now we
have n-m (6-3) non-basic variables and 3 basic variables.
See the Tables below, the process is the same as the simplex method already treated.
First substitute the two artificial variables R1 and R2. After this calculate 𝒁𝒋 and 𝑪𝒋 − 𝒁𝒋
Thirdly Check optimality value Min = 𝐶𝑗 − 𝑍𝑗 ≥ 0 then get the most negative value from 𝐶𝑗 − 𝑍𝑗 in this
case is 4 - 7M. To find Key row calculate the ratio using the entering column values. Then select the least
positive value form ratio column
Initial Table
No of 𝐶𝑗 4 1 0 M M 0 solution Ratio
iterations
0 starting CB B. V. 𝑥1 𝑥2 𝑥3 𝑅1 𝑅2 𝑥4

M 𝑅1 3 1 0 1 0 0 3 3
=1 ÷𝟑
3
M 𝑅2 4 3 -1 0 1 0 6 6
= 1.5
4
0 𝑥4 1 2 0 0 0 1 4 4
=4
1
𝑍𝑗 7m 4M -M M M 0 9M
𝐶𝑗 − 𝑍𝑗 4-7M 1-4M M 0 0 0
 Page 28 of 45

1st iteration Table

No of 𝐶𝑗 4 1 0 M M 0 solution Ratio
iterations
1st CB B. 𝑥1 𝑥2 𝑥3 𝑅1 𝑅2 𝑥4
iteration V.
4 𝑥1 3 1 0 1 0 0 3 3
=1 =1
3 3 3 3
M 𝑅2 0 5 -1 4 1 0 2 6
−
3 3 5
0 𝑥4 0 5 0 1 0 1 3 9
−
3 3 5
𝑍𝑗 4 4 5 -M 4 4 M 0 4 +2M
+ 𝑀 − 𝑀
3 3 3 3
𝐶𝑗 − 𝑍𝑗 0 1 5 M 4 𝑀 0 0
− − 𝑀 − +
3 3 3 3

2nd iteration Table

No of 𝐶𝑗 4 1 0 M M 0 solution Ratio
iterations
2nd CB B. 𝑥1 𝑥2 𝑥3 𝑅1 𝑅2 𝑥4
iteration V.
4 𝑥1 1 0 1 3 1 0 3 3
−
5 5 5 5 5
1 𝑥2 0 1 3 4 3 0 6 6
− − 5
5 5 5 5 5 ÷
3
0 𝑥4 0 0 1 1 −1 1 1 1

𝑍𝑗 4 1 1 8 1 0 18
− −
5 5 5 5
𝐶𝑗 − 𝑍𝑗 0 0 1 8 1
− 𝑀+ 𝑀+
5 5 5

No of 𝐶𝑗 4 1 0 M M 0 solution Ratio
iterations
3rd CB B. 𝑥1 𝑥2 𝑥3 𝑅1 𝑅2 𝑥4
iteration V.
4 𝑥1 1 0 0 2 0 1 2
−
5 5 5
1 𝑥2 0 1 0 1 0 3 9
−
5 5 5
0 𝑥3 0 0 1 1 −1 1 1

2 9 17
Therefore, the optimum solution gives 𝑥1 = 5 ; 𝑥2 = 5 ; 𝑥3 = 1, 𝑍𝑗 = 5
 Page 29 of 45

Application of LP to power systems problems economic dispatch, automatic load shedding

generation expansion studies.

Fuel scheduling
Considering the limitations imposed by long-term yearly fuel contracts, the objective is to schedule fuel
Deliveries and storage to meet plant requirements
Dispatching
The objective is to schedule the committed generators to meet the load, maintain voltages, frequency within
prescribed tolerances and minimize operating cost without unduly stressing equipment.

1.0 Operational Cost of a thermal Plant

There are three factor that affect power generation at minimum cost. The factors are – efficiency of
generators, fuel cost and transmission losses. A very highly efficient plant may be located in a place with
high cost of fuel and this does not indicate minimum cost. Also, if a plant is located very far from the load
center, the transmission losses along the line may be so high that the cost of running the generator becomes
extremely exorbitant. So, the usual problem is to determine the generation of different plants such that the
total cost is minimum. Below is a brief discussion on the role of operating cost plays in economic
scheduling of plants.
Let the input to the thermal plant be Btu/h (British thermal unit/hour) and the power output measured in
Megawatts (MW). Figure 1 (a) shows the input-output curve of a thermal plant unit known as heat curve.
If the cost of fuel is so many Naira per million Btu, we get fuel cost curve.
(N/hr)

Fuel
input
(Btu/hr)

(a) Pi (MW) (b)

Figure 1 (a) Heat rate curve (b) Operating Cost

Observation of the fuel cost curve shows that cost can be approximated to the quadratic curve.
𝐶𝑖 = 𝛼𝑖 + 𝛽𝑖 𝑃𝑖 + 𝛾𝑖 𝑃𝑖2 (1)
Another important curve that usually shows a measure of how costly it will be to produce the next
increment of power is known as the incremental cost curve; it is the plotting of the derivative of the
fuel cost curve against the real power (Figure 2) that is
𝑑𝐶𝑖
= 2𝛾𝑖 𝑃𝑖 + 𝛽𝑖 (2)
𝑑𝑃𝑖

λi
N/MW

Pi (MW)
Figure 2: Incremental fuel cost curve
 Page 30 of 45

Note that total operating cost usually includes fuel cost and cost of labour, supplies and maintenance.
These other costs were assumed fixed and to be a percentage of the fuel cost and for this they were
generally included in the incremental fuel cost curve.
Input - Output Characteristic of Thermal Units
For thermal units, we call the input - output characteristic the generating unit fuel consumption function,
or operating cost function. The unit of the generator fuel consumption function is Btu per hour heat input
to the unit (or MBtu/h). The fuel cost rate times Btu/h is the $ per hour ($/h) input to the unit for fuel. The
output of the generating unit will be designed by PG, the megawatt net power output of the unit.

In addition to the fuel consumption cost, the operating cost of a unit includes labor cost, maintenance cost,
and fuel transportation cost. It is difficult to express these costs directly as a function of the output of the
unit, so these costs are included as a fixed portion of the operating cost.

