Wright State University

CORE Scholar
Browse all Theses and Dissertations Theses and Dissertations

2008

A Stochastic Production Planning Model Under

Uncertain Demand
Meenakshi Prajapati
Wright State University

Follow this and additional works at: [Link]

Part of the Operations Research, Systems Engineering and Industrial Engineering Commons

Repository Citation
Prajapati, Meenakshi, "A Stochastic Production Planning Model Under Uncertain Demand" (2008). Browse all Theses and Dissertations.
Paper 885.

This Thesis is brought to you for free and open access by the Theses and Dissertations at CORE Scholar. It has been accepted for inclusion in Browse all
Theses and Dissertations by an authorized administrator of CORE Scholar. For more information, please contact
corescholar@[Link].
A Stochastic Production Planning Model

Under Uncertain Demand

A thesis submitted in partial fulfillment

of the requirements for the degree of the
Master of Science in Engineering

by

MEENAKSHI PRAJAPATI

B. E., Jai Narain Vyas University, 2003

M. E., Malaviya National Institute of Technology, 2005

2008
Wright State University
 Wright State University

SCHOOL OF GRADUATE STUDIES

November 6, 2008

I HEREBY RECOMMEND THAT THE THESIS PREPARED UNDER

MY SUPERVISION BY MEENAKSHI PRAJAPATI ENTITLED A Stochastic

Production Planning Model Under Uncertain Demand BE ACCEPTED IN

PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DE-

GREE OF Master of Science in Engineering

Dr. Xinhui Zhang

Thesis Director

Dr. S. Narayanan
Department Chair

Committee on
Final Examination

Dr. Xinhui Zhang

Dr. Frank Ciarallo

Dr. Yan Liu

Joseph F. Thomas, Jr., Ph.D.

Dean, School of Graduate Studies
Abstract

Prajapati, Meenakshi. [Link]., Department of Biomedical, Industrial, and

Human Factors Engineering, Wright State University, 2008.
A Stochastic Production Planning Model Under Uncertain Demand.

Production planning plays a vital role in the management of manufactur-

ing facilities. The problem is to determine the production loading plan-

consisting of the quantity of production and the workforce level - to fulfill a

future demand. Although the deterministic version of the problem has been

widely studied in the literature, the stochastic production planning problem

has not. The application of production planning models could be limited if

the stochastic nature of the problem, for example, uncertainty in future de-

mand, is not addressed. This study addresses such a stochastic production

planning problem under uncertain demand and its application in an enclo-

sure manufacturing facility.

The thesis first addresses the forecast of the demand where seasonal fluc-

tuation is present. A decomposition model is utilized in the forecast and

iii
compared with other forecasting methods. Although forecast models could

be used to improve the accuracy of forecast, error and uncertainty still ex-

ists. To deal with this uncertainty, a two stage stochastic scenario based

production planning model is developed to minimize the total cost consist-

ing of production cost, labor cost, inventory cost and overtime cost under

uncertain demand.

The model is solved with data from a local manufacturing facility and the

results are compared with various deterministic production models to show

the effectiveness of the developed stochastic model. Parametric analyses are

performed to derive managerial insights related to issues such as overtime

usage and inventory holding cost and the proper selection of scenarios un-

der pessimist, neutral and optimist forecasts. An extension of the stochastic

model, i.e., a robust model is also solved in an effort to minimize changes

in the solutions under various scenarios.

The stochastic production planning model has been implemented in the

manufacturing facility, provided guidance for material acquisition and pro-

duction plans and has dramatically increased the company’s bottom line.

As a result, it’s estimated an approximately annual savings of $340,000 in

inventory cost can be achieved for the company in the next few years.

iv
Contents

1 Introduction 1

1.1 Production Planning in Manufacturing Environment . . . . . 1

1.2 Motivation and Contribution . . . . . . . . . . . . . . . . . 3

1.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . 4

2 Literature Survey 6

2.1 Deterministic Demand Production Planning . . . . . . . . . 6

2.2 Stochastic Programming . . . . . . . . . . . . . . . . . . . 9

2.2.1 Scenario Analysis . . . . . . . . . . . . . . . . . . . 10

2.2.2 Two-stage Stochastic Production Planning Model . . 10

2.2.3 Multi-Scenario Stochastic Model . . . . . . . . . . 13

3 Demand Forecast 18

3.1 Seasonal Decomposition/Time-Series Model . . . . . . . . . 19

3.2 Forecast Demand and Results . . . . . . . . . . . . . . . . . 21

4 Stochastic Production Planning Models and Its Formulation 29

4.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . 30

v
 4.2 A Two-stage Stochastic Production Planning Model . . . . . 33

4.2.1 Mathematical Model . . . . . . . . . . . . . . . . . 33

4.3 Robust Production Planning Formulation . . . . . . . . . . . 37

5 Computational Results 41

5.1 Input Parameters . . . . . . . . . . . . . . . . . . . . . . . 41

5.2 Optimal Production Plan and Cost Analysis . . . . . . . . . 43

5.2.1 Solution Under Perfect Information . . . . . . . . . 43

5.2.2 Two-Stage Stochastic Solution with the Same Four

Month Production . . . . . . . . . . . . . . . . . . 44

5.2.3 Two-Stage Stochastic Model with the Same 12 Month

Production . . . . . . . . . . . . . . . . . . . . . . 45

5.2.4 Expected Value Solution . . . . . . . . . . . . . . . 46

5.2.5 Solution Under Expected Demand . . . . . . . . . . 46

5.3 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . 48

5.3.1 Change of Probability Associated with the Scenar-

ios . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.3.2 Change from Scenario Generation Method . . . . . 50

5.3.3 Change of Overtime Ratio . . . . . . . . . . . . . . 52

5.3.4 Change of Inventory Holding Cost . . . . . . . . . . 53

5.4 Robust model . . . . . . . . . . . . . . . . . . . . . . . . . 55

6 Conclusions and Implementation 57

6.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 57
vi
6.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . 58

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

vii
List of Figures

3.1 Observed Value of Series . . . . . . . . . . . . . . . . . . . 20

3.2 Trend of Series . . . . . . . . . . . . . . . . . . . . . . . . 20

3.3 Seasonal Component of Series . . . . . . . . . . . . . . . . 20

3.4 Random Component of Series . . . . . . . . . . . . . . . . 20

3.5 Forecast Comparison of Company Data . . . . . . . . . . . 24

5.1 Mean and Variance of two random samples . . . . . . . . . 53

5.2 Sensitivity Analysis-Overtime Ratio Varies . . . . . . . . . 54

viii
List of Tables

3.1 Demand Data & Decomposition Forecast Results for 2007 . 26

3.2 Forecast Result Comparison with Company’s Judgmental

Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.3 Forecast Errors of the Models . . . . . . . . . . . . . . . . . 27

3.4 Forecast Data Comparison with SAS Results . . . . . . . . 27

3.5 Forecast Data & Results of the Decomposition Model for

September 2007 to August 2008 . . . . . . . . . . . . . . . 28

5.1 Company’s Demand Data of the Scenarios . . . . . . . . . . 42

5.2 Deterministic Production Planning Model . . . . . . . . . . 43

5.3 Two-Stage Stochastic Model Production Plan (4 Month) . . 44

5.4 Two-Stage Stochastic Model Production Plan with Same 12

Month Production . . . . . . . . . . . . . . . . . . . . . . . 45

5.5 EVIP and VSS Value for the Two Two-Stage Stochastic

Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5.6 Inventory and Overtime Based on Expected Value . . . . . . 47

ix
5.7 Production Plan Based on Expected Demand . . . . . . . . 47

5.8 Inventory and Overtime for All Scenarios Based on Ex-

pected Demand Production . . . . . . . . . . . . . . . . . . 48

5.9 Cost Comparison of Various Models . . . . . . . . . . . . . 49

5.10 Probability Associated with Each Scenario . . . . . . . . . . 49

5.11 Sensitivity Analysis- Change of Probability Distribution . . 50

5.12 Inventory, Overtime and Production Under Case II . . . . . 50

5.13 Inventory, Overtime and Production Under Case III . . . . . 51

5.14 Inventory, Overtime and Production Under 3 Random Sce-

narios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5.15 Sensitivity Analysis-Overtime Ratio Varies . . . . . . . . . 52

5.16 Sensitivity Analysis-Inventory Holding Cost Varies . . . . . 54

5.17 Inventory, Overtime and Production with Inventory Holding

Cost 20% . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5.18 Inventory, Overtime and Production with Inventory Holding

Cost 30% . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5.19 Robust Model Production Plan . . . . . . . . . . . . . . . . 56

5.20 Cost Difference of Various Models . . . . . . . . . . . . . . 56

6.1 Production Plan From September 2008 to August 2009 . . . 58

x
 Acknowledgement
I would like to thank my thesis committee members, Dr. Xinhui Zhang,

Dr. Frank Ciarallo and Dr. Yan Liu for their time, suggestions and valuable

comments. I would also like to thank Dr. Xinhui Zhang for the graduate

scholarship.

This thesis is a celebration of my cordial relationship with Dr. Xinhui

Zhang as my adviser, mentor and as a friend. I have benefited intellectually

from his piercing insights and tastes for good research. His perpetual en-

couragement and motivation have illuminated my graduate life.

I am immensely indebted to my parents for their nurture, upbringing

and unconditional love. I am also thankful to my sisters Sarla and Renu and

my parents-in-law for their moral support and encouragement.

Finally, I would like to express my heartfelt gratitude, love and admi-

ration to my husband, Mr. Gulshan Singh. Without his incredible support,

encouragement, and loving kindness, I would not have made it through the

program so smoothly. We have made it together as a family.

xi
 Dedicated to:

My Husband and My Parents

xii
Chapter 1

Introduction

1.1 Production Planning in Manufacturing Environment

Manufacturing planning helps direct the acquisition and allocation of avail-

able resources to production activities so as to satisfy customer demand over

a specific time horizon. The production planning problem aims to match

production and sourcing decisions to meet future customer demand subject

to production capacity, workforce availability and overtime restrictions and

is inherently an optimization problem. The objective of the problem is to

minimize the total cost or to maximize profit.

Mathematical models for production planning problems can be broadly

classified into two categories: deterministic models and stochastic models.

