You are on page 1of 22

Computers and Chemical Engineering 140 (2020) 106966

Contents lists available at ScienceDirect

Computers and Chemical Engineering


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

A stochastic approach for integrated production and distribution


planning in dairy supply chains
Armando Guarnaschelli a, Héctor Enrique Salomone b, Carlos A. Méndez c,∗
a
Pontificia Universidad Católica de Valparaíso. Avda, Brasil 2147, Valparaíso, Chile
b
INGAR-CONICET-UTN, Avellaneda 3657, S3002GJC Santa Fe, Argentina
c
INTEC (UNL - CONICET), Güemes 3450, (3000) Santa Fe, Santa Fe, Argentina

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

Article history: This work addresses production and distribution planning for a real-world dairy supply chain. The plan-
Received 5 February 2020 ning model accounts for the production and distribution of Cheese, Yogurt, Powdered Milk and UHT milk
Revised 4 May 2020
products across a two-echelon Supply Chain. This task is undermined by the inherent variability of raw
Accepted 3 June 2020
materials and finished products demand. The integrated production and distribution planning method-
Available online 8 June 2020
ology introduced is based on a two-stage stochastic mixed integer linear programming formulation. In
Keywords: real-world settings the number of scenarios grows substantially; thus, a scenario reduction strategy based
Dairy supply chains on clustering techniques is given. A decomposition and solving strategy is also introduced and applied to
Supply chain planning a real-world case study. This study showed that the value of the stochastic solution might rise to 21.1%
Stochastic programming above the deterministic solution. This indicates the importance of considering uncertainty for dairy pro-
Production and Distribution Planning duction and distribution. Besides, different-size instances are tested to study the scalability of the solution
approach.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction supply chain performance Prajogo and Olhager (2012) . Moreover,


this demand for integrated production and distribution planning
Milk and dairy supply chains play an important role in a world has been recognized and analyzed for perishable products supply
with a continuously growing demand for food. The high nutritional chains Amorim et al. (2012).
value of milk and its by-products has been increased through In response to this fact, the present work addresses the
the use of fortifying additives. This has made them a primary operations planning (Supply Chain Operations Planning de Kok
source of nutrition for a significant portion of the world popula- et al. (2003)) in an integrated decision model that is intended to
tion, OECD/FAO 2018. . facilitate the coordination of supply chain functions and incorpo-
Only in recent years, milk and dairy supply chain production rates supply and demand variability.
and distribution planning have caught attention in the mainstream A real-world dairy supply chain motivates this research work;
technical literature Sel and Bilgen (2015). There are several issues its facilities are composed of 4 manufacturing plants and a set of
and complexities that arise when trying to model dairy supply 9 geographically distributed distribution centers (DCs) sourced by
chains, such as high variability in raw material supply, a large land transportation using dedicated trucks. The supply chain is af-
number of products, their perishability, and fluctuating demand. fected considerably by supply and demand variability.
Furthermore, there is a strong interaction between the sourcing, The planning model proposed for this supply chain, and others
production, and logistics departments for planning and scheduling with similar configurations are formulated as a two-stage stochas-
supply chain operations. Coordinating supply chain plans is usu- tic programming model, represented by a mixed-integer linear pro-
ally a difficult task since many tradeoffs between the functions gramming (MILP) formulation. The formulation accounts for most
carried out by these departments are to be considered, and of- relevant planning decisions in a planning horizon of one month.
ten these departments have different objectives. It is widely rec- The planning model decides on the following aspects of the
ognized that considering production and distribution planning in business cycle: production campaigns sizes, lengths, and work-
an integrated and optimized fashion will greatly improve current ing shifts for every production plant; inventory level decisions
for factory and DCs warehouses. Required refrigerated and non-

refrigerated trucks for replenishment shipments to DCs and the re-
Corresponding author.
E-mail address: cmendez@intec.unl.edu.ar (C.A. Méndez).

https://doi.org/10.1016/j.compchemeng.2020.106966
0098-1354/© 2020 Elsevier Ltd. All rights reserved.
2 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

