Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Procedia
Engineering
Procedia Engineering
Procedia 00 (2011)
Engineering 000000
15 (2011) 1154 1159
www.elsevier.com/locate/procedia

Advanced in Control Engineeringand Information Science

Model Predictive Control for inventory Management in

Supply Chain Planning
DONG Haia,b, ZHENG Haob, LI Yan Pingb
a
School of Mechanical Engineering, Shenyang University, Shenyang, P. R. China
b
Equipment manufacturing integrated automation key lab, Shenyang University, Shenyang, P. R. China

Abstract

In supply chain planning, it is very important to forecast the changes in the market in order to maintain an inventory
level that is just enough to satisfy customer demand. According to the demand uncertainty in supply chain network,
model predictive control was applied to solve dynamic optimization of the inventory. Firstly, the mechanism of
model predictive control and adaptability of dynamic control in supply chain were presented, and the problem of
difficult determining of the cost of change made in the control in supply chain related systems was studied. Secondly,
a comprehensive model of the cost objective optimization was established, and a general optimal control law and
control algorithm were proposed, and a new supply chain simulator interface was also presented and used in the
simulations of this study. Finally, simulation results show that this approach can track customer demand and maintain
a proper inventory level without causing a bullwhip effect, and effectively control dynamic and uncertain supply
chain system.
2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
Selection and/or peer-review under responsibility of [CEIS 2011]
Keywords: Supply chain planning; model predictive control; inventory management; simulation

1. Introduction

In recent years many different fields of science have been used to manage the problematic dynamics of
supply chains(SC). Control and systems theory have been among the most popular tools to handle the
problem consisting of uncertainties, delays and lack of information, and have shown to be very
applicable.
_________

* Corresponding author. Tel.: 13898802977; fax: +86-024-62268611.

1877-7058 2011 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
doi:10.1016/j.proeng.2011.08.213
2 DONG HaiHai,et
DONG et al. /al/
Procedia Engineering
Procedia 1500
Engineering (2011) 1154
(2011) 1159
000000 1155

E-mail address: sydxdh@163.com

Recent work using model predictive control (MPC) has shown it as an attractive method for SC
management and inventory control. Such as Braun M W studied application of MPC to robust
management of multiechelon demand networks in semiconductor manufacturing supply chains[1]. Alberto
B. analyzed MPC of Hybrid Systems with applications to supply chain management[2]. William B. studied
distributed MPC for dynamic supply chain management [3].The purpose of this paper is to demonstrate the
application of MPC to the problem of dynamic management of supply chain planning. Method of
automation, especially model predictive control are being studied,and the suitability to supply chain and
inventory management examined.

2. Problem Description

Supply chain management(SCM) is a process, which can quick response to demand of market, and
plan, coordinate, control and optimaize whole the network system of supply chain, and optimization
control of dynamic inventory is a difficulty in SC control[4]. In real world demand variance magnifying
and the uncertainty of demand forecast in SC result in difficulty increase to enterprise control inventory.
Basically there are two problems, delays and uncertainties, which together make supply chains very
problematic to control. All decisions are made on the basis of forecasts and therefore uncertainties are
constantly present. Demand forecasting is not the only phase of the supply chain where uncertainty can be
forund. Altogether uncertainty is involved with demand forcasting, stockholding, production.

3. Control Strategy

3.1. Mathmatical model

The key issue is the foundation of reasonable model in MPC control dynamic inventory of SC. MPC
use optimization to design an optimal control law by minimizing an objective function. The general
formation of the objective function can be written as

N2 Nu

J (N1, N 2 , Nu ) = ( j )[ y (k + j k ) w ( k + j )] +
2
( i )[ u ( k + i 1)] 2
(1)
j = N1 i =1

where N1 and N2 are the minimum and maximum cost horizons and Nu is the control

horizon. ( j ) and (i ) can be seen as the unit costs of the control. ( k ) , y ( k ) and u(k) are the reference
trajectory, the predicted outputs and the change between current predicted control signal and previous
predicted control signal, respectively.
As can be seen from the name of the cost function, quadratic cost functions use quadratic costs. The
most basic form of an inventory control based quadratic cost function can be written as