The thermal unit system generally consists of the boiler, the steam turbine, and the generator. The input
of the boiler is fuel, and the output is the volume of steam. The relationship of the input and output can be
expressed as a convex
curve. The input of the turbine - generator unit is the volume of steam, and the output is the electrical
power. A typical boiler - turbine - generator unit consists of a single boiler that generates steam to drive a
single turbine - generator set. The input - output characteristic of the whole generating unit system can be
obtained by combining directly the input - output characteristic of the boiler and the input - output
characteristic of the turbine - generator unit. It is a convex curve, which is shown in Figure 1

FIGURE 1 Input - output characteristic of the generating unit

Figure 2 Simple Model of a fossil plant

Economic Dispatch (Neglecting losses and no generator limits)

If transmission losses were neglected, it all implies that the grid does not have impedances and that
all the generated power gets to the common bus. Figure 3 shows the diagrammatic representation of
such simplified economic dispatch problem.
 Page 31 of 45

C1 C2 Cng

P2
P1 Pn

Figure 3 Plants connected to a common bus

Let PD be the total power demand equal to the total generation as a result of no losses. Also, let Ci be
the known fuel cost of each plant.
The problem is to find the real power generation for each plant such that the objective function (i.e.,
total production cost) as defined by the Eqn(3) is a minimum, subject to the constraint.
𝑛𝑔

𝐶𝑡 = ∑ 𝐶𝑖 = 𝛼𝑖 + 𝛽𝑖 𝑃𝑖 + 𝛾𝑖 𝑃𝑖2 (3)
𝑖=1
s.t,
𝑛𝑔

∑ 𝑃𝑖 = 𝑃𝐷 (4)
𝑖=1
Where 𝐶𝑡 = Total production Cost; 𝐶𝑖 = production cost of ith plant, 𝑃𝑖 = generation of ith plant,
𝑃𝐷 = total load demand, ng = total number of dispatchable generating plants.
The method of Lagrange multipliers to convert the constrained optimization (non-linear) problem to
unconstrained one.
𝑛𝑔

ℒ = 𝐶𝑡 + 𝜆 (𝑃𝐷 − ∑ 𝑃𝑖 ) (5)
𝑖=1
As usual, the minimum is obtained at the point where the partial derivative of the function to its
variables are zero.
𝜕ℒ
=0 (6)
𝜕𝑃 𝑖
𝜕ℒ
=0 (7)
𝜕𝜆
𝐸𝑞𝑢𝑎𝑡𝑖𝑜𝑛 6 𝑦𝑖𝑒𝑙𝑑𝑠
𝜕 𝐶𝑡
+ 𝜆 (0 − 1)
𝜕 𝑃𝑖
𝑏𝑢𝑡 𝐶𝑡 = 𝐶1 + 𝐶2 + . . . + 𝐶𝑛𝑔
𝜕𝐶𝑡 𝑑𝐶
Then = 𝑑𝑃𝑡 = 𝜆
𝜕𝑃𝑖 𝑖
This gives the condition for optimum dispatch as
𝑑𝐶𝑡
= 𝜆, 𝑖 = 1, 2, . . . 𝑛𝑔 (8)
𝑑𝑃𝑖
𝑜𝑟 𝛽𝑖 + 2𝛾𝑖 𝑃𝑖 = 𝜆 (9)
𝑤ℎ𝑒𝑟𝑒 𝐶𝑖 = 𝛼𝑖 + 𝛽𝑖 𝑃𝑖 + 𝛾𝑖 𝑃𝑖2
Applying Eqn. (7) to (5)
𝑛𝑔

∑ 𝑃𝑖 = 𝑃𝐷 (10)
𝑖=1
We can summarize by saying that when losses were neglected with no generation limits, for most
economic operation; all plants must operate at equal incremental production cost while satisfying the
equality constraints. From Eqn. (9)
 Page 32 of 45

𝜆−𝛽𝑖
𝑃𝑖 = where i= 1, 2, . . . ng (11)
2𝛾𝑖
Substituting Eqn. (11) into (10) and solving for 𝜆 gave;
𝑛𝑔
𝜆 − 𝛽𝑖
∑ = 𝑃𝐷 (12)
2𝛾𝑖
𝑖=1
𝑀aking 𝜆 the subject
𝑛𝑔 𝑛𝑔

∑ 𝜆 − 𝛽𝑖 = 𝑃𝐷 × ∑ 2𝛾𝑖
𝑖=1 𝑖=1
𝑛𝑔 𝑛𝑔

𝜆 = 𝑃𝐷 (∑ 2𝛾𝑖 ) + ∑ 𝛽𝑖
𝑖=1 𝑖=1

Multiple both Sides by ∑ 1/2𝛾𝑖

𝑛𝑔 𝑛𝑔 𝑛𝑔 𝑛𝑔 𝑛𝑔
1 1
𝜆 ∑ 2𝛾𝑖 = 𝑃𝐷 (∑ 2𝛾𝑖 ) (∑ ) + (∑ 𝛽𝑖 ) (∑ )
2𝛾𝑖 2𝛾𝑖
𝑖=1 𝑖=1 𝑖=1 𝑖=1 𝑖=1
𝑛𝑔 1
Divide the RHS by (∑𝑖=1 2𝛾 )
𝑖

𝑛𝑔 𝛽
𝑃𝐷 + ∑𝑖=1 2𝛾𝑖
𝑖
∴𝜆= (13)
1
∑𝑛𝑔
𝑖=𝑖 2𝛾
𝑖
The value of 𝜆 found in Eqn. (13) is substituted in Eqn. (11) to obtain the optimal dispatch of
generation. This is the analytical method of solving this economic dispatch problem.

ITERATIVE TECHNIQUES

When the system equations become more complicated as a result of inclusion of losses, a better
approach may be used, this is the iterative techniques.
Equation (12) can be written as
𝑓(𝜆) = 𝑃𝐷 (14)
If the left-hand side is expanded in Taylor series about an operating point 𝜆(𝑘) , and ignoring higher-order
terms gave
𝑑𝑓(𝜆) (𝑘)
𝐹(𝜆)𝑘 + ( ) ∆𝜆(𝑘) = 𝑃𝐷 (15)
𝑑𝜆
Or
∆𝑃(𝑘) ∆𝑃 (𝑘)
(𝜆)𝑘 = = 𝑑𝑃𝑖 (𝑘)
(16)
𝑑𝑓 (𝑘)
( ) ∑( )
𝑑𝜆 𝑑𝜆
Or
∆𝑃 (𝑘)
∆𝜆𝑘 = 1 (See eqn. (11) (17)
∑
2𝛾𝑖
and so,
𝜆(𝑘+1) = 𝜆(𝑘) + ∆𝜆(𝑘) (18)
 Page 33 of 45

𝑛𝑔 (𝑘)
where ∆𝑃(𝑘) = 𝑃𝐷 − ∑𝑖=1 𝑃𝑖 (19)
(𝑘)
the process continues until ∆𝑃 is less than a specified accuracy.