sulting expected customer service levels. It also considers export scheme for some days. This scheme establishes the number of pro-
decisions. duction shifts taking place on a production day. Length-based cam-
The objective is to maximize the expected net benefits and paigns have a minimum and maximum run length for the pro-
maintain a stable operation as a result of maximizing incomes mi- duction plant. Each manufacturing plant produces only one fam-
nus operation costs, while satisfying inventory and perishability re- ily of dairy products, and there are four families: powdered milk
strictions. The novelty of the model and the approach to the solu- (PM), ultra-high temperature processing milk (UHT milk), yogurt,
tion is based on the following aspects: and cheese.
Production time and distribution lead times are considered to
• Raw milk transformation into the most relevant families of
be discrete in daily periods. If a product is manufactured on day
dairy products is considered. The model balances the distribu-
t and becomes available on the same day, its production time
tion of available milk into the different families of products,
is zero. If a production time is greater than zero, then work-in-
considering its impact along the supply chain.
process factory inventories are considered.
• There are two methods for campaign definition in manufac-
As this supply chain includes the production and distribution
turing plants according to two planning policies, length-based
of two families of fresh products, namely yogurt, and cheese, per-
campaign definition and working shifts-based campaign defini-
ishability needs to be considered when modeling their inventories
tion.
both at factory warehouses and DC warehouses.
• The model considers distribution centers replenishment opera-
Stochastic demand and supply are described through indepen-
tions and costs, consolidating shipments by land transportation.
dent demand distributions for each family of products and raw
• Additionally, the model includes variability in supply (raw milk)
milk. These distributions are represented by a set of scenarios and
and demand (of finished products) through the definition of
their probabilities. Demand allocation to DCs is considered to be
different scenarios conducting a two-stochastic programming
rule-based, that is, for every product family, the demand for a
formulation.
given DC is calculated as a percentage of total demand. Forecasting
In real world SCs the number of products, DCs, and stochastic techniques recommend this procedure to avoid the negative effects
scenarios grows significantly. Consequently, a twofold strategy is of the well-known bullwhip effect Chen and Lee (2017).
proposed. To cope with the vast number of scenarios, a scenario The objective for the planning problem is to maximize expected
reduction strategy based on well-known clustering algorithms is net benefits while maintaining a stable operation (including setup
introduced. Moreover, to account for the numerous dairy products tasks and costs). A feasible plan will also have to correctly use
a decomposition approach based on solving an aggregated version production capacity, fulfill inventory cycles constraints, and define
of the model and then refining it into a detailed product-specific production campaigns and necessary DCs replenishments. We em-
plan is used. phasize that the focus of this work is to optimize the material
The paper is organized as follows: first, the problem to be and product flows throughout a dairy SC. In the search for this
solved is stated, and related works are discussed. Then, the de- optimization we have intentionally neglected through abstraction
cision model is described together with the decomposition strat- many details of day to day manufacturing which are concisely pre-
egy and scenario reduction algorithm. Afterward, a case study of a sented in the work of Banaszewska et al. (2013).
real-world supply chain is described together with its set of sup-
ply and demand scenarios describing its stochastic nature. Numer- 3. Related works
ical experimentation on a set of different-sized instances is given.
Planning results are provided and analyzed. The applicability of the On top of day to day and month-to-month production plan-
approach is discussed, and finally, conclusions are given. ning and distribution, there is the need to optimally design risk
and uncertainty resilient Supply Chains. The following works an-
2. Problem statement swer this need. Design problems can be traced back to semi-
nal papers such as the one of Geoffrion and Graves (1974). Upon
This work is focused on dairy supply chains planning, an impor- this work, a relevant feature included is demand uncertainty. A
tant type of food supply chain. Dairy supply chains encompass the MILP 2-stage stochastic programming for multi-echelon SC net-
functions of daily sourcing of raw milk, processing raw milk into work design is introduced by Tsiakis et al. (2001), the authors use
dairy products, inventory management and logistics management a case study approach to incorporate stochastic demand. An effi-
of the physical flow of products from factory storage to distribu- cient scheme based on both Sample Average Approximation and
tion centers or customers. Benders decomposition is used to tackle a large-scale SC network
The problem addressed in this work deals with building an in- design problem by Santoso et al. (2005). The authors propose a 2-
tegrated production and distribution plan for a type of dairy supply stage stochastic formulation and solve real-world sized problems
chains. In this type of supply chain there is one industrial complex instances. A model that considers Large-Scale Supply Chain Design
with several manufacturing plants, or specific production lines for with Stochastic Inventory Management is introduced by You and
different families of dairy products. This industrial complex is sup- Grossmann (2008). The authors propose a non-linear integer for-
plied with raw milk on a daily basis by a set of associated milking mulation and solution strategies to cope with its complexity. Un-
yards. Raw milk is stored for a limited amount of time (shelf-life) certainty is considered by modeling safety stock constraints that
before being distributed to one of the manufacturing plants in the exploits risk pooling under normally distributed demand. In the
complex. Raw milk production quantities are not certain. Gener- same line of thought, the work of Brunaud et al. (2019), intro-
ally, they are influenced by weather and other difficult to measure duces continuous and periodic review inventory policies to general
variables. This work considers demand and supply variability to ac- SC planning problems.
count for this uncertainty. The work of Salema et al. (2007), introduces a comprehensive
Every manufacturing plant processes raw milk in-house to meet model for a generic multi-product reverse logistics network. Their
its corresponding production requirements. approach introduces uncertainty in demand and returns as a 2-
Manufacturing plants operate through the definition of produc- stage stochastic programming formulation. A design problem for
tion campaigns that can be defined, according to production and resilient supply chains under disruption risks at distribution cen-
management aspects, either as shift-based campaigns or length- ters is formulated as a two-stage stochastic program by Garcia-
based campaigns. Shift-based campaigns use a daily production Herreros et al. (2014). The method includes a strengthened multi-
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 3