Deterministic models assumes that the data are known and typically model

the uncertainty using ”best guesses” of uncertain values. Although vari-

ous human judgment based and quantitative models have been developed to

forecast these variables with uncertainty such as demand, these determin-

istic models typically end up solving ”mean-value” or ”worst-case” prob-

1
lems. The solution to such ”worst-cast” or ”mean-value” problems are of-

ten inadequate – large error bounds arise when one solves ”mean value”

problems and ”worst-case” formulations that can produce very conservative

and expensive solutions (Birge (1982)). Without considering uncertainty,

the deterministic production planning models, though widely studied in the

literature, are less acceptable and deployed in practice. For some examples,

please see Qiu and Burch (1997) and Tabucanon (1988).

Several uncertainties in manufacturing exist and can be categorized into

two categories: environmental uncertainty and system uncertainty. The for-

mer refers to uncertainties that are beyond the scope of control of the pro-

duction process, such as supply and demand uncertainty, while the latter

refers to uncertainties that relate to the production process, such as opera-

tion yield uncertainty, production lead time uncertainty, quality uncertainty

and production failures. Though various analytic and simulation models

have been proposed, see Mula et al. (2006) for a survey, performing an all-

inclusive study of production planning with all the uncertainties is almost

impossible – simulation models are descriptive in nature and generally do

not provide optimal solutions. In the case of production planning, a more

viable approach is to identify and address one or a few uncertainties within

a model which can be used to derive optimal solutions. This is the focus of

this study.

2
1.2 Motivation and Contribution

This thesis focuses on the stochastic production planning with uncertain

demand. The problem arises in a local enclosure manufacturing facility

producing thermal cooling units where seasonal fluctuation in demand and

uncertainty in company growth are present. Seasonal demand creates more

complex production problems than stationary demand because firms that

face seasonal demands typically do not have sufficient capacity to meet de-

mand in the high demand season and must build inventory in the low de-

mand season in anticipation of high demand later (Krane and Braun (1991)

and Fair (1989)). The uncertainty of company and market growth raises

even more uncertainty in the demand of products.

To make things more complicated, the company purchase raw material

and semi-finished goods from overseas and a lead time of 40-45 working

days is the norm. With more manufacturing are being outsourced overseas,

the long lead time in purchasing materials has a dramatic impact on produc-

tion planning – purchasing decisions and production plans for future (3∼4)

months has to be made before demand is realized. The uncertainty in de-

mand and company growth calls for the solution of a stochastic production

planning problem. The goal is to develop a production plan for the next year

to minimize the expected total inventory, overtime and workforce costs by

considering the manufacturing capacity and workforce level under different

3
scenarios.

This thesis develops a two stage stochastic production planning model

that yields solutions which serve as the baseline for different economic

growth scenarios. The model is solved with real data from a local com-

pany and compared with deterministic production models on the basis of

robustness and effectiveness. A parametric analysis is used to derive man-

agerial insights related to issues such as overtime usage, inventory holding

cost and the proper selection of scenarios under pessimistic, neutral and op-

timistic forecasts. An extension of the stochastic model i.e., a robust model

is also formulated and solved in an effort to minimize changes in the solu-

tions under various scenarios.

The stochastic production planning model is currently implemented in

the manufacturing facility, provided guidance for material acquisition and

production plans and has significantly increase the company’s bottom line.

It’s estimated an approximately annual savings of $340,000 in inventory can

be achieved for the company in the next few years.

1.3 Thesis Organization

This thesis is organized as follows. Chapter 2 reviews the relevant literature.

Chapter 3 presents the decomposition forecast model and comparison with

the company’s current human judgmental forecast. Chapter 4 describes the

two stage stochastic production planning model, followed by a robust pro-

4
duction planning model. Chapter 5 presents the computational results using

data from the company. The results show the effectiveness and efficiency of

the stochastic model as compared with the deterministic model. The various

parametric analysis have been preformed. Finally, conclusions and imple-

mentation are outlined in Chapter 6.

It is noteworthy that although the stochastic programming has been suc-

cessfully applied on a wide variety of problems ranging from financial plan-

ning (Carino et al. (1998)), airline crew scheduling (Schaefer et al. (2005)),

the solution to production planning using stochastic programming has only

recently received some attention (Karabuk and Wu (2008)). While most of

these studies focus on the strategic issues such as capacity planning, this

thesis focuses more on an operational production planning model with long

purchasing lead time that has not been addressed in the literature.

5
Chapter 2

Literature Survey

This chapter presents an overview of the literature on deterministic produc-

tion planning, stochastic programming and stochastic production planning

problems.

2.1 Deterministic Demand Production Planning

The classical deterministic production planning problem has been the focus

of and received a significant attention in the optimization literature. For a

recent survey, please see Fogarty et al. (1984), for some of the well-known

models, their computational characteristics and solutions approaches. The

basic mathematical model for deterministic production planning, also known

as the lot sizing problem, can be formulated as follows.

Parameters
6
 f (t) fixed cost of producing in period t

C p (t) unit production cost in period t

Ch (t) storage cost in period t

C manufacturing capacity

D(t) demand in period t

Decision variable

P(t) amount produced in period t

I(t) inventory level at the end of period t

y(t) 1 if production occurs in t; 0 otherwise

Deterministic Production Planning or Lot Sizing Model

Minimize
n n n
∑ CpP(t) + ∑ f (t)y(t) + ∑ ChI(t) (2.1)
t=1 t=1 t=1

Subject to

• Inventory balance constraints

I(t − 1) + P(t) = D(t) + I(t) ∀t = 1..n (2.2)

• Capacity constraints

P(t) <= Cy(t) ∀t = 1..n (2.3)

7
 • Non-negativity and integer constraints

P(t) ≥ 0, y(t) ∈ {0, 1} ∀t = 1..n (2.4)

The objective (2.1) here is to minimize the total cost including the labor

cost per year (term 1), production cost per year (term 2) and inventory cost

(term 3). Constraint (2.2) ensures that the available inventory in any period

plus the demand in that period equals to the summation of inventory from

the previous period and production during the current period. Constraint

(2.3) states the capacity limitation while constraint (2.4) ensures that the

decision variables P(t) is non-negative and y(t) is binary. The determinis-

tic production planning problem under capacity limitation is proven to be

a NP-hard problem (Garey and Johnson (1979)). Deterministic production

planning (Swoveland (1975)) has received much attention from the litera-

ture, some applications are listed below.

Homburg (1996) modeled a production planning model for decentral-

ized organization. The model was formulated as a linear program with

multiple objectives. Lee et al. (2005) used production planning model to

develop an optimal production schedule for a manufacturer of hard-disk

drives. Liu and Tu (2008) solve a production planning problem where the

production quantity is limited by inventory storage capacity. The problem

further assumes that stock outs are allowed and that inventory storage capac-

ity is constant. Huang and Xu (1998) transformed the multi-stage, multi-

8
item aggregate scheduling problem into a static job-assignment problem.

They used an adapted Frank Wolfe algorithm to solve their modified prob-

lem. Qiu and Burch (1997) modeled a hierarchical production planning

problem for a yarn-fiber facility. Kanyalkar and Adil (2005) formulated a

linear programming model for an integrated multi-item, parallel multi-plant

production and dynamic distribution problem. It also considered safety

stocks and capacity aggregation to minimize the effect of uncertainties in

demand and supply.

The deterministic production planning model and its solution quality

depends heavily on the accuracy of forecasted demand requirements – a

difficult goal in a noisy and uncertain information environment (see Tabu-

canon (1988)). As mentioned, ”average demand” or ”worse case” demand

estimates often leads to inferior solutions. To model uncertainty, stochastic

production planning have to be developed.

2.2 Stochastic Programming

Stochastic programs, originated in the late fifties with the pioneering works

of Dantzig (1955), Beale (1955), and Birge (1959), are mathematical pro-

grams where some of the data in the objective or constraints are uncertain.

Uncertainty is usually characterized by a probability distribution on the pa-

rameters. Although the uncertainty is rigorously defined, in practice it can

9
range in detail from a few scenarios to specific probability distributions.

2.2.1 Scenario Analysis

Rockafellar and Wets (1991) introduced a method for optimizing against an

uncertain future. This technique, called scenario analysis, requires the user

to specify a finite number of scenarios, each representing a different envi-

ronment for the time period considered. Rockafellar and Wets (1991) then

posed the problem of optimizing the expected objective value of this collec-

tion of problems: that is, the sum of the objective values of the individual

scenario problems, each multiplied by the probability of that scenario’s oc-

currence. To analyze how the optimal solution would change with respect

to different scenarios, scenario analysis can be carried out by solving a set

of optimization models with different sets of data as well as the blending of

all of the scenario-dependent solutions into one solution.

2.2.2 Two-stage Stochastic Production Planning Model

Each decision and its execution can be subdivided into several stages, so that

the stochastic problems will represent a multi-stage optimization problem.

The two-stage stochastic model makes decisions using a two stage frame-

work. The first stage decision variables are optimized before a realization

of the uncertain and random variables. After realization of the random vari-

ables, the second stage variables are optimized.

The first stage decision variables are called the structural component
10
that is fixed in the second stage and free of any uncertainty in its input data.

The second stage decision variables are called the control component that is

subjected to uncertain input data.

In this study, two sets of variables are introduced to define the two-stage

stochastic model:

x, denotes the vector of decision variables whose optimal value is not

conditioned on realization of the uncertain parameters. These are design

variables. Variables in this set cannot be adjusted once a specific realization

of the data is observed.

y, denotes the vector of control decision variables that are subjected to

adjustment once the uncertain parameters are observed. Their optimal value

depends both on the realization of uncertain parameters, and on the optimal

value of the design variables.

The definition of design and control variables is borrowed from the flex-

ibility analysis of production and distribution processes Seider et al. (1991).

In a manufacturing environment, a facility location, facility capacity, sup-

plier selection, product purchasing decisions and demand realization comes

in phases. Depending on the problem and context, a two-stage models can

be defined in several ways. For example, the first stage could incorporate

variables related to the location and capacity of the facility such as work-

force size, production capacity and the second stage could represent how,

given these decisions, the detailed production plans should be carried out

11
or inventory levels managed under the realized demand. This results in a

strategic production capacity planing model or the stochastic capacity plan-

ning model as seen in Eppen et al. (1989) and Karabuk and Wu (2008).