Example 1:
The fuel-cost function for three thermal plants in N/hr were given by
𝐶1 = 500 + 5.3𝑃1 + 0.004𝑃12
𝐶2 = 400 + 5.5𝑃2 + 0.006𝑃22
𝐶3 = 200 + 5.8𝑃1 + 0.009𝑃32
When P1, P2 and P3 are in MW. The total load, PD is 800 MW. Neglecting line losses and generator
limits, find the optimal dispatch and the total cost in N/hr. Using both analytical and iterative methods.
(Hint use guess of 6.0 where applicable)
Solution
(i) By analytical/gradient method
Applying Eqn. (13)
𝑛𝑔 𝛽
𝑃𝐷 +∑𝑖=1 𝑖
2𝛾𝑖
𝜆= 𝑛𝑔 1
∑𝑖=𝑖
2𝛾𝑖
800+ 5.3⁄
0.008+5.5⁄0.012+5.8⁄0.018
𝜆 = 1⁄ 1 1 (where 𝛾𝑖 = 0.004, 0.006, 0.009)
0.008+ ⁄0.012+ ⁄0.018
800+1443.0555
= 263.8889 = 8.5 𝑁/𝑀𝑊ℎ
Applying the value of 𝜆 in Eqn. (11) gives

𝜆−𝛽𝑖
𝑃𝑖 =
2𝛾𝑖

8.5−5.3
𝑃1 = 2(0.004) = 400𝑀𝑊
8.5−5.5
𝑃2 = 2(0.006) = 250𝑀𝑊
8.5−5.8
𝑃3 = 2(0.009) = 150 𝑀𝑊

(ii) Numerical solution using iterative method.

𝜆−𝛽
Making initial estimation of 𝜆(1) = 6.0 and apply it in Eqn. (11) 𝑃𝑖 = 2𝛾 𝑖
𝑖
(1) 6.0 −5.3
𝑃1 = 2(0.004) = 87.5000
(1) 6.0 −5.5
𝑃2 = 2(0.006) = 41.6667
(1) 6.0 −5.8
𝑃3 = 2(0.009) = 11.1111
But PD = 800 MW given

To find the error in Power use Eqn. (19)

𝑛𝑔
(𝑘)
∆𝑃(𝑘) = 𝑃𝐷 − ∑ 𝑃𝑖
𝑖=1
∆𝑃(1) = 800 − (87.5 + 41.6667 + 11.1111) = 659.7222
Substitute this error in power in Eq. (17)
659.7222
∆𝜆(1) = 1 1 1 = 2.5
+ +
2(0.004) 2(0.006) 2(0.009)
(2) 8.5 −5.3
𝑃1 = 2(0.004) = 400
 Page 34 of 45

(2) 8.5 −5.5

𝑃2 = 2(0.006) = 250
(2) 8.5 −5.8
𝑃3 = 2(0.009) = 150
Hence, ∆𝑃(2) = 800 − (400 + 250 + 150) = 0

Therefore, the new value of 𝜆 is

𝜆(2) = 6.0 + 2.5 = 8.5
Continuing the process, for the second iteration, we obtain
Since, ∆𝑃(2) = 0, the equality constraint is met in two iterations, therefore, this gives the optimal
dispatch as
𝑃1 = 400𝑀𝑊
𝑃2 = 250𝑀𝑊
𝑃3 = 150 𝑀𝑊
Therefore, the total fuel cost becomes
𝐶𝑡 = 500 + 5.3(400) + 0.004(400)2
+ 400 + 5.5(250) + 0.006(250)2
+ 200 + 5.8(150) + 0.009(150)2

Take Home

Two units of the system have the following cost curves

2
𝑃𝑔1 = 0.05𝑃𝑔1 + 22𝑃𝑔1 + 120 𝑁/ℎ𝑟
2
𝑃𝑔2 = 0.06𝑃𝑔2 + 16𝑃𝑔2 + 120 𝑁/ℎ𝑟
where 𝑃𝑔 is in MW. Both the units operate at all times and the maximum and minimum loads on each unit are 100
MW and 20 MW, respectively. Determine the economic operating schedule of the plants for loads of 80 MW, 120
MW, and 180 MW, neglecting the transmission line losses.

Economic Dispatch neglecting losses and including Generator limits.

The generators’ output power was usually descripted to lie within certain minimum and maximum limits.
The problem is to find the real power generation for each plant such that the objective function
(i.e., total production cost) as defined by Eq. (3) is minimum, subject to the constraint given by Eq. (4)
and the inequality constraints given by;
𝑃𝑖 (min) ≤ 𝑃𝑖 ≤ 𝑃𝑖 (𝑚𝑎𝑥), 𝑖 = 1, 2, . . 𝑛𝑔 (20)
where 𝑃𝑖 (min) and 𝑃𝑖 (𝑚𝑎𝑥) were the minimum and maximum generating limits respectively, for plant i.
Kuhn Tucker is used in complementing the Lagrange conditions to include the inequality constraints as
additional terms. The necessary conditions for the optimal dispatch with losses neglected are:
𝑑𝐶𝑖
= 𝜆 𝑓𝑜𝑟 𝑃𝑖 (min) ≤ 𝑃𝑖 ≤ 𝑃𝑖 (𝑚𝑎𝑥),
𝑑𝑃 𝑖
𝑑𝐶𝑖
≤ 𝜆 𝑓𝑜𝑟 𝑃𝑖 = 𝑃𝑖 (𝑚𝑎𝑥),
𝑑𝑃𝑖
𝑑𝐶𝑖
≥ 𝜆 𝑓𝑜𝑟 𝑃𝑖 = 𝑃𝑖 (𝑚𝑖𝑛),
𝑑𝑃𝑖
The Gradient/iterative method is also employed here i.e., for estimated value of 𝜆, 𝑃𝑖 is found from the
coordinate equation and iteration is continued until ∑ 𝑃𝑖 = 𝑃𝐷 . If any plant reaches a maximum or
minimum, the plant is pegged at the limit. The unviolated plant operates at equal incremental cost.
Example 2:
Find the optimal dispatch and the total cost in N/hr. for the thermal plants of Example 1 when the total
load is 975 MW with the following generator limits in MW;
200 ≤ 𝑃1 ≤ 450
 Page 35 of 45