cut Benders decomposition algorithm and the derivation of deter- by Touil et al. (2016). The first objective maximizes the benefits
ministic bounds based on the optimal solution over reduced sets considering costs of setups, storage, distribution, and pending or-
of scenarios. The results demonstrate the importance of including ders. The second objective maximizes customers service levels. The
DC capacity in the design problem and anticipating the distribution authors use a short-term planning horizon of 5 days and explic-
strategy in adverse scenarios. itly model transportation lead-times, perishability and production
The specific subject of this work is mainly related to three re- capacity constraints. The multi-objective nature is tackled using
search areas in Production and Distribution Planning (PDP) in Sup- fuzzy goal programming.
ply Chains (SCs), these are: deterministic approaches for PDP in
Dairy SCs, PDP in food and consumer goods SCs and specifically 3.2. PDP in food and consumer goods SCs
to stochastic approaches for PDP in SCs with a focus on stochastic
programming. The work of van Elzakker et al. (2014a) introduces a tactical
planning model for fast-moving consumer goods supply chains.
3.1. Deterministic approaches for in dairy SCs The model includes sourcing, production, and warehousing in fac-
tories and distribution centers and minimizes a global cost func-
Only recently, dairy supply chain planning has become the sub- tion (summing operational, setups, and penalizations: missed sales
ject of active research in mainstream literature. A thorough review and safety stock violations). Production capacities are considered
of models for supply chain management in dairy industries encom- in both mixing and packing stages. These stages define the pro-
passing works on different quantitative approaches and problem duction process. They propose an efficient SKU-decomposition al-
settings has been published by Sel and Bilgen (2015). The work gorithm that allows solving problems with lots of SKUs within a
of Akkerman et al. (2010) reviews quantitative operation manage- 52-week planning horizon (one year). The same authors, van Elza-
ment approaches for food distribution management. They focus on kker et al. (2014b), using the same planning model, introduce an
works where models consider manufacturing in different plants extension that considers three methods for implementing shelf life
and the shipment of products from these plants to retailers, possi- restrictions, augmenting the applicability of their model to perish-
bly through distribution centers. able SKUs. These methods are: direct shelf life implementation,
A majority of works in dairy SC planning are focused on which thoroughly tracks the lifetime of a product but requires
short-term (around a week or so) operations scheduling for complex mathematical models; indirect shelf life implementation
yogurt production, the works of Kopanos et al. (2011) and that gives an “aggregate” tracking of a product’s lifetime and re-
Kopanos et al. (2012) are representative of this research. In the duces the size of the model; and finally, a hybrid shelf life im-
work of Kopanos et al. (2011), the schedule is achieved by grouping plementation that attempts to use the best of both approaches.
products into families. Timing and sequencing decisions are made These important works are highly related to the production and
between families, and material balances are considered for every distribution settings (excluding uncertainty) in the proposal of this
product in a batch production environment. The same authors in manuscript, and its generality allows its application for dairy sup-
the work by Kopanos et al. (2012) extend the previously men- ply chain planning. Nevertheless, there are substantial differences
tioned production planning model from a broader perspective, fo- in the modeling of production processes and transportation. In
cusing on the supply chain of yogurt and similar semi-continuous those works weekly production levels are set, and setups decrease
food industries by including distribution operations scheduling. the available packaging time. In the present work, production cam-
The model also innovates on the solving strategy as they integrate paigns are defined with daily time grid considering the work shifts
a hybrid continuous-discrete timeline representation. and the total campaign lengths, while setups are explicitly mod-
A framework that through multi-objective optimization ex- eled and minimized. On the transportation side, their works as-
plores the advantages of integrating production and distribution sume that lead-times are zero. This last assumption is strong as
planning for perishable goods is presented by Amorim et al. (2012). it prevents the application of their proposal in geographically dis-
The addressed problem considers a set of production plants with tributed supply chains where lead-times have significant effects.
dedicated lines and producing multiple perishable items (which in Also, in the present work, transportation costs are associated with
fact could be dairy products). A block planning formulation is used, the number and type of trucks required for delivering products to
and different shelf life strategies are