The general form of the two-stage stochastic programming model is ex-

pressed as follows:

Minx cT x + E[Q(x, ξ)] (2.5)

s.t Ax = b (2.6)

x≥0 (2.7)

Where Q(x, ξ) is the optimal value of the second stage problem

Miny q(ω)T y (2.8)

s.t T (ω)x +Wy = h(ω) (2.9)

y≥0 (2.10)

The second stage problem depends on the data ξω ≡ (q(ω), h(ω), T (ω)),

elements of which can be random, while the matrix W is assumed to be

known beforehand. The matrices T (ω) and W are called the technology

and recourse matrices, respectively. The expectation E[Q(x, ξ)] is taken

with respect to the random vector ξ = ξ(ω) whose probability distribution

is assumed to be known. The above formulation originated in the works

of Dantzig (1955) and Beale (1955). Equation (2.6) denotes the structural

constraints whose coefficients are fixed and free of uncertainty, so they are

processed in the first stage. Equation (2.9) denotes the control constraints
12
and they are processed in the second stage. Therefore, Equations (2.5)-

(2.7) represent the first-stage model and Equations (2.8)-(2.10) represent the

second-stage model. c is the vector of cost coefficient at the first stage. A is

the first-stage coefficient matrix and b is the corresponding right-hand side

vectors. q is the vector of cost (recourse) coefficient vectors at the second

stage. W is the second-stage (recourse) coefficient matrix and h(ω) is the

corresponding right-hand side vector. T (ω) is the matrix that ties the two

stages together. In the second-stage model, the random constraint defined in

(2.9), h(ω) − T (ω)x, is the goal constraint: violations of this constraint are

allowed, but the associated penalty cost, qT y, will influence the choice of x.

2.2.3 Multi-Scenario Stochastic Model

Consider the special case when random data have a discrete distribution

with a finite number K of possible realizations ξk = (qk , hk , Tk ), k = 1, ..., K,

(called scenarios), with the corresponding probabilities pk . In that case

E[Q(x, ξ)] = ∑K
k=1 pk Q(x, ξk ), where

Q(x, ξk ) = Min qTk yk : Tk x +Wyk , yk ≥ 0 (2.11)

The above two-stage problem can be formulated as:

K
Min c x + ∑ pk (qTk yk )
T
(2.12)
k=1
s.t Ax = b (2.13)

Tk x +Wyk = hk s = 1, ...., S (2.14)

x, yk ≥ 0 k = 1, ...., K (2.15)
13
The vector of first-stage decision variables, x, is scenario independent. The

vectors of second-stage decision variables, yk are introduced to control the

random constraints with minimal recourse penalty cost.

Stochastic programming and scenario analysis has been applied in a

number of different areas, such as financing (Carino et al. (1994), Carino

et al. (1998), Carino and Ziemba (1998) and Consigli and Dempster (1998)),

agriculture planning (Maatman et al. (2002)), logistics planning (Dempster

et al. (2000)) and airline crew scheduling (Schaefer et al. (2005) and Yen

and Birge (2001)). In the manufacturing domain, several stochastic pro-

gramming models have been proposed on capacity planning and production

planning.

Eppen et al. (1989) studied the capacity planning issues for a General

Motors plants. They formulated a model to aid in making decisions about

capacity for four of GM’s auto lines. The model incorporates elements of

scenario planning, integer programming, and risk analysis.

Escudero and Kamesam (1955) presented two approaches for modeling

the aggregate production planning problem under stochastic non-stationary

demand, capacity constraints and dual sourcing.

Kira et al. (1997) proposed a stochastic linear programming approach

for hierarchical production planning under uncertain demand. They demon-

strated that the shape of the distribution that describes demand is an impor-

tant component in production decision-making.

14
 Karabuk and Wu (2003) develop a strategic capacity planning model

for a major semi-conductor manufacturer. They formulate a multi-stage

stochastic program with recourse where demand and capacity uncertainties

are incorporated via a scenario structure.

Karabuk and Wu (2008) develop a stochastic programming model that

explicitly includes uncertainty in the form of discrete demand scenarios for

a production planning problem in textile manufacturing. The paper describe

the development of the model and illustrate its application with numerical

examples.

Alonso-Ayuso et al. (2003) presented a modeling framework for two-

stage production planning under uncertainty in the main parameters.

Clay and Grossmann (1997) solves a stochastic linear programming

models for production planning where cost coefficient and RHS term uncer-

tainties are considered. The decisions are taken in two-stages for deciding

production levels.

Escuderoa et al. (1999) proposes a two-stage model for a manufactur-

ing, assembly and distribution supply chain planning problem under uncer-

tainty in product demand, component supplying cost and delivery time.

Hill and Sawaya (2004) presents a stochastic model for finding the op-

timal dates to stop production of an existing product and to start production

of a new product in the presence of an uncertain approval date of medical

devices.

15
 Fleten and Kristoffersen (2008) develop a short-term multi-stage stochas-

tic production plan for a price-taking hydro-power plant operating under un-

certainty.

Feiring and Sastri (1990) includes an aggregate production planning model

of manufacturing resources in order to satisfy stochastic demand for a fam-

ily of products to minimize total costs that include production and inventory

holding costs over a rolling horizon. The demand in this paper is assumed

to be normally distributed.

Leung et al. (2006) addresses the production planning problem with ad-

ditional constraints, such as production plant preference selection. To deal

with the uncertain demand data, a stochastic programming approach is pro-

posed to determine optimal medium-term production loading plans under an

uncertain environment. A set of data from a multinational lingerie company

in Hong Kong is used to demonstrate the robustness and effectiveness of the

proposed model.

Chen et al. (2002) addressed the solution of linear stochastic planning

problems. In this paper, they consider the role of product mix flexibility,

defined as the ability to produce a variety of products, in an environment

characterized by multiple products, uncertainty in product life cycles and

dynamic demands. Using a scenario based approach for capturing the evo-

lution of demand, they develop a stochastic programming model for deter-

mining technology choices and capacity plans.

16
 Reaychen and Fang (2001) solve the production planning problem where

unit cost to subcontract, work force level, production capacity and market

demands are uncertain and random.

17
Chapter 3

Demand Forecast

An accurate forecast of demand is critical to an effective operation of a

manufacturing facility. It affects the decisions regarding the capacity of the

facility, the workforce size and composition, the production plans and the

setting of safety stock. For production planning, it serves as an input to

an optimization model and significantly affects the final result – an inaccu-

rate forecast results in unreliable production plans with over- or under-stock

problems. The set of data provided to us by the manufacturing facility re-

veals a strong seasonality in the market. When a demand is highly seasonal,

it is unlikely an accurate forecast can be obtained without the use of an ap-

propriate model.

There are generally two families of forecasting methods: quantitative

and qualitative. Quantitative methods are those which rely on quantitative

data such as previous sales figures to make predictions about the future. The

forecast methods in this group include Time-Series and Causal models and

are typically used for middle term forecast, from a few months to a year.
18
Qualitative methods mainly rely on subjective data gathered from execu-

tives, sales team and consumers that predict future economic conditions for

a business.

In this study, a quantitative decomposition model is used to predict the

seasonal demand for the manufacturing company to find the trend and sea-

sonality factors. The forecast model was implemented in Microsoft Excel

and was able to obtain reasonable results with a Mean Absolute Percent-

age Error (MAPE) of 7%. These results are comparable to those provided

by the Statistics Analysis System (SAS) software package. This forecast

was combined with subjective forecasts from the sales team to derive pes-

simistic, neutral and optimistic scenarios used in the stochastic production

planning model.

3.1 Seasonal Decomposition/Time-Series Model

In time series analysis, demand data are considered as the combination of a

systematic pattern and some random noise superimposed to that pattern. A

forecasting approach can be improved if the underlying factors of a data pat-

tern can be identified and forecasted separately. Breaking down the data into

its component parts is called decomposition. A decomposition approach to a

forecast model typically assumes that sales are affected by four factors: the

general trend, general economic cycles, seasonality and irregular or random

occurrences. In most analyses, the random noise is considered zero-mean,

19
white and Gaussian. A forecast model need to find these separate compo-

nents - trend, seasonal, cyclical and random - and combine then to provide

the systematic pattern for a time series (Box et al. (1976) and Yaffee and

McGee (2000)).

1400 1100

1000
1200

900
1000
Demand Data

800

Trend
800
700

600
600

400
500

200 400
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Time Period Time Period

Figure 3.1: Observed Value of Series Figure 3.2: Trend of Series

1.4 150

1.3
100

1.2

50
1.1
Seasonal

Random

1 0

0.9
−50

0.8

−100
0.7

−150
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30
Time Period Time Period

Figure 3.3: Seasonal Component of Series Figure 3.4: Random Component of Series

Specifically the components can be represented by:

• Trend-Cycle(TCt ): A trend is a general long-term pattern observed over

the entire data set. The cycle is a cyclical variations around the trend

line.

• Seasonal(St ): A seasonal component is a cyclical variation with a fixed

20
 frequency, such as yearly cycles in the sales of lawnmowers. Seasonal

effects repeat on a regular and predictable basis.

• Random(Rt ): A random component is a residual component left behind

after the other three components have been accounted for.

The time series Yt can be a function of these components:

Yt = f (TCt , St , Rt ) (3.1)

where function f may take several forms:

Additive Yt = TCt + St + Rt (3.2)

Multiplicative Yt = TCt × St × Rt (3.3)

Pseudo-Additive Yt = TCt (St + Rt − 1) (3.4)

Figures (3.1) to (3.4) illustrate how a series may be decomposed into three

components. In this study, a multiplicative model is used for the forecast of

the company’s data.

3.2 Forecast Demand and Results

The company manufactures a variety of products in the thermal cooling unit

group, however, individual product in the group shares sub-assemblies or

components with other products and follows a similar seasonal demand pat-

tern. As a result, it is more effective to forecast the aggregate demand of

the thermal cooling products as a group. The forecasting of an aggregate

21
demand is not only more accurate than that of the individual product (Nah-

mias (1993)), but also more suited for production planning purpose, since

these products share components and subassemblies.

Demand Data and Forecast Calculation: The aggregate demand data for

the past two and a half years of the manufacturing company from January

2005 to June 2007 are presented in Table (3.1). Here the first column repre-

sents the time period and the second column the demand in that period.

The decomposition forecast is developed and implemented in Microsoft

Excel to find the trend and seasonality associated with data.