150 ≤ 𝑃2 ≤ 350
100 ≤ 𝑃3 ≤ 225
Assume 𝜆(0) = 6.0

Solution
𝜆−𝛽
Using 𝑃𝑖 = 2𝛾 𝑖
𝑖
(1) 6.0 −5.3
𝑃1 = 2(0.004) = 87.5000
(1) 6.0 −5.5
𝑃2 = 2(0.006) = 41.6667
(1) 6.0 −5.8
𝑃3 = 2(0.009) = 11.1111
But PD = 975 MW given
∆𝑃(1) = 975 − (87.5 + 41.6667 + 11.1111) = 834.7222
Substitute this error in power in Eq. (17)
834.7222
∆𝜆(1) = 1 1 1 = 3.1632
+ +
2(0.004) 2(0.006) 2(0.009)
Therefore, the new value of 𝜆 is
𝜆(2) = 6.0 + 3.1632 = 9.1632
(2)
Then compute 𝑃𝑖
(2) 9.1632 −5.3
𝑃1 = 2(0.004) = 482.8947
(2) 9.1632 −5.5
𝑃2 = = 305.2632
2(0.006)
(2) 9.1632 −5.8
𝑃3 = = 186.8421
2(0.009)
(2)
Hence, ∆𝑃 = 975 − (482.8947 + 305.2632 + 186.8421) = 0
Since, ∆𝑃(2) = 0, the equality constraint is met in two iterations, but observe that P1 exceeds it
maximum limit; so, we have to peg P1 to the upper limit say P1 = 450 and continue the iteration. This
implies that
∆𝑃(2) = 975 − (450 + 305.2632 + 186.8421) = 32.8947
Then from Eq. (17);
32.8947
∆𝜆(2) = 1 1 = 0.2368
+
2(0.006) 2(0.009)
The new value of 𝜆 becomes
𝜆(3) = 9.1632 + 0.2368 = 9.4
For the 3rd iteration, we obtain
(3)
𝑃1 = 450
(2) 9.4 −5.5
𝑃2 = 2(0.006) = 325
(2) 9.4 −5.8
𝑃3 = 2(0.009) = 200
∆𝑃(2) = 975 − (450 + 325 + 200) = 0
Now, ∆𝑃(3) = 0, showing that the equality constraint is met and P2 and P3 are within their limits given
rise to the optimal dispatch as
𝑷𝟏 = 𝟒𝟓𝟎𝑴𝑾
𝑷𝟐 = 𝟑𝟓𝟎𝑴𝑾
𝑷𝟑 = 𝟐𝟎𝟎 𝑴𝑾
𝝀 = 𝟗. 𝟒 𝑵/𝑴𝑾𝒉𝒓
Therefore, the total fuel cost becomes
𝐶𝑡 = 500 + 5.3(450) + 0.004(450)2
 Page 36 of 45

+ 400 + 5.5(325) + 0.006(325)2

= 8, 235. 25𝑁/ℎ𝑟

Economic Dispatch including losses

For short transmission line distances and very high load density area, the effect of transmission loss is
usually negligible and this makes the plants to operate at the same or equal incremental production cost
resulting to the optimal dispatch of generation. But for large interconnected networks where power is
transmitted to long distance and with low density areas, transmission losses are a major factor and effect
the optimum dispatch of generator. There are two ways in which the effect of transmission losses can be
represented. First of all, is to express the total transmission loss as a quadratic function of the generator
power outputs.
Such as
𝑛𝑔 𝑛𝑔

𝑃𝐿 = ∑ ∑ 𝑃𝑖 𝐵𝑖𝑗 𝑃𝑗 (21)
𝑖=1 𝑗=1
Secondly, by using a more general formula, containing a linear term and constant term generally referred
to as Kron’s loss formula is
𝑛𝑔 𝑛𝑔 𝑛𝑔

𝑃𝐿 = ∑ ∑ 𝑃𝑖 𝐵𝑖𝑗 𝑃𝑗 + ∑ 𝐵0𝑖 𝑃𝑖 + 𝐵00 (22)

𝑖=1 𝑗=1 𝑖=1
where
𝐵𝑖𝑗 = loss coefficient or called B-coefficients
If the function given by Eq. (22) is provided, then the economic dispatch problem is
Minimize the overall generating cost Ct which the function of plant output is
𝑛𝑔

𝐶𝑡 = ∑ 𝐶𝑖
𝑖=1
𝑛𝑔

= ∑ 𝛼𝑖 + 𝛽𝑖 𝑃𝑖 + 𝛾𝑖 𝑃𝑖2 (23)
𝑖=1
Subject to the constraint that generation should be equal to total demands plus losses, i.e.,
𝑛𝑔

∑ 𝑃𝑖 = 𝑃𝐷 + 𝑃𝐿 (24)
𝑖=1
Satisfying the inequality constraints, expressed as follows
𝑃𝑖 (min) ≤ 𝑃𝑖 ≤ 𝑃𝑖 (𝑚𝑎𝑥), 𝑖 = 1, 2, . . 𝑛𝑔 (25)
where 𝑃𝑖 (min) and 𝑃𝑖 (𝑚𝑎𝑥) were the minimum and maximum generating limits respectively for plant i.
Employing the Lagrangian’s multiplier and adding additional terms to include the inequality constraints,
we have
𝑛𝑔 𝑛𝑔 𝑛𝑔

ℒ = 𝐶𝑡 + 𝜆 (𝑃𝐷 + 𝑃𝐿 − ∑ 𝑃𝑖 ) + ∑ 𝜇𝑗 (𝑚𝑎𝑥) {𝑃𝑖 − 𝑃𝑖 (𝑚𝑎𝑥)} + ∑ 𝜇𝑗 (𝑚𝑖𝑛) 𝑃𝑖 − 𝑃𝑖 (𝑚𝑖𝑛) (26)

𝑖=1 𝑖=1 𝑖=1
Observe that 𝜇𝑗 (𝑚𝑎𝑥) = 0 when 𝑃𝑖 < 𝑃𝑖 (𝑚𝑎𝑥) and 𝜇𝑗 (𝑚𝑖𝑛) = 0, if 𝑃𝑖 > 𝑃𝑖 (𝑚𝑖𝑛). This implies that variable 𝜇
becomes active if the limit is violated.
As usual, the minimum of the unconstrained function of Eq. (26) is found at the point where the partials of the
function to its variables are zero.

𝜕ℒ
=0 (27)
𝜕𝑃𝑖
 Page 37 of 45

𝜕ℒ
=0 (28)
𝜕𝜆
𝜕ℒ
=0 (29)
𝜕𝜇𝑗(𝑚𝑎𝑥)
Equation (29) and (30) imply that Pi should be allowed to go beyond its limits and when Pi is within the
limits 𝜇𝑗 (𝑚𝑖𝑛) = 𝜇𝑗 (𝑚𝑎𝑥) = 0 and Khun Tucker function becomes the same as Lagrange.
Eq (27) gives
𝜕𝐶𝑡 𝜕𝑃
+ 𝜆 (0 + 𝐿 − 1) = 0 (30)
𝜕𝑃𝑖 𝜕 𝑃𝑖
Since, 𝐶𝑡 = 𝐶1 + 𝐶2 + . . . + 𝐶𝑛𝑔
Then,
𝜕𝐶𝑡 𝑑𝐶𝑡
=
𝜕𝑃𝑖 𝑑𝑃𝑖
This gives the condition for optimum dispatch to be
𝑑𝐶𝑖 𝜕𝑃
+ 𝜆 𝜕𝑃𝐿 = 𝜆 𝑖 = 1, 2, . . . 𝑛𝑔 (31)
𝑑𝑃𝑖 𝑖
𝝏 𝑷𝑳
The term = incremental transmission loss
𝝏 𝑷𝒊
Second condition Eq. (28) results to
𝑛𝑔