N2 Nu

m in J ( N 1 , N 2 , N u ) = ( i )( I d
( i ) I ( i )) + ( i ) ( O ( i ))
2 2
(2)
i= N1 i =1

where N1, N2 and Nu are the prediction and control horizon limitations similar to the ones in Equation

(1), I (i ) is the target inventory level at instant i, I (i ) is the predicted inventory level, dependent on the
d
1156 DONGname
Author Hai et/ Procedia
al. / Procedia Engineering
Engineering 15 (2011)
00 (2011) 1154 1159
000000 3

order level, at instant i, (i ) and (i ) are weight parameters at instant i and O (i ) is the change made in
ordering level between instants i-1 and i.
Linear cost functions operate quite different from quadratic cost functions as all costs increase linearly.
Therefore, for example, inventory unit costs can be used as such to penalize held inventory.In this case
the trade-off cost would be penalizing stock-out[5]. Based on this, one can write the most basic form of an
inventory control based linear cost function as

Nu

min J ( N u ) = ( (i ) I (i ) + (i ) S (i )) (3)
i =1

Where Nu is the control horizon limitation, I (i ) is predicted inventory level at instant i, S ( i ) is

predicted stock-out level at instant i, and (i ) and (i ) are weight parameters or actual unit costs,

penalizing held inventory and stock-out situations,respectively, at instant i. Both I (i ) and S ( i ) are
dependent on order levels, and need to be positive.

3.2. The desired inventory level

The paper set the main goal of SCM is to satisfy certain customer demand with a level of certainty, this
make the meaning of the cost function changes dramatically. So it is no longer necessary to implement
every single unit cost into the control strategy, but one can concentrate more on the issues affecting
customer satisfactory levels and robustness of the control. In both of these cases inventories have a
significant role, it can be regarded as a buffer to guarantee a certain customer demand satisfactory (CDS)
level. The CDS level is often expressed with a certain percent which denotes the percentage which is
satisfied of the total customer demand. This can be turned around and calculated, how much safety stock
is needed to meet this percentage of customer demand, its function form can be written as

CDS CDS (k )
(4)
100%
=
p ( D ( k )) d D ( k )

Where CDS is the desired customer demand satisfactory level percentage, CDS (k ) is the upper
confidence limit,set by the CDS level,at instant k and p(D(k)) is the density function of demand D(k) at
instant k. After solving the upper confidence limit, the desired inventory level Id (k) can be calculated with

I d ( k ) = CDS ( k ) E ( D ( k )) (5)

Where CDS (k ) is the upper limit of the confidence interval set by the desired CDS level and E(D(k))
is the expectation value of forecasted demand D(k).

3.3. Control algorithm

The cost functions presented in Equations (2) are completed with the desired inventory level
calculation presented in Equation (5), the actual algorithm implemented can be written. In a very general
form, the algorithm can be written, step by step, as follows:
Step 1
4 DONG HaiHai,et
DONG et al. /al/
Procedia Engineering
Procedia 1500
Engineering (2011) 1154
(2011) 1159
000000 1157

Generate the demand forecast D so that the density function p(D) of related uncertainty is obtained.
Determine the required CDS level the company needs to satisfy.
Step 2
Calculate the desired inventory levels Id for each product and for the whole prediction horizon with
Equations (4) and (5) on the basis of the CDS level and uncertainty density function p(D) determined in
Determine the cost function J to be optimized, with related parameters, such as, for example, unit
costs and and tuning parameters p1 and p2.
Determine the constraints related to the optimization.
Step 3
Calculate the optimal order levels for the control horizon by minimizing the cost function J
determined in Step 2 while taking constraints into consideration.
For each product, use only the first element of the optimal order level as the actual order to the
subcontractor.
In this thesis, the algorithm has been used in all simulations. For the demand forecaster, the simple
method of multiple regression was used.The model of the multiple regression forecaster can be written as

Yi = 1 + 2 X 2 i + 3 X 3 i + + k X ki + i (6)

where Yi is the dependent variable, which in this case is the demand forecast, X s are the independent
variables and is the error term. Each Xji represents the ith observation on explanatory variable Xj ,1 is
the constant term of the equation. The model is used to forecast the demand for the next instant.