• Based on the data from January, 2005 to December, 2006, the values

of intercept and slope are calculated as 417.5 and 13.58 respectively by

using the INTERCEPT and SLOPE function embedded in Microsoft

Excel. The trend of the series is then calculated as:

Trend = Intercept + Slope × TimePeriod

• The seasonal factors are calculated by dividing the consumption (ac-

tual demand) value by trend value. The average seasonal factors are

calculated as the average of seasonal factors in the respective months.

• Finally, the forecast is calculated by multiplying the trend with the av-

erage seasonal factors.

The trend, seasonal factors, the average factor and the forecast of 2007 are

listed in Table (3.1) under the corresponding columns.

22
Decomposition Forecast versus Company Forecast: To evaluate the ac-

curacy of the forecast model, we adopt the absolute percentage error (APE)

as the measure of the accuracy. The absolute percentage error (APE) is

defined as:
kActual Demand − Forecast Datak
APE =
Actual Demand
Table (3.2) gives the forecast results and the absolute percentage errors

of the decomposition forecast as well as the company’s judgmental forecast

for the first six month of 2007. The decomposition forecast model was

providing reasonable good results.

As we can see from Table (3.2), the maximum APE of the decomposi-

tion forecast model was 17%, with an average of 7%. The maximum APE of

the company’s judgement forecast was 38% with an average APE of 23%.

As a side story, at the end of June 2007, we decided to re-run the model us-

ing data from January, 2005 to June, 2007 and updated the forecast for July

and August of 2007. The updated forecast for July 2007 was 878 units and

the real demand was 881 units – a 3 unit difference. The updated forecast

for August 2007 was 994 units and the real demand was 991 units – a 13

unit difference. The forecast has gained the trust of the company.

Figure (3.5) shows the results of the various forecast models from Jan-

uary 2005 to June 2008. Here, the line with plus sign represents the actual

demand, the dotted line the company’s forecast, and the line with circle the

decomposition forecast. As we can see, the decomposition model forecast,

23
the line with circle, matches closely with actual demand.

1600

1400
Company Judgemental Forcast
1200 Actual Demand
Seasonal Decompotion Forcast

1000

800

600

400

200
0 5 10 15 20 25 30 35 40 45 50
Time Period in Months from 2005 to 2008

Figure 3.5: Forecast Comparison of Company Data

The Decomposition Forecast versus SAS Forecast: To see how the de-

composition forecast model compares with other forecast models, we input

the same demand data from January 2005 to December 2006 into Statistics

Analysis System (SAS) software package. SAS uses different types of fore-

cast models and the various forecast results and their mean square errors

provided are listed in (3.3). For details of these modeles, please refer to

(Brocklebank and Dickey (2003)). Based on the mean square error (MSE),

it can be seen that the seasonal decomposition model has the lowest forecast

error, provides the best result and is used in the comparison.

Table (3.4) shows the results of the the decomposition forecast results

and that of the decomposition model from SAS. As we can see, the er-

rors from both the forecast model are rather close although SAS provided
24
slightly better results, the reason seems to be the additional feature embed-

ded in the SAS system such as smoothing of the demand to derive the trend

etc.

Final Forecast of Demand from September 2007 to August 2008: Based

on the data from January, 2005 to June, 2007, we updated the forecast of the

demand from September 2007 to August 2008. Here, the values of intercept

and slope are calculated as 409.71 and 14.22 respectively. The trend, sea-

sonal factor, average factor and the forecast of the year 2007 and 2008 are

listed in Table (3.5).

Conclusion: Although more complicated forecast model can be developed

to discover the root of the fluctuation, there is no guarantee that they will

provide more accurate forecasts than simple models. Beside, market changes

could significantly change these forecast values – the key here is to develop

stochastic models to be used under various scenarios. The above model

provides insight in the business trend and is used together with company’s

forecast to form pessimistic, neutral and optimistic scenarios as inputs to the

stochastic production planning model to be discussed in the next chapter.

25
 Table 3.1: Demand Data & Decomposition Forecast Results for 2007
Year Consumption Trend Seasonal Factors Average Factors Forecast
Jan-05 405 431 0.94 0.84 362
Feb-05 352 445 0.79 0.79 353
Mar-05 474 458 1.03 1.08 494
Apr-05 445 472 0.94 0.98 461
May-05 465 485 0.96 1.06 513
Jun-05 727 499 1.46 1.33 665
Jul-05 519 513 1.01 1.03 530
Aug-05 607 526 1.15 1.18 620
Sept-05 517 540 0.96 0.97 525
Oct-05 515 553 0.93 1.00 553
Nov-05 603 567 1.06 0.99 564
Dec-05 418 581 0.72 0.74 428
Jan-06 439 594 0.74 0.84 499
Feb-06 485 608 0.80 0.79 483
Mar-06 697 621 1.12 1.08 670
Apr-06 641 635 1.01 0.98 620
May-06 749 648 1.16 1.06 685
Jun-06 799 662 1.21 1.33 882
Jul-06 712 676 1.05 1.03 698
Aug-06 830 689 1.20 1.18 813
Sept-06 693 703 0.99 0.97 683
Oct-06 766 716 1.07 1.00 716
Nov-06 675 730 0.92 0.99 726
Dec-06 562 743 0.76 0.74 549
Jan-07 757 0.84 635
Feb-07 771 0.79 613
Mar-07 784 1.08 846
Apr-07 798 0.98 779
May-07 811 1.06 857
Jun-07 825 1.33 1099
Jul-07 839 1.03 866
Aug-07 852 1.18 1005
Sept-07 866 0.97 841
Oct-07 879 1.00 879
Nov-07 893 0.99 888
Dec-07 906 0.74 669

26
 Table 3.2: Forecast Result Comparison with Company’s Judgmental Model
Year 2007 Consumption Our Forecast APE Company APE
(Forecast Model) Forecast (Company’s Model)
Jan-07 615 635 0.03 542 0.12
Feb-07 701 613 0.13 596 0.15
Mar-07 800 846 0.06 1026 0.28
Apr-07 944 779 0.17 1236 0.31
May-07 910 857 0.06 1253 0.38
Jun-07 1013 1099 0.08 1152 0.14
Jul-07 881 878 0.00 1099 0.25
Aug-07 981 994 0.01 1181 0.20
Sum (Error) 0.54 1.83
Average (Error) 0.07 0.23

Table 3.3: Forecast Errors of the Models

Forecast Model MSE
Linear Trend with Autoregressive Errors 77.58
Seasonal Decomposition 41.79
Seasonal exponential Smoothing 59.40
Winters Method-Additive 43.76
Winters Method-Multiplicative 61.83
ARIMA(0,1,1)(1,0,0)s NO INT 92.63
ARIMA(2,1,0)(0,1,1)s NO INT 85.74
Log Linear Trend with Autoregressive Errors 77.29
Log Linear Trend with Seasonal Terms 44.19
Log Seasonal Exponential Smoothing 64.51
Log Winters Method-Additive 46.27
Log Winters Method-Multiplicative 65.80
Log ARIMA(0,1,1)(1,0,0)s NO INT 92.86
Log ARIMA(2,1,0)(0,1,1)s NO INT 108.65

Table 3.4: Forecast Data Comparison with SAS Results

Year 2007 Consumption Decomposition APE SAS APE
Forecast (Forecast Model) Forecast (SAS’s Model)
Jan-07 615 635 0.03 663 0.07
Feb-07 701 613 0.13 689 0.02
Mar-07 800 846 0.06 833 0.04
Apr-07 944 779 0.17 853 0.11
May-07 910 857 0.06 884 0.03
Jun-07 1013 1099 0.08 1023 0.01
Jul-07 881 878 0.00 880 0.00
Aug-07 981 994 0.01 983 0.00
Sum (Error) 0.54 0.28
Average (Error) 0.07 0.03

27
Table 3.5: Forecast Data & Results of the Decomposition Model for September 2007 to August 2008
Year Consumption Trend Seasonal Factors Average Factors Forecast
Jan-05 405 424 0.96 0.83 353
Feb-05 352 438 0.80 0.83 365
Mar-05 474 452 1.05 1.06 479
Apr-05 445 467 0.95 1.04 487
May-05 465 481 0.97 1.07 517
Jun-05 727 495 1.47 1.29 640
Jul-05 519 509 1.02 1.03 526
Aug-05 607 523 1.16 1.18 616
Sep-05 517 538 0.96 0.97 522
Oct-05 515 552 0.93 1.00 550
Nov-05 603 566 1.07 0.99 561
Dec-05 418 580 0.72 0.73 426
Jan-06 439 595 0.74 0.88 520
Feb-06 485 609 0.80 0.88 533
Mar-06 697 623 1.12 1.06 658
Apr-06 641 637 1.01 1.03 659
May-06 749 651 1.15 1.04 677
Jun-06 799 666 1.20 1.36 903
Jul-06 712 680 1.05 1.03 702
Aug-06 830 694 1.20 1.15 798
Sept-06 693 708 0.98 0.94 664
Oct-06 766 722 1.06 0.93 669
Nov-06 675 737 0.92 0.92 676
Dec-06 562 751 0.75 0.69 516
Jan-07 615 765 0.80 0.88 670
Feb-07 701 779 0.90 0.88 682
Mar-07 800 794 1.01 1.06 838
Apr-07 944 808 1.17 1.03 836
May-07 910 822 1.11 1.04 854
Jun-07 1013 836 1.21 1.36 1134
Jul-07 850 1.03 878
Aug-07 865 1.15 994
Sept-07 879 0.94 823
Oct-07 893 0.93 827
Nov-07 907 0.92 833
Dec-07 921 0.69 633
Jan-08 936 0.88 819
Feb-08 950 0.88 831
Mar-08 964 1.06 1018
Apr-08 978 1.03 1013
May-08 993 1.04 1032
Jun-08 1007 1.36 1365
Jul-08 1021 1.03 1054
Aug-08 1035 1.15 1190

28
Chapter 4

Stochastic Production Planning Models

and Its Formulation

Product demand is one of the key inputs and a major source of uncertainty

to a production planning problem. As we have seen, forecast will never

be 100% accurate and market changes could significantly change the de-

mand in the future. As such, developing stochastic production planning

that accounts for demand uncertainity is necessary for effectively running a

production facility and to deal with any specific realization of the demand

uncertainty.

In this chapter, a two-stage stochastic production planning model is pre-

sented. The model is then extended to a robust production planning model.