∑ 𝑃𝑖 = 𝑃𝐷 + 𝑃𝐿 (32)
𝑖=1
Equation (31) is classically rearranged as
1 𝑑𝐶𝑖
[ 𝜕𝑃 ] =𝜆 𝑖 = 1, 2, . . . 𝑛𝑔 (33)
1− 𝐿 𝑑𝑃𝑖
𝜕𝑃𝑖

Or

𝑑𝐶
𝐿𝑖 𝑑𝑃𝑖 = 𝜆 𝑖 = 1, 2, . . . 𝑛𝑔n (34)
𝑖
Where
𝐿𝑖 = 𝑝𝑒𝑛𝑎𝑙𝑡𝑦 𝑓𝑎𝑐𝑡𝑜𝑟 𝑜𝑓 𝑝𝑙𝑎𝑛𝑡 𝑖
1
= 𝜕𝑃𝐿 (35)
1−
𝜕𝑃𝑖
Observe that effect of transmission loss is to introduce a penalty factor with a value that depends on the
location of the plant. Eq (34) shows that the minimum cost is obtained when the incremental cost of each
plant multiplied by its penalty factor is the same for all the plants.
Eq. (2) gives the incremental production cost while from Eq. (22) the incremental transmission loss is
obtained to be
𝑛𝑔
𝜕𝑃𝐿
= 2 ∑ 𝐵𝑖𝑗 𝑃𝑗 + 𝐵0𝑖 (36)
𝜕𝑃𝑖 𝑗=1
Using both the incremental production cost and incremental transmission loss in Eq. (31) yields
𝑛𝑔

𝐵𝑖 + 2𝛾𝑖 𝑃𝑖 + 2𝜆 ∑ 𝐵𝑖𝑗 𝑃𝑗 + 𝐵0𝑖 𝜆 = 𝜆

𝑗=1
Or
 Page 38 of 45

𝑛𝑔
𝛾𝑖 1 𝐵𝑖
[ + 𝐵𝑖𝑖 ] 𝑃𝑖 + ∑ 𝐵𝑖𝑖 𝑃𝑗 = [1 − 𝐵0𝑖 − ] (37)
𝜆 2 𝜆
𝑗=1
𝑗≠𝑖
Extending Eq. (37) to all plants yield the following linear equations in matrix form
𝛾1
+ 𝐵11 𝐵12 …. 𝐵1𝑛𝑔 𝐵1
𝜆 𝑃1 1 − 𝐵01 −
𝛾2 𝜆
𝐵21. + 𝐵22. … 𝐵2𝑛𝑔 𝑃2. 1 𝐵2
.. 𝜆 .. ; [ . ]= 1 − 𝐵02 − (38)
. . ; ; 2 𝜆
. 𝐵𝑛𝑔
𝑃𝑛𝑔
𝛾𝑛𝑔
[1 − 𝐵0𝑛𝑔 − ]
[ 𝐵𝑛𝑔1 𝐵𝑛𝑔2 . .. 𝜆
+ 𝐵𝑛𝑔𝑛𝑔 ] 𝜆

As usual, the iterative process is continued using the gradient method. In doing this Eq. (37), Pi at the Kth
iteration is expressed as:
𝑛𝑔
(𝑘) (𝑘)
𝑃𝑖 = 𝜆(𝑘) (1 − 𝐵0𝑖 ) − 𝐵𝑖 − 2𝜆(𝑘) ∑ 𝐵𝑖𝑗 𝑃𝑗 (39)
𝑗≠𝑖
Using Pi from Eq. (39) in (32) results in
𝑛𝑔 𝑛𝑔
(𝑘) (𝑘) (𝑘) (𝑘)
∑𝜆 (1 − 𝐵0𝑖 ) − 𝐵𝑖 − 2𝜆 ∑ 𝐵𝑖𝑗 𝑃𝑗 = 𝑃𝐷 + 𝑃𝐿 (40)
𝑗≠𝑖 𝑗≠𝑖

Or
(𝑘) (𝑘)
𝑓(𝜆) = 𝑃𝐷 + 𝑃𝐿 (41)
Expanding left hand side of Eq. (41) in Taylor’s series about an operating point 𝜆(𝑘) , and neglecting the
higher-order terms results in

𝑑𝑓(𝜆) (𝑘)
𝑓(𝜆)(𝑘) + [ ] ∆𝜆(𝑘) = 𝑃𝐷 + 𝑃𝐿 (42)
𝑑𝜆
Or
∆𝑃(𝑘) ∆𝑃 (𝑘)
∆𝜆(𝑘) = = 𝑑𝑃𝑖 (𝑘)
(43)
𝑑𝑓 (𝑘)
[ ] ∑[ ]
𝑑𝜆 𝑑𝜆
where
𝑛𝑔 (𝑘) 𝑛𝑔 𝑛𝑔 (𝑘)
𝑑𝑃𝑖 𝛾𝑖 (1 − 𝐵0𝑖 ) + 𝐵𝑖𝑗 𝐵𝑖 − 2𝛾𝑖 ∑𝑗≠𝑖 𝐵𝑖𝑗 𝑃𝑗
∑[ ] =∑ 2 (44)
𝑑𝜆 2(𝛾𝑖 + 𝜆 (𝑘)
𝐵𝑖𝑗 )
𝑖=1 𝑖=1

Therefore,
𝜆(𝑘+1) = 𝜆(𝑘) + ∆𝜆(𝑘) (45)
where
𝑛𝑔
(𝑘) (𝑘)
∆𝑃(𝑘) = 𝑃𝐷 + 𝑃𝐿 − ∑ 𝑃𝑖 (46)
𝑖=1
The process continued until ∆𝑃(𝑘) is less than a specified accuracy.
Note that if approximate loss formula is expressed by
 Page 39 of 45

𝑛𝑔

𝑃𝐿 = ∑ 𝐵𝑖𝑖 𝑃𝑖2 (47)

𝑖=1
is used, 𝐵𝑖𝑖 = 0, 𝐵00 = 0, and the solution of the simultaneous equation given by Eq. (39) reduces to the
following simple expressions;

(𝑘) 𝜆(𝑘) − 𝐵𝑖
𝑃𝑖 = (48)
2(𝛾𝑖 + 𝜆(𝑘) 𝐵𝑖𝑗 )
And reduces to
𝑛𝑔 𝑛𝑔
𝑑𝑃𝑖 (𝑘) 𝛾𝑖 − 𝐵𝑖𝑖 𝐵𝑖𝑗
∑[ ] =∑ 2 (49)
𝑑𝜆 2(𝛾𝑖 + 𝜆(𝑘) 𝐵 )
𝑖=1 𝑖=1 𝑖𝑗