4. Simulations

In this paper,the actual simulations are implemented based on a semiconductor enterprise which
mainly produces C3,and C3 consists of three core element including M10,M20 and M40.The goal in the
simulations demonstrate that MPC can handle dynamic inventory of SC under uncertain conditions.The
simulations in this paper were made using MATLAB and Simulink.

4.1. Simulators and date used

The supply chain simulations are made with more realistic data with three different products, distortion
and seasonal behavior. The distortions are normally distributed in order to keep the demand forecasting as
simple as possible since the main focus is not on the forecaster. The demand pattern used in the
simulations can be seen in Fig. 2. This weekly demand data is covers six years as it consists of total 312
weeks. First two years of the data is used as training data for the demand forecaster in the system.
1158 DONGname
Author Hai et/ Procedia
al. / Procedia Engineering
Engineering 15 (2011)
00 (2011) 1154 1159
000000 5

1300
1200 M10 M20 M40
1100
1000
900
800

Units
700
600
500
400
300
50 100 150 200 250 300
Time(weeks)

Fig. 2. Demand data used in the supply chain simulations

4.2. Simulations with a more realistic demand pattern

The demand involved also noise which made the control task even more realistic. The parameters
and needed to be retuned as those parameters used in the previous simulation resulted as very poor
responses. New parameters were chosen as follows: = 0.3, = 0.7, N1 = 3, N2 = 12, Nu = 20,the target
inventory level was kept constant in the level of 100 units. This is propably not the most cost efficient
way of managing an inventory but was used nevertheless to keep the results easy to understand. The
demand curve and inventory response can be seen in Fig. 3. As can be seen,when demand quick increase,
inventory level also suddenly changes to response to the change of market. The penaling of MPC control
actions forced the controller to order excessive amounts in order to avoid stock-out.

450 Demand
Inventory
350
Unites

250

150

50
20 40 60 80 100 120 140
Time (days)

Fig.3. The inventory responses of MPC in a Gaussian-shaped demand pattern

5. Conclusions

In this paper, firstly,MPC is applied to dynamic SCM planning, simulation results show MPC can
effectively control the SC system of high dynamic and uncertain, improve robust of the system, and
satisfy customers demand by the flexible transaction of constrains and reasonable safety stock. Secondly,
MPC technology adapts agility and collaborative in supply chain planning, itself can be acted as a good
support technology of supply chain network. Finally, in the future research, we hope to apply model
reference control to SCM, it can handle some problem of MPC in the complexity of mission,heavy
computing load, and dynamic oscillation of the data.
6 DONG HaiHai,et
DONG et al. /al/
Procedia Engineering
Procedia 1500
Engineering (2011) 1154
(2011) 1159
000000 1159

References

[1] Braun M.W.,et al. Application of MPC to robust management of multiechelon demand networks in semiconductor
manufacturing.Simulation, 2003,79(3):139-156.
[2] Alberto B.,et al. Model predictive control of hybrid systems with applications to supply chain management. 49th ANIPLA
National Congress,Napoli,Italy, 2005:15-30.
[3] William B.,et al. Distributed model predictive control for dynamic supply chain management. Proceedings of the
International Workshop on Assessment and Future Directions of NMPC, Freudenstadt-Lauterbad, Germany, 2005,p.26-30.
[4] Wang W S,Gong Y D,Yu P L. Networked manufacturing. Shenyang:Northeastern University Press,2003,p.7-11.
[5] Ydstie B E. A Model Predictive Control Strategy for Supply Chain Optimization. Computers and Chemical
Engineering,2003,27(8):1201-1218.