While stochastic programming model is designed to minimize only the ex-

pected costs, the robust programming model is designed to provide solutions

that are consistent across all scenarios to reduce solution variability.

29
4.1 Problem Statement

The Definition of The First and Second Stage Decisions: In a two-stage

stochastic model, the definition of the first and the second stage variable

are problem and context dependent and accordingly a two-stage stochastic

models can be defined in several ways. For example, the first stage could

incorporate variables related to the location and capacity of the facility and

the second stage could represent how, given these decisions, the detailed

production plans should be carried out or inventory levels managed under

the realized demand. This results in a strategic capacity planing model as

seen in Eppen et al. (1989) and Karabuk and Wu (2008).

In an operation environment, when the capacity is fixed, purchasing and

production decisions also come in phases. For our client, a US subsidiary

of a global manufacture, most of the materials and sub-assemblies are ei-

ther purchased from its parent company in Europe or from east Asia due to

global outsourcing activities to utilize low labor cost. As a result, a long

lead time of more than 40∼45 working days is the norm, which means pur-

chasing decision and production decisions has to be made (3∼4 ) months

in advance before the demand is realized. The first stage decision in this

operational environment is therefore related to the purchasing decisions and

production decision for the next few (3∼4) months and the second stage de-

cision is to decide how, given these decisions, the detailed inventory, over-
30
time and production schedules under realized demand. This results in an

stochastic operational production planning model.

To be specific, it is assumed that the maximum production capacity,

defined as maximal number of units per time period, remain unchanged.

The first stage decision is how much material to purchase for the production

of the next 4 months, with an eye of the future seasonal demand. When

a possible realization of demand is revealed, the second stage decisions are

made to find, given these first stage purchasing and production decisions, the

detailed production schedules, overtime usage and inventory level managed

for these months.

In practice, of course, only the purchasing to support the next 4 months

are taken and the model needs to be solve every few months. The problem is

therefore a rolling plan capacity allocation model under uncertain demand.

The Generation of Pessimistic, Neutral and Optimistic Scenarios: The

decomposition forecast, though accurate, seems to gives the lowest values

and is used as the pessimistic forecast of future. The company’s judgmental

forecast, although always seems to be an overestimate, does represent an

optimistic view of the future and therefore as the optimistic forecast. A third

forecast was obtained by randomly choosing a value between pessimistic

and optimistic forecast and serves as a neutral forecast. An equal probability

is given to all these three forecasts or scenarios.

Other Operations: In developing the production planning model, the fol-

31
low options are used:

1. Overtime. Overtime can be used to create a temporary increase in ca-

pacity without the added expense of hiring additional workers. Due to

the labor union regulation, overtime is limited to 20% of the maximum

regular production capacity.

2. Inventory. Finished-goods inventory is used to fill demand during pe-

riods of high demand. A minimum safety stock of 200 units has been

decided to hedge against spikes in a period.

3. Hire/lay off. Laying off and rehiring are strategic decisions and is thus

not considered in this study. Workers cross trained and in a low demand

seasons, can be transferred in other production groups.

4. Back-Log. Back-log orders and subcontracting are not considered in

this study.

Notice again that the production plan for first four months are kept the

same, this represents the first stage decisions. When the purchasing decision

are known and demand are realized, the second-stage decisions are the de-

tailed overtime decisions and inventory level decisions that changes under

each scenario.
32
4.2 A Two-stage Stochastic Production Planning Model

4.2.1 Mathematical Model

The above production planning model belongs to the class of multi-scenario,

two-stage stochastic production planning model. The goal is to find a op-

timal production plan that minimizes the overall production, inventory and

overtime cost to satisfy the demand.

The following notation are defined in the development of the model.

Parameters

Indices:

s Index of scenario

t Index of time period

Sets:

S Set of scenario

T Set of time period (12 months)

33
Deterministic Parameters:

d(s,t) Demand forecast of thermal cooling unit product in period t

under scenario s

Cr Regular production cost of worker per year ($36000/year)

Co Overtime production cost of worker ($/man-hour)

Ch Annual holding cost per unit of product

N Average number of days in the period

I0 Initial inventory

τ Daily production of units per person

α Overtime ratio

β Average Inventory cost

δ Minimum Production quantity in time period t

w Number of workers

ρ Efficiency of workforce (ρ = 0.9)

Recourse Parameters:

p(s) Probability of occurrence of scenario

SS(s,t) Safety stock of products in period t under scenario s

34
Decision variables

P(s,t) Number of units produced in the regular hours in period t under

scenario s (Units/hours)

O(s,t) Number of units produced in the overtime hours in period t under

scenario s (Units/hours)

I(s,t) Inventory of product in end of period t under scenario s (Units)

y(s,t) 1 if overtime is used, otherwise 0 in period t under scenario s

A Two Stage Stochastic Production Planning Model

Minimize
" #
S T T
∑ p(s) Cr .w + ∑ O(s,t).Co + ∑ (I(s,t)/12).Ch (4.1)
s=1 t=1 t=1

Subject to

• Inventory balance constraints

I(s, 0) + P(s,t) + O(s,t) = d(s,t) + I(s,t) ∀s ∈ S,t = 1 (4.2)

I(s,t − 1) + P(s,t) + O(s,t) = d(s,t) + I(s,t) ∀s ∈ S,t = 2..T (4.3)

• Capacity constraints

P(s,t) ≤ N.τ.ρ.w ∀s ∈ S,t ∈ T (4.4)

35
 • Overtime and production occurrence constraints

P(s,t) ≥ N.τ.ρ.w.y(s,t) ∀s ∈ S,t ∈ T (4.5)

O(s,t) ≤ α.N.τ.ρ.w.y(s,t) ∀s ∈ S,t ∈ T (4.6)

• Minimum production constraints

P(s,t) ≥ δ.N.τ.ρ.w ∀s ∈ S,t ∈ T (4.7)

• Safety stock of finished product constraints

I(s,t) ≥ SS(s,t) ∀s ∈ S,t ∈ T (4.8)

• Same production constraints

P(s,t) = P(s + 1,t) ∀s = 1..S − 1,t = 1..4 (4.9)

• Non-negativity conditions

P(s,t), O(s,t), I(s,t) ≥ 0 ∀s ∈ S,t ∈ T (4.10)

y(s,t) is binary ∀s ∈ S,t ∈ T (4.11)

The objective (4.1) is to minimize the total cost, including the fixed an-

nual labor cost (term 1), overtime cost (term 2), and inventory cost (term 3).

Constraints (4.2) and (4.3) ensure that the beginning inventory plus produc-

tion and overtime during the current period equals the demand plus ending

inventory. Constraint (4.4) ensures that the total production by regular-time

workers during period t and scenarios s is limited by the production ca-

pacity. Constraints (4.5) and (4.6) ensure that the overtime occurs only
36
when the regular production is at its maximum level and overtime cannot

not exceed the maximum limit, which is set at α% of maximum regular

production. Constraint (4.7) ensures that the production under all scenar-

ios remains greater than the minimum production. Constraint (4.8) ensures

that the ending inventory at each period is not less than the minimum safety

stock. Constraint (4.9) ensures that the productions under all scenarios re-

mains the same for first four months–the first stage decisions. Constraint

(4.10) ensures that all decision variables are non-negative and constraint

(4.11) ensures that the decision variables y(s,t) is binary. This problem is

an integer linear programming (INLP) problem and is NP-hard (Garey and

Johnson (1979)).

4.3 Robust Production Planning Formulation

The above stochastic program aims to minimize the expected value, the first

moment, of inventory and production cost and does not include any high or-

der moments or risk attribute of the decision maker, or the distribution of the

objective values. In an effort to minimize the second or higher moments of

the inventory and production cost and to reduce the variability of solutions

under different scenarios, a robust model is used. The robust model aims to

obtain a solution that will not differ substantially among different scenarios

and to achieve this, minimizes fist order as well as higher order moments.

Two popular approaches for handling risk are: mean/variance mod-

37
els (Markowitz (1959)), and von Neumann-Morgenstern expected utility

models (Keeney and Raiffa (1976) and Bai et al. (1997)). The former

includes minimizing variance as part of the objective whereas the latter

presents a more general approach for handling risk aversion. In this study, a

mean/variance model by (Mulvey and Vladimirou (1991)) with a additional

variance constraint is used. This approach requires that the distribution of

the random variable be symmetric around its mean. Third and higher mo-

ments are ignored in the model.

To do so, a variance of cost is added to the model for robustness and the

following additional notations and variable are used in the development of

the model.

Parameters

λ Weighting scale

ε1 Lower bound

ε2 Upper bound

Variables

ξs Variable with probability ps under scenario s

Mathematical Model
38
 The robust production planning model can be formulated as follows.
!2
S S S
∑ ps ξs + λ ∑ ps ξs − ∑ psξs
′ ′
Min (4.12)
s=1 s=1 ′
s =1
K
W here ξs = cT x + ∑ pk (qTk yk ) (4.13)
k=1
′ ′
(∑s=1 ps ξs − ∑s′ =1 ps ξs )
S S
Sub ject to ε1 ≤ ≤ ε2 (4.14)
(∑Ss=1 ps ξs )
Equation(2.13) − (2.15) (4.15)

The first term of the objective function is the same as the stochastic model,

the expected total cost.

′ ′ 2
 
The second term in the objective function, λ ∑Ss=1 ps ξs − ∑s′ =1 ps ξs , is
S

used to penalize variations of production plan cost under various scenar-

ios. Constraint (4.14) is used to keep the difference between scenarios cost

within ε1 and ε2 . The objective function here is to minimize the deviation

in an effort to get solutions that are less sensitive to change in the demand

data under all scenarios.

Objective function

• Regular production cost

PC = Cr .w (4.16)

• Overtime production cost

T
OC(s) = ∑ O(s,t).Co (4.17)
t=1

39
 • Inventory cost
T
IC(s) = ∑ (I(s,t)/12).Ch.β (4.18)
t=1

Minimize
S
∑ p(s) [PC + OC(s) + IC(s)] (4.19)
s=1
" #2
S S
+λ ∑ p(s) (PC + OC(s) + IC(s)) − ∑
′
 ′ ′

p(s ) PC + OC(s ) + IC(s )
s=1 ′
s =1

Subject to

Equation (4.2)-(4.11)

• Scenario difference constraints

h ′ ′
i
(∑Ss=1 p(s) [PC + OC(s) + IC(s)] − ∑Ss′ =1 PC + OC(s ) + IC(s ) )
ε1 ≤ ≤ ε2
(∑Ss=1 p(s) [PC + OC(s) + IC(s)])
∀s ∈ S (4.20)

The first part of objective function in equation (4.19) is the total cost as-

sociated with the labor cost per year, overtime workers cost and the in-

ventory cost. The second part in objective function is the square of the

difference between first part of the cost and each individual scenario cost.