Efficiency of the system is given as

𝑃𝑜𝑢𝑡𝑝𝑢𝑡
𝜂=
𝑃𝑖𝑛𝑝𝑢𝑡
𝑃𝐿𝑜𝑠𝑠
=1−( )
∑𝑖 𝑃𝑖

Worked Example 3:
The fuel cost in N/hr. of three thermal plants of a power system were given as:
𝐶1 = 200 + 7.0𝑃1 + 0.008𝑃12 𝑁/ℎ𝑟.
𝐶2 = 180 + 6.3𝑃2 + 0.009𝑃22 𝑁/ℎ𝑟.
𝐶3 = 140 + 6.8𝑃1 + 0.007𝑃32 𝑁/ℎ𝑟.
where P1, P2 and P3 were in MW. Plant output were subjected to the following limits
10 𝑀𝑊 ≤ 𝑃1 ≤ 85 𝑀𝑊
10 𝑀𝑊 ≤ 𝑃2 ≤ 80 𝑀𝑊
10 𝑀𝑊 ≤ 𝑃3 ≤ 70 𝑀𝑊
Assume for this problem, that the real power loss is given by the simplified expression
𝑃𝐿 (𝑝. 𝑢) = 0.0218𝑃12 (𝑝. 𝑢) + 0.0228𝑃22 (𝑝. 𝑢) + 0.0179𝑃32 (𝑝. 𝑢)
Where the loss coefficients were specified in per unit on a 100 MVA base. Determine the optimal
dispatch of generation when the total system load is 150 MW. Assume initial value of 𝜆(0) = 8.0.

Solution

In the cost function Pi were expressed in MW. Therefore, the real power loss in terms of MW
generation should be

𝑃12 𝑃22 𝑃32

𝑃𝐿 (𝑝. 𝑢) = [0.0218 ( ) + 0.0228 ( ) + 0.0179 ( )] 𝑀𝑊
100 100 100

= 0.000218𝑃12 + 0.000228𝑃22 + 0.000179𝑃32 MW

In carrying out the numerical solution using the gradient method, using Eq. (48)

(𝑘) 𝜆(𝑘) −𝐵
𝑃𝑖 = 2(𝛾 +𝜆(𝑘)𝐵𝑖 and substituting the values appropriately gave rise to
𝑖 𝑖𝑗 )
 Page 40 of 45

(1) 8.0 − 7.0

𝑃1 = = 51.3136 𝑀𝑊
2(0.008 + (8.0 × 0.000218))

(1) 8.0 − 6.3

𝑃2 = = 78.5292 𝑀𝑊
2(0.009 + (8.0 × 0.000228))

(1) 8.0 − 6.8

𝑃3 = = 71.1575 𝑀𝑊
2(0.007 + (8.0 × 0.000179))

The real power loss using Eq (47)

(1)
𝑃𝐿 = 0.000218(51.3136)2 + 0.000228(78.5292)2 + 0.000179(71.1575)2 = 2.8864
Since PD = 150 MW, the error ∆𝑃(1) from Eq. (46) is
∆𝑃(1) = 150 + 2.8864 − (51.3136 + 78.5292 + 71.1575) = −48.1139
From Eq. (49)
3 (1)
𝑑𝑃𝑖 0.008 + (0.000218 × 7.0) 0.009 + (0.000228 × 6.3)
∑[ ] = 2
+
𝑑𝜆 2(0.008 + 8.0 × 0.000218) 2(0.009 + 8.0 × 0.000228)2
𝑖=1
0.007 + (0.000179 × 6.8)
+ = 152.4924
2(0.007 + 8.0 × 0.000179)2
From Eq. (43)
−48.1139
∆𝜆(1) = 152.4924 = −0.31552
Therefore, the new value of 𝜆 is
𝜆(2) = 8.0 − 0.31552 = 7.6845
The process is continued for the second iteration, i.e.,
(2) 7.6845 − 7.0
𝑃1 = = 35.3739 𝑀𝑊
2(0.008 + (7.6845 × 0.000218))

(2) 7.6845 − 6.3

𝑃2 = = 64.3830 𝑀𝑊
2(0.009 + (7.6845 × 0.000228))

(2) 7.6845 − 6.8

𝑃3 = = 52.8027 𝑀𝑊
2(0.007 + (7.6845 × 0.000179))
The real power loss is;
(2)
𝑃𝐿 = 0.000218(35.3739)2 + 0.000228(64.3830)2 + 0.000179(52.8027)2 = 1.7170 MW
Since PD = 150 MW, the error ∆𝑃(2) from Eq. (46) is
∆𝑃(2) = 150 + 1.7170 − (35.3728 + 64.3821 + 52.8015) = −0.8426
From Eq. (49)
3 (2)
𝑑𝑃𝑖 0.008 + (0.000218 × 7.0) 0.009 + (0.000228 × 6.3)
∑[ ] = 2
+
𝑑𝜆 2(0.008 + 7.6845 × 0.000218) 2(0.009 + 7.6845 × 0.000228)2
𝑖=1
0.007 + (0.000179 × 6.8)
+ = 154.588
2(0.007 + 7.6845 × 0.000179)2

From Eq. (43)

−0.8426
∆𝜆(2) = 154.588 = −0.005451
Therefore, the new value of 𝜆 is
𝜆(3) = 7.6845 − 0.005451 = 7.679
 Page 41 of 45

For the third iteration, we have

(3) 7.679 − 7.0
𝑃1 = = 35.0965 𝑀𝑊
2(0.008 + (7.679 × 0.000218))

(2) 7.679 − 6.3

𝑃2 = = 64.1369 𝑀𝑊
2(0.009 + (7.679 × 0.000228))

(2) 7.679 − 6.8

𝑃3 = = 52.4834 𝑀𝑊
2(0.007 + (7.679 × 0.000179))

The real power loss is;

(3)
𝑃𝐿 = 0.000218(35.0965)2 + 0.000228(64.1369)2 + 0.000179(52.4834)2 = 1.6995 MW
Since PD = 150 MW, the error ∆𝑃(3) from Eq. (46) is
∆𝑃(2) = 150 + 1.6995 − (35.0965 + 64.1369 + 52.4834) = −0.01742
From Eq. (49)
3 (3)
𝑑𝑃𝑖 0.008 + (0.000218 × 7.0) 0.009 + (0.000228 × 6.3)
∑[ ] = +
𝑑𝜆 2(0.008 + 7.679 × 0.000218)2 2(0.009 + 7.679 × 0.000228)2
𝑖=1
0.007 + (0.000179 × 6.8)
+ = 154.624
2(0.007 + 7.679 × 0.000179)2

From Eq. (43)

−0.01742
∆𝜆(3) = 154.624 = −0.001127
Therefore, the new value of 𝜆 is
𝜆 = 7.679 − 0.001127 = 7.6789
Since ∆𝜆(3) , is small equality constraint is then met in the fourth iterations, and the optimal dispatch for
𝜆 = 7.6789 are;

(3) 7.6789 − 7.0

𝑃1 = = 35.08890 𝑀𝑊
2(0.008 + (7.6789 × 0.000218))

(2) 7.6789 − 6.3

𝑃2 = = 64.1302 𝑀𝑊
2(0.009 + (7.6789 × 0.000228))

(2) 7.6789 − 6.8

𝑃3 = = 52.4746 𝑀𝑊
2(0.007 + (7.6789 × 0.000179))
The real power loss is;
(4)
𝑃𝐿 = 0.000218(35.08890)2 + 0.000228(64.1302)2 + 0.000179(52.4746)2 = 1.699 MW

Since PD = 150 MW, the error ∆𝑃(4) from Eq. (46) is

∆𝑃(2) = 150 + 1.699 − (35.08890 + 64.1302 + 52.474) = 𝟎. 𝟎𝟎𝟓𝟑
This is above the value hence we peg iteration at the fourth iteration.

Therefore, total fuel cost is given by

𝐶𝑡 = 200 + 7.0(35.08890) + 0.008(35.08890)2
+ 180 + 6.3(64.1302) + 0.009(64.1302)2
+ 140 + 6.8(52.4746) + 0.007(52.4746)2
= 455.4721 + 621.0344 + 516.1023655
 Page 42 of 45

= 1592.6088 𝑁/ℎ𝑟
Efficiency of the system is given as
𝑃𝑜𝑢𝑡𝑝𝑢𝑡
𝜂= × 100%
𝑃𝑖𝑛𝑝𝑢𝑡
𝑃𝐿𝑜𝑠𝑠
=1−( )
∑𝑖 𝑃𝑖

1.699 × 106
𝜂 =1− × 100%
(35.08890 + 64.1302 + 52.4746) × 106
= 𝟗𝟖. 𝟖𝟖%
Example 3:
Derive the loss function Pl and hence the B-coefficients for the network shown in Figure Example 3.

1 3 2
Unit 1 I1 I2 Unit 2

a b
Ra Rb

c Rc

Figure Example 3
Load
Solution
Computing losses in each line section adding together gives
𝑃𝐿 = 3{|𝐼1 |2 𝑅𝑎 + |𝐼2 |2 𝑅𝑏 + |𝐼1 + 𝐼2 |2 𝑅𝑐 } (1)

Assuming that 𝐼1 𝑎𝑛𝑑 𝐼2 are in phase, then

|𝐼1 + 𝐼2 | = |𝐼1 | + |𝐼2 |
∴ (|𝐼1 | + |𝐼2 |)2 = |𝐼1 |2 + 2|𝐼1 ||𝐼2 | + |𝐼2 |2 (2)
Substituting (2) into (1) gives
𝑃𝐿 = 3{|𝐼1 |2 (𝑅𝑎 + 𝑅𝑐 ) + |𝐼2 |2 (𝑅𝑏 + 𝑅𝑐 ) + 2|𝐼1 ||𝐼2 |𝑅𝑐 }
From power injected into Bus 1 and 2
𝑃1 𝑃
|𝐼1 | =
|𝑉 |.𝑝.𝑓
; |𝐼2 | = |𝑉 2|.𝑝.𝑓
√3 1 1 √3 2 2
Where 𝑝. 𝑓1; 𝑝. 𝑓2 were power factors at buses 1 and 2 respectively.

𝑃2 (𝑅 +𝑅 )
𝑎 𝑐 𝑃22 (𝑅𝑏 +𝑅𝑐 ) 2𝑃 𝑃 𝑅
1
𝑃𝐿 = [ (|𝑉 |.𝑝.𝑓 )2
+ (|𝑉2 |.𝑝.𝑓2 )2
+ (|𝑉 |.𝑝.𝑓1)(|𝑉
2 𝑐
|.𝑝.𝑓 )
]
1 1 1 1 2 2

𝑅𝑎 +𝑅𝑐
Let 𝐵11 = 2
(|𝑉1 |.𝑝.𝑓1 )
𝑐 𝑅
𝐵12 = 𝐵21 = (|𝑉 |.𝑝.𝑓 )(|𝑉 |.𝑝.𝑓 )
1 1 2 2
𝑅𝑏 +𝑅𝑐
𝐵22 = 2
(|𝑉2 |.𝑝.𝑓2 )
Therefore,
𝑃𝐿 = 𝑃12 𝐵11 + 2𝑃1 𝑃2 𝐵12 + 𝑃22 𝐵22
 Page 43 of 45

3 3

𝑃𝐿 = ∑ ∑ 𝑃𝑖 𝐵𝑖𝑘 𝑃𝑘
𝑖=1 𝑘=1
Principle of Equal Incremental Rate
Given a system that consists of two generators connected to a single bus serving a received electrical
load PD. The input - output characteristics of two generating units are F1(PG1) and F2(PG2), respectively.
The total fuel consumption of the system F is the sum of the fuel consumptions of the two generating
units. Assuming there is no power output limitation for both generators, the essential constraint on the
operation of this system is that the sum of the output powers must equal the load demand. The economic
power dispatch problem of the system, which is to minimize F under the above – mentioned constraint,
can be expressed as:
min F
min 𝐹 = 𝐹1 (𝑃𝐺1 ) + 𝐹2 (𝑃𝐺2 ) (1)
S.t.
𝑃𝐺1 + 𝑃𝐺2 = 𝑃𝐷 (2)
According to the principle of equal incremental rate, the total fuel consumption F will be minimal if the
incremental fuel rates of two generators are equal, that is,
𝑑𝐹1 𝑑𝐹2
= =𝜆 (3)
𝑑𝑃𝐺1 𝑑𝑃𝐺2
𝑑𝐹
where 𝑑𝑃 1 is the incremental fuel rate of generating unit i, which corresponds to the slope of the input -
𝐺1
output curve of the generating unit.

If two generators operate under the different incremental fuel rate, and
𝑑𝐹1 𝑑𝐹2
>
𝑑𝑃𝐺1 𝑑𝑃𝐺2
the total output powers maintain the same rate. If generator 1 reduces output power ∆P , generator 2 will
𝑑𝐹
increase output power ∆P . Then generator 1 will reduce fuel consumption 𝑑𝑃 1 ∆P, and generator 2 will
𝐺1
𝑑𝐹
increase fuel consumption 𝑑𝑃 2 ∆P. The total savings of fuel consumption will be
𝐺2

𝑑𝐹1 𝑑𝐹2 𝑑𝐹1 𝑑𝐹2

∆𝐹 = ∆𝑃 − ∆𝑃 = ( − ) ∆𝑃 > 0 (4)
𝑑𝑃𝐺1 𝑑𝑃𝐺2 𝑑𝑃𝐺1 𝑑𝑃𝐺2
𝑑𝐹 𝑑𝐹
It can be observed from equation (4) that ∆𝐹 will be zero when 𝑑𝑃 1 = 𝑑𝑃 2 that is, the incremental fuel
𝐺1 𝐺2
rates of two generators are equal.

Example 1
Suppose the input - output characteristics of two generating units which are as follows:
2
𝐹1 = 0.0008𝑃𝐺1 + 0.02𝑃𝐺1 + 5 𝐵𝑡𝑢/ℎ
2
𝐹2 = 0.0005𝑃𝐺2 + 0.03𝑃𝐺2 + 4 𝐵𝑡𝑢/ℎ
Determine the economic operation point for these two units when delivering a total of 500 MW power demand.

Solution
Firstly, obtain the incremental fuel rate of two generating units as follows:
𝑑𝐹1
𝜆1 = = 0.0016𝑃𝐺1 + 0.02
𝑑𝑃𝐺1
 Page 44 of 45

𝑑𝐹2
𝜆2 = = 0.001𝑃𝐺2 + 0.03
𝑑𝑃𝐺2
According to the principle of equal incremental rate we have
𝜆1 = 𝜆2
0.0016𝑃𝐺1 + 0.02 = 0.001𝑃𝐺2 + 0.03
When one multiply through by 1000
1.6𝑃𝐺1 − 𝑃𝐺2 = 10 (𝑖)
𝑃𝐺1 + 𝑃𝐺2 = 500 (𝑖𝑖)
Solving eqn. (i) &(ii) simultaneously gives
From Eqn. (i) 𝑃𝐺2 = 1.6𝑃𝐺1 − 10
Substitute in eqn.(ii) yields
𝑃𝐺1 + 1.6𝑃𝐺1 − 10 = 500
𝑃𝐺1 + 1.6𝑃𝐺1 = 510
2.6𝑃𝐺1 = 510
510
𝑃𝐺1 = = 196.154 𝑀𝑊
2.6
Solving for 𝑃𝐺2 from eqn. (ii)
196.154 + 𝑃𝐺2 = 500

𝑃𝐺2 = 303. 846 𝑀𝑊

The corresponding system incremental fuel rate under this load is obtained as follows:

𝑑𝐹1
𝜆1 = = 0.0016𝑃𝐺1 + 0.02
𝑑𝑃𝐺1
𝜆 = 0.0016(196.154 ) + 0.02

𝜆 = 0.3338

Exercises
1. The fuel function in N/hr. for two 800 MW thermal plants are given by
𝐶1 = 400 + 6.0𝑃1 + 0.004𝑃12
𝐶2 = 500 + 𝛽𝑃2 + 𝛾𝑃22
Where P1 and P2 were in MW
(a) The incremental cost of power 𝜆 is N8/ MWhr when the total power demand is 550 MW.
Neglecting losses, determine the optimal generation of each plant.
(b) The incremental cost of power 𝜆 is N10/MWhr when the total power demand is 1300 MW. Neglecting
losses, determine the optimal generation of each plant.
(c) From the results of (a) and (b), find the fuel cost coefficient 𝛽 𝑎𝑛𝑑 𝛾 of the second plant.
(d) The fuel cost functions in N/hr for three thermal plants are given by

𝐶1 = 350 + 7.20𝑃1 + 0.0040𝑃12

𝐶2 = 500 + 7.30𝑃2 + 0.0025𝑃22
𝐶3 = 600 + 6.74𝑃3 + 0.0030𝑃32
Where 𝑃1 , 𝑃2 𝑎𝑛𝑑 𝑃3 were in MW. The governors were set such that generators share the load
equally. Neglecting line losses and generator limits, find the total cost in N/hr when the total load
is
(𝑖) 𝑃𝐷 = 450 𝑀𝑊 ; (𝑖𝑖) 𝑃𝐷 = 745 𝑀𝑊; (𝑖𝑖𝑖) 𝑃𝐷 = 1335 𝑀𝑊
2. The fuel cost function in N/hr of two thermal plants were
 Page 45 of 45

𝐶1 = 320 + 6.2𝑃1 + 0.004𝑃12

𝐶2 = 200 + 6.0𝑃2 + 0.003𝑃22
Where 𝑃1 , 𝑃2 𝑎𝑛𝑑 𝑃3 were in MW. Plant outputs were subjected to the following limits in MW
50 ≤ 𝑃1 ≤ 250
50 ≤ 𝑃2 ≤ 350
The per unit system real power loss with generation expressed in per unit on a 100 MW base is
given by
𝑃𝐿 (𝑝. 𝑢) = 0.0125𝑃12 (𝑝. 𝑢) + 0.00625𝑃22 (𝑝. 𝑢)
The total load is 412.35 MW. Determine the optimal dispatch of generation. Start with an initial
estimation of 𝜆 = 7 𝑁/𝑀𝑊ℎ𝑟
3. Figure P3 shows a system having two plants 1 and 2 connected to buses 1 and 2, respectively.
There were two loads and a network of three branches. Bus 1 is the reference bus with voltage of
1.0∠0° 𝑝. 𝑢. the branch currents and impedances were given as follows:
𝐼𝑎 = 2 − 𝑗0.5 𝑝. 𝑢
𝐼𝑏 = 1.6 − 𝑗0.4 𝑝. 𝑢
𝐼𝑐 = 1.8 − 𝑗0.45 𝑝. 𝑢
𝑍𝑎 = 0.06 + 𝑗0.24 𝑝. 𝑢
𝑍𝑏 = 0.03 − 𝑗0.12 𝑝. 𝑢
𝑍𝑐 = 0.03 − 𝑗0.12 𝑝. 𝑢 calculate then loss formula coefficients of the system in per unit and in
reciprocal megawatts, if the base is 100 MVA.

1 Ref. bus 2

I1 I2
V1=1 00 a

Plant 1 Plant 2
b c

Figure P3 Load 1 Load 2

4. Suppose the input - output characteristics of three generating units are as follows:
2
𝐹1 = 0.0006𝑃𝐺1 + 0.5𝑃𝐺1 + 6 𝐵𝑡𝑢/ℎ
2
𝐹2 = 0.0005𝑃𝐺2 + 0.6𝑃𝐺2 + 5 𝐵𝑡𝑢/ℎ
2
𝐹3 = 0.0007𝑃𝐺3 + 0.4𝑃𝐺3 + 3 𝐵𝑡𝑢/ℎ
Determine the economic operation point for these three units when delivering a total of 500 MW
and 800 MW power demand, respectively.