Constraint (4.20) ensures that the difference between each individual sce-

nario cost should remain between ratio ε1 and ε2 and average cost.

This problem is a integer non-linear integer programming (NILP) prob-

lem, where all unknown variables are all required to be integers and some

of the constraints or objective is non-linear.

40
Chapter 5

Computational Results

All the stochastic and robust production planning models are implemented

in Xpress and solved using its integer linear and nonlinear solvers on a Pen-

tium IV 1.8MHZ computer with 512 MB of RAM. The computation time

for most of the models are within 4 to 5 seconds and has not posed a great

challenge. The results from various models are presented derive managerial

insights.

5.1 Input Parameters

The Generation of Pessimistic, Neutral and Optimistic Scenarios:

As mentioned, the decomposition forecast, although accurate, seems to pro-

vide the smaller forecast values and is used as the pessimistic forecast (S(L))

of the future. The company’s judgmental forecast, although always seems

to be an overestimate, does represent an optimistic view of the future and

therefore as the optimistic forecast (S(H)). A third forecast was obtained by

randomly choosing a value between pessimistic and optimistic forecast and

serves as a neutral forecast(S(M)). An equal probability is given to each of

41
these three forecasts or scenarios. The product quantities required under

different scenarios in each month are shown in Table (5.1).

Table 5.1: Company’s Demand Data of the Scenarios

Scenario (s) 1 2 3 4 5 6 7 8 9 10 11 12
S1 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
S2 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
S3 869 885 824 633 836 950 1012 795 1036 1375 1385 1248

Initial Inventory and Inventory Cost: The initial inventory, Is0 , is the

finished goods inventory at the end of August 2007. This value is currently

set at 1500 units. Notice that this is a rather large value due to the lack of

proper forecast and stochastic production planning models in the past. The

annual inventory holding cost is provided by company and is set at 10% per

year and the average cost per unit is $1000. The minimum inventory is set

at 200 units as per company requirement.

Production Capacity: The number of workers for the production group

(w) is 18 and is fixed. The average daily production rate, τ, is 2.89 unit and

with 90% efficiency, it becomes 2.60 unit. The company operates 8 hours a

day, 5 days a week, thus the regular production capacity is calculated as 983

units per month with an average of 21 working days.

Production and Overtime Costs: The regular production cost (Cr ) and

overtime production cost (Co ) are constant throughout the planning horizon

and are estimated to be $18 per hour and $27 per hour respectively. The

overtime ratio is set at 20%.

42
5.2 Optimal Production Plan and Cost Analysis

Given these parameters, the optimal level of production, inventory and over-

time for different models are discussed presently.

5.2.1 Solution Under Perfect Information

Under perfect information, it is assumed that each of the scenario can hap-

pen with the given probability, but the manager knows beforehand exactly

what scenario will happen. If this is the case, the manager would then take

optimal solution for each scenario as given in Table (5.2). This would leave

him with a cost of $731,475 for the pessimistic scenario, $750,608 for the

optimistic scenario, and $745,267 for the neutral scenario. The average cost

in the long run would be the average of three costs, namely $742,450 per

year. This is the cost that incurs under perfect information when the future

demand is perfectly known.

Table 5.2: Deterministic Production Planning Model

Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Produce 650 650 650 650 657 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1327 1150 967 984 822 974 939 909 860 478 407 200
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Produce 650 650 891 983 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1168 924 994 1133 1279 1273 1254 1447 1300 874 467 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Produce 650 650 650 734 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1281 1046 872 973 1120 1153 1124 1312 1259 867 465 200

43
5.2.2 Two-Stage Stochastic Solution with the Same Four Month Production

Of course, to know future demand is impossible, in that case, the best a

manager can do is to follow the production plan as given by the stochastic

production planning model shown in Table (5.3). This gives an optimal cost

of $747,145 per year, which includes production cost, inventory cost and

overtime cost. From the Table (5.3), we can see that most of the products

are produced from regular-time. The production for first four month is the

same in all scenarios.

Table 5.3: Two-Stage Stochastic Model Production Plan (4 Month)

Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Produce 650 650 650 983 650 731 983 909 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1327 1150 967 1317 1148 1048 1013 909 860 478 407 200
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Produce 650 650 650 983 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 44 197
Inventory 1168 924 753 892 1038 1032 1013 1206 1059 633 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Produce 650 650 650 983 650 956 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 111
Inventory 1281 1046 872 1222 1036 1042 1013 1201 1148 756 354 200

The expected value of perfect information (EVIP) is defined the value

of knowing the future with certainty. The value can be calculated as the cost

of stochastic model ($747,145) minus the cost of solutions under perfect

information ($742,450), which equals $4,695. This is the cost we have to

pay for not knowing the future or the maximum money we would like to
44
pay to know the future.

5.2.3 Two-Stage Stochastic Model with the Same 12 Month Production

A two stage stochastic model can also be derive to find solutions where all

the 12 month productions remain the same. Such a model would be useful to

provide purchase forecast for supplier or for supply contract negotiation. In

such a model, we again set the first stage decision as the regular production,

but set all the production to be the same across the planning horizon. When

demand are realized, overtime and inventory can be adjusted to meet the

requirements. The cost of such a two stage stochastic model, as shown

in Table (5.4), is $754,169. The value is higher than two-stage stochastic

model with same 4 moth production as it has more constraints. The expected

value of perfect information and value of stochastic solution for the two two-

stage models are given in Table (5.5)

Table 5.4: Two-Stage Stochastic Model Production Plan with Same 12 Month Production
Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 650 650 650 650 966 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1327 1150 967 984 1131 1283 1248 1218 1169 787 716 509
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 650 650 650 650 966 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 197 197 197
Inventory 1168 924 753 559 688 682 663 856 709 480 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 650 650 650 650 966 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 101
Inventory 1281 1046 872 889 1019 1052 1023 1211 1158 766 364 200

45
 Table 5.5: EVIP and VSS Value for the Two Two-Stage Stochastic Models
Model Total Cost EVIP VSS VSS
(Expected Value) (Expected Demand)
Two-Stage (Same 12 Month) 754169 11719 -3305 -1730
Two-stage (Same 4 Month) 747145 4695 -10329 -8754

5.2.4 Expected Value Solution

Yet another approach to derive production plans is to choose the production

plan as the average of the production plan in all three scenarios, when they

are solved individually. This results in what is called the expected value

solution. This approach is common in optimization field but can have un-

favorable results. Table (5.6) shows the inventory level and overtime usage

for such a solution under the three scenarios, $763,033 in the first scenario,

$760,598 in the second scenario and $748,792 in the third scenario. The

average total cost in long run is $757,474 per year.

The value of the stochastic solution(VSS) is the possible gain from solv-

ing the stochastic model and is calculated as the expected value solution cost

($757,474) minus stochastic model cost ($747,145) and equals $10,329.

This is the benefit of solving the stochastic model over the average model,

in this case, the expected value solution.

5.2.5 Solution Under Expected Demand

Lastly, the most common approach, the deterministic model, is to solve

the problem with the average of the demand of the three scenarios. This

approach is also called the solution under expected demand. Table (5.7)
46
 Table 5.6: Inventory and Overtime Based on Expected Value
Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 650 650 730 789 874 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1327 1150 1047 1203 1258 1410 1375 1345 1296 914 843 636
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 650 650 730 789 874 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 70 197 197
Inventory 1168 924 833 778 815 809 790 983 836 480 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 650 650 730 789 874 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1281 1046 952 1108 1146 1179 1150 1338 1285 893 491 226

shows the production plan under such a model. The cost of this expected

demand solution for the three scenarios are shown in Table (5.8).

The value of the stochastic solution (VSS) is calculated as the expected

demand solution cost ($755,899) minus stochastic model cost ($747,145)

and equals $8,754. This is the benefit of solving the stochastic model over

the average model, in this case, the expected demand solution.

It has to mention that, the expected value solution was to combine a

solution from the each individual solution, where as the solution under ex-

pected demand is to solve the problem with the expected demand.

Table 5.7: Production Plan Based on Expected Demand

Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Demand 891 869 826 703 831 923 1011 866 1066 1383 1276 1229
Production 650 650 650 760 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1259 1040 864 921 1073 1133 1105 1222 1139 739 446 200

47
 Table 5.8: Inventory and Overtime for All Scenarios Based on Expected Demand Production
Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 650 650 650 760 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1327 1150 967 1094 1258 1410 1375 1345 1296 914 843 636
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 650 650 650 760 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 70 197 197
Inventory 1168 924 753 669 815 809 790 983 836 480 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 650 650 650 760 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1281 1046 872 999 1146 1179 1150 1338 1285 893 491 226

As a conclusion, Table (5.9) shows the inventory cost, overtime cost,

production cost, total cost and expected cost for different models discussed

in this section.

5.3 Sensitivity Analysis

5.3.1 Change of Probability Associated with the Scenarios

In this section, we considered the changes in the total costs by changing the

probability associated with the scenarios. Three cases have been considered

in this study. In case I, each scenario has equal probability. In case II, the

probability distribution is skewed towards the pessimistic forecast (S(L)). In

case III, the probability distribution is skewed towards the optimistic fore-

cast (S(H)). Table (5.10) shows the probability of each scenarios and Table

(5.11) shows the optimal costs in these cases.

As can be seen, the total cost is reduced in Case II since a large prob-
48
 Table 5.9: Cost Comparison of Various Models
Model Scenario Production Overtime Inventory Total Expected
Cost Cost Holding Cost cost cost
Perfect Information Solution
S(L) 648000 0 83475 731475
S(H) 648000 0 102608 750608 742450
S(M) 648000 0 97267 745267
Two-Stage Stochastic Model
(4 Month)
S(L) 648000 0 90200 738200
S(H) 648000 20022 84900 752922 747145
S(M) 648000 9221 93092 750313
Two-Stage Stochastic Model
(12 Month)
S(L) 648000 0 104075 752075
S(H) 648000 49098 66267 763365 754169
S(M) 648000 8391 90675 747066
Expected Value Solution
S(L) 648000 0 115033 763033
S(H) 648000 38548 74050 760598 757474
S(M) 648000 0 100792 748792
Expected Demand Solution
S(L) 648000 0 113458 761458
S(H) 648000 38548 72475 759023 755899
S(M) 648000 0 99217 747217

Table 5.10: Probability Associated with Each Scenario

Case Scenario S(L) Scenario S(H) Scenario S(M)
Case I 0.33 0.33 0.33
Case II 0.5 0.2 0.3
Case III 0.2 0.5 0.3

ability of a low demand is present. In contrast, the total cost is increased

in case III since a large probability of high demand is present. Not surpris-

ingly, changes of scenario probability changes the mean of future demand,

high (low) demand will result in high (low) costs because more units are

produced, more (less) overtime used, and more (fewer) inventory required

to avoid shortages. The detailed production plans associated with case II

and case III are shown in Table (5.12) and (5.13) respectively.
49
 Table 5.11: Sensitivity Analysis- Change of Probability Distribution
Case Production Cost Overtime Cost Inventory Holding Cost Total cost
Case I 648000 9748 89397 747145
Case II 648000 10293 85936 744229
Case III 648000 9853 91241 749094

Table 5.12: Inventory, Overtime and Production Under Case II

Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 650 650 650 882 650 758 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1327 1150 967 1216 1047 974 939 909 860 478 407 200
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 650 650 650 882 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 27 0 0 0 118 197
Inventory 1168 924 753 791 937 931 939 1132 985 559 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 650 650 650 882 650 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 185
Inventory 1281 1046 872 1121 935 968 939 1127 1074 682 280 200

5.3.2 Change from Scenario Generation Method

In all previous experiments, scenarios S(L) and S(H) represents lower and

upper bound and scenario S(M) has random values between S(L) and S(H).

In this experiment, rather than choose lower and higher demand S(L) and

S(H) as the scenarios, three random scenarios are generated with demand

uniformly distributed between S(L) and S(H). The detailed production plans

are shown in Table (5.14). The cost associated with the three scenarios are

$688,883, $689,542, $688,292 and the average is $688,906.

It is interesting to notice that a significant lower cost of $688,906 is ob-

tained. The reason we believe is as follows. Supper a low demand of 0 and

50
 Table 5.13: Inventory, Overtime and Production Under Case III
Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 650 650 694 983 650 731 983 865 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1327 1150 1011 1361 1192 1092 1057 909 860 478 407 200
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 650 650 694 983 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 197
Inventory 1168 924 797 936 1082 1076 1057 1250 1103 677 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 650 650 694 983 650 956 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 67
Inventory 1281 1046 916 1266 1080 1086 1057 1245 1192 800 398 200

a high demand of 1, if the three scenarios (S(L),S(H),S(M)) are used, it rep-

resents the lower, upper and medium scenario, shown in Figure (5.1) with

a discrete distribution. The mean is of the demand 0.5 and the variance is

0.408. The random sample of scenarios from low and high, however, repre-

sents the medium demand, (shown in Figure (5.1) by uniform distribution).

The mean of the demand is 0.5 and variance is 0.288. Though the mean

demand of the two experiments are the same, the latter has a much smaller

variance, thus a much smaller cost.

As we can see, the magnitude of variance in the uncertainty have a dra-

matic impact on the solution of a stochastic program. While in practice, it is

common to specify optimistic, neutral, and pessimistic scenarios, whether

to use optimistic and pessimistic or to choose random samples between the

optimistic and pessimistic requires detailed investigation.

51
 Table 5.14: Inventory, Overtime and Production Under 3 Random Scenarios
Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (M1)
Demand 970 854 892 968 905 916 866 878 943 969 825 915
Production 650 650 650 650 689 916 866 878 943 969 825 915
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1180 976 734 416 200 200 200 200 200 200 200 200
Scenario (M2)
Demand 940 855 972 846 965 974 937 899 942 971 853 941
Production 650 650 650 650 678 974 937 899 942 971 853 941
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1210 1005 683 487 200 200 200 200 200 200 200 200
Scenario (M3)
Demand 904 978 929 857 947 825 923 901 941 872 850 894
Production 650 650 650 650 715 825 923 901 941 872 850 894
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1246 918 639 432 200 200 200 200 200 200 200 200

5.3.3 Change of Overtime Ratio

In this experiment we study the change of costs with respect to different

overtime ratios. These overtime ratios are set 0%, 5%, 10%, 15%, 20%,

25% and 30% respectively. Table (5.15) and Figure (5.2) shows the changes

the costs associated with overtime ratio.

Table 5.15: Sensitivity Analysis-Overtime Ratio Varies

Case Production Cost Overtime Cost Inventory Holding Cost Total cost
0% 648000 0 103228 751228
5% 648000 7062 93161 748223
10% 648000 9748 89833 747581
15% 648000 9748 89536 747284
20% 648000 9748 89397 747145
25% 648000 9748 89397 747145
30% 648000 9748 89397 747145

As we can see, an overtime ration of 50%, representing no overtime

allowed, gives the highest cost of $751,228. When overtime ratio increases,

the total cost drops from $751,228 to $ 747,145. However, when overtime
52
 Figure 5.1: Mean and Variance of two random samples

ratio exceeds 20%, the total cost remains the same.

This experiment shows that certain use of overtime is beneficial. How-

ever, as excessive use of overtime may not be optimal – as overtime ratio

raise up to 20%, an 100% preminum overtime pay is incurred, raising the

overtime cost to $36 per hour. These overtime is costly and will not appear

in the optimal solution.

5.3.4 Change of Inventory Holding Cost

In this experiment, we studied the changes of total cost with respect to dif-

ferent inventory holding costs. The inventory holding cost are set at 0%,

5%, 10%, 15%, 20%, 25%, and 30% respectively. Table (5.16) shows the

total costs associated with the inventory holding values and the detailed pro-

duction plans for inventory holding cost of 20% and 30% are shown in Table
53
 5
x 10
7.515

7.51

7.505

Total Cost 7.5

7.495

7.49

7.485

7.48

7.475

7.47
0 5 10 15 20 25 30
Overtime Ratio

Figure 5.2: Sensitivity Analysis-Overtime Ratio Varies

(5.17) and Table (5.18) respectively.

Table 5.16: Sensitivity Analysis-Inventory Holding Cost Varies

Case Production Cost Overtime Cost (Units) Inventory Holding Cost (Units) Total cost
5% 648000 ( 10575 ) 0(0) 51614 ( 13887 ) 699614
10% 648000 ( 10457 ) 9748 ( 117 ) 89397 ( 12228 ) 747145
15% 648000 ( 10192 ) 31957 ( 383 ) 108833 ( 10222 ) 788790
20% 648000 ( 10149 ) 35363 ( 426 ) 140883 ( 9955 ) 824246
25% 648000 ( 10016 ) 46578 ( 559 ) 164174 ( 9392 ) 858752
30% 648000 ( 9939 ) 53141 ( 636 ) 191292 ( 9131 ) 892433

As we can see, when inventory holding cost increases, the total cost

increases, more units are produced in overtime and the average units stored

for future demand is decreased. It is interesting to notice that as inventory

holding cost increases, even regular production for first few months is kept

at the minimum and overtime is used to fulfill the demand. This seems

to correspond to a lean manufacturing notion that ”extra production, if not

needed, is considered as a waste”.

54
 Table 5.17: Inventory, Overtime and Production with Inventory Holding Cost 20%
Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 650 650 650 650 650 731 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 62 197
Inventory 1327 1150 967 984 815 715 680 650 601 219 210 200
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 650 650 650 650 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 180 197 197
Inventory 1168 924 753 559 705 699 680 873 726 480 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 650 650 650 650 650 956 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 50 197 197
Inventory 1281 1046 872 889 703 709 680 868 815 473 268 200

5.4 Robust model

In the computation of the robust model, parameters λ is set at 0.01, ε1 and ε2

at 0.05. The optimal level of production, inventory and overtime is shown

in Table (5.19) and the optimal total cost is $750,714.

Though the robust solution has a slightly high total cost, it however,

has dramatically less changes in the cost for each scenario. The cost for

S(L) is $750,674, for S(M), $750,667, for S(H), $750,801. The small cost

difference among each scenarios may be preferable if a manager does not

want dramatic changes in its cost over years. The cost difference for all

models are listed in Table (5.20)

55
 Table 5.18: Inventory, Overtime and Production with Inventory Holding Cost 30%
Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 650 650 650 650 650 650 850 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 195 81 197
Inventory 1327 1150 967 984 815 634 466 436 387 200 210 200
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 650 650 650 650 769 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 197 197 197 197
Inventory 1168 924 753 559 491 485 466 659 709 480 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 650 650 650 650 650 742 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 67 197 197 197
Inventory 1281 1046 872 889 703 495 466 654 668 473 268 200

Table 5.19: Robust Model Production Plan

Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 650 650 791 983 707 753 904 983 934 983 828 964
Overtime 0 0 0 0 0 0 0 4 0 4 0 0
Inventory 1327 1150 1108 1458 1346 1268 1154 1128 1030 652 426 200
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 650 650 791 983 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 100
Inventory 1168 924 894 1033 1179 1173 1154 1347 1200 774 367 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 650 650 791 983 650 956 983 953 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 1281 1046 1013 1363 1177 1183 1154 1312 1259 867 465 200

Table 5.20: Cost Difference of Various Models

Model Total Cost cost under S(L) cost under S(H) cost under S(M)
Perfect Information Solution 742450 731475 750608 745267
Two-Stage Stochastic Model (4 Month) 747145 738200 752922 750313
Two-Stage Stochastic Model (12 Month) 754169 752075 763365 747066
Robust Stochastic Model 750674 750801 750667 750714
Expected Value Solution 757474 763033 760598 748792
Expected Demand Solution 755899 761458 759023 747217

56
Chapter 6

Conclusions and Implementation

6.1 Conclusions

Production planning plays a vital role in manufacturing companies, but un-

certainty and randomness in demand pattern makes it difficult for the man-

agers to take full advantage of production planning model. In this study,

a two-stage stochastic production planning model is proposed to deal with

uncertain demand.

The model is solved with real data from a local company and compared

with deterministic production models on the basis of robustness and effec-

tiveness. A parametric analysis is used to derive managerial insights related

to issues such as overtime usage, inventory holding cost and the proper se-

lection of scenarios under pessimistic, neutral and optimistic forecasts. An

extension of the stochastic model i.e., a robust model is also formulated

and solved in an effort to minimize changes in the solutions under various

scenarios. The stochastic model provide the best overall results and help

making sound decision under uncertainty.

57
6.2 Implementation

The initial inventory of thermal cooling unit , as collected from real data of

work-in-progress or finished good inventories at the beginning of the plan-

ning horizon (September 2007), is 1500 units. This makes the average in-

ventory from September 2007 to August 2008 to 894 units.

Table 6.1: Production Plan From September 2008 to August 2009

Time 1 2 3 4 5 6 7 8 9 10 11 12
Period
Scenario (L)
Demand 823 827 833 633 819 831 1018 1013 1032 1365 1054 1190
Production 982 894 983 983 715 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 0
Inventory 359 426 576 926 822 974 939 909 860 478 407 200
Scenario (H)
Demand 982 894 821 844 837 989 1002 790 1130 1409 1390 1250
Production 982 894 983 983 983 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 120 197 0 0 0 118 197
Inventory 200 200 362 501 647 761 939 1132 985 559 270 200
Scenario (M)
Demand 869 885 824 633 836 950 1012 795 1036 1375 1385 1248
Production 982 894 983 983 940 983 983 983 983 983 983 983
Overtime 0 0 0 0 0 0 0 0 0 0 0 185
Inventory 313 322 481 831 935 968 939 1127 1074 682 280 200

To find the optimal level of initial inventory, we resolve the model and

leave I0 as a variable and kept it equal to final inventory. The optimal level of

initial inventory is 200 units. As such, the average inventory from Septem-

ber 2008 to August 2009, as shown in Table (6.1) will be 551 units. This

translates into a reduction of inventory by 38% and more than $350,000 of

inventory cost savings for the company. This is achieved through the use

of better forecast and stochastic production planning model with uncertain

demand. In addition, due to the purchasing of materials at the right time,

58
air frieght expition cost has been reduced and service level has slightly in-

creased.

59
Bibliography

Alonso-Ayuso, A., L. Escudero, A. Garin, M. Ortuno, and G. Perez, 2003.

An approach for strategic supply chain planning under uncertainty based

on stochastic 0-1 programming. Journal of Global Optimization, 26:97–

124.

Bai, D., T. Carpenter, and J. Mulvey, 1997. Making a case for robust opti-

mization models. Management Science, 43:895–907.

Beale, E., 1955. On minimizing a convex function subject to linear inequal-

ities. Journal of Royal Statistics Society, 17:173–184.

Birge, J. R., 1959. Chance constrained programming. Management Science,

5:73–79.

Birge, J. R., 1982. The value of the stochastic solution in stochastic linear

programs with fixed recourse. Mathematical Programming, 24:314–325.

Box, G., G. M. Jenkins, and G. Reinsel, 1976. Time Series Analysis: Fore-

casting and Control (3rd ed.). Holden-Day, San Francisco.

60
Brocklebank, J. C. and D. A. Dickey, 2003. SAS for Forecasting Time Series

(2nd ed.). SAS Publishing.

Carino, D. R., T. Kent, D. H. Meyers, C. Stacy, M. Sylvanus, A. L. Turner,

K. Watanabe, and W. T. Ziemba, 1994. The russel-yasuda kasai model: An

asset liability model for a japanese insurance company using multistage

stochastic programming. Interfaces, 24:29–49.

Carino, D. R., D. H. Meyers, and W. T. Ziemba, 1998. Concepts, techni-

cal issues, and uses of the russell- yasuda kasai financial planning model.

Operations Research, 46:450–462.

Carino, D. R. and W. T. Ziemba, 1998. Formulation of the russell-yasuda

kasai financial planning model. Operations Research, 46:433–449.

Chen, Z. L., L. Shanling, and D. Tirupati, 2002. A scenario based stochastic

programming approach for technology and capacity planning. Computers

and Operations Research, 29:781–806.

Clay, R. L. and I. E. Grossmann, 1997. A disaggregation algorithm for the

optimization of stochastic planning models. Computer Chemical Engi-

neering, 21:751–774.

Consigli, G. and M. A. H. Dempster, 1998. Dynamic stochastic program-

ming for asset-liability management. Annals of Operations Research,

81:131–161.
61
Dantzig, G. B., 1955. Linear programming under uncertainty. Management

Science, 1:197–206.

Dempster, M. A. H., N. H. Pedron, E. A. Medova, J. E. Scott, and A. Sem-

bos, 2000. Planning logistics operations in the oil industry. Journal of the

Operational Research Society, 51(11):1271–1288.

Eppen, G. D., R. K. Martin, and L. Schrage, 1989. A scenario approach to

capacity planning. Operations Research, 37:517–527.

Escudero, L. and P. Kamesam, 1955. On solving stochastic production plan-

ning problems via scenario modelling. An Official Journal of the Spanish

Society of Statistics and Operations Research, 3:69–95.

Escuderoa, L. F., E. Galindoa, G. Garcaa, E. Gmeza, and V. Sabau, 1999.

Schumann, a modeling framework for supply chain management under

uncertainty. European Journal of Operational Research, 119:14–34.

Fair, R. C., 1989. The production-smoothing model is alive and well. Jour-

nal of Monetary Economics, 24:353–370.

Feiring, B. R. and T. Sastri, 1990. Improving production planning by utiliz-

ing stochastic programming. Computers & Industrial Engineering, 19:53–

56.

Fleten, S.-E. and T. K. Kristoffersen, 2008. Short-term hydropower pro-

62
 duction planning by stochastic programming. Computer and Operation

Research, 35:2656–2671.

Fogarty, D. W., J. H. Blackstone, and T. R. Hoffmann, 1984. Production

and Inventory Management. Prentice-Hall, New Jersey.

Garey, M. R. and D. S. Johnson, 1979. Computers and Intractability: A

Guide to the Theory of NP-completeness. W. H. Freeman & Company,

San Francisco.

Hill, A. and W. Sawaya, 2004. Production planning for medical devices with

an uncertain regulatory approval date. IIE Transactions, 36:307–317.

Homburg, L., 1996. Production planning with multiple objectives in decen-

tralized organizations. International Journal of Production Economics,

56-57:243–52.

Huang, H.-J. and G. Xu, 1998. Aggregate scheduling and network solving

of multistage and multi-item manufacturing systems. International Jour-

nal of Production Economics, 105:52–65.

Kanyalkar, A. P. and G. K. Adil, 2005. An integrated aggregate and de-

tailed planning in a multi-site production environment using linear pro-

gramming. International Journal of Production Research, 43:4431–4454.

Karabuk, S. and S. D. Wu, 2003. Coordinating strategic capacity planning

in the semiconductor industry. Operations Research, 51:839–849.

63
Karabuk, S. and S. D. Wu, 2008. Production planning under uncertainty

in textile manufacturing. Journal of the Operational Research Society,

59:510–520.

Keeney, R. and H. Raiffa, 1976. Decisions with multiple objectives-

preferences and value tradeoffs. John Wiley and Sons, New York.

Kira, D., M. Kusy, and I. Rakita, 1997. A stochastic linear programming

approach to hierarchical production planning. Journal of the Operational

Research Society, 48:207–211.

Krane, S. and S. Braun, 1991. Production smoothing evidence from

physical-product data. Journal of Political Economy, 99:558–581.

Lee, L. H., E. P. Chew, and T. S. N, 2005. Production planning with ap-

proved vendor matrices for a hard-disk drive manufacturer. European

Journal of Operational Research, 162:310–324.

Leung, S., Y. Wu, and K. Lai, 2006. A stochastic programming approach

for multi-site aggregate production planning. Journal of the Operational

Research Society, 57:123–132.

Liu, X. and Y. Tu, 2008. Production planning with limited inventory capac-

ity and allowed stockout. International Journal of Production Economics,

111:180–191.

Maatman, A., C. Schweigman, A. Ruijs, and van M. H. Der Vlerk, 2002.

64
 Modeling farmers’ response to uncertain rainfall in burkina faso: A

stochastic programming approach. Operations Research, 50:399–414.

Markowitz, H., 1959. Efficient Diversification of Investments. John Wiley

and Sons, New York.

Mula, J., R. Poler, J. Garca-Sabater, and F. Lario, 2006. Models for pro-

duction planning under uncertainty: A review. International Journal of

Production Economics, 103:271–285.

Mulvey, J. M. and H. Vladimirou, 1991. Applying the progressive hedg-

ing algorithm to stochastic generalized networks. Annals of Operations

Research, 31:399–424.

Nahmias, S., 1993. Production and Operations Analysis (2nd ed.). Irwin,

New York.

Qiu, M. M. and E. E. Burch, 1997. Hierarchical production planning and

scheduling in a multi-product, multi-machine environment. International

Journal of Production Research, 35:3023–42.

Reaychen, W. and H.-H. Fang, 2001. Aggregate production planning with

multiple objectives in a fuzzy environment. European Journal of Opera-

tional Research, 133:521–536.

Rockafellar, R. T. and R. J. B. Wets, 1991. Scenarios and policy aggregation

65
 in optimization under uncertainty. Mathematics of Operations Research,

16:119–147.

Schaefer, A. J., E. L. Johnson, A. J. Kleywegt, and G. L. Nemhause,

2005. Airline crew scheduling under uncertainty. Transportation Science,

39:340–348.

Seider, W. D., D. D. Brengel, and S. Widagdo, 1991. Nonlinear analysis in

process design: A review. AIChE Journal, 37:1–38.

Swoveland, C., 1975. A deterministic multi-period production planning

model with piecewise concave production and holding backorder costs.

Management Science, 21:1007–1013.

Tabucanon, M. T., 1988. Multiple criteria decision making in industry. Jour-

nal of the Operational Research Society, 41:883–883.

Yaffee, R. and M. McGee, 2000. Introduction to Time Series Analysis and

Forecasting with Applications to SAS and SPSS. Academic Press, Inc.,

San Diego.

Yen, J. and J. Birge, 2001. A stochastic programming approach to the airline

crew. scheduling problem. Transportation Science, 40:3–14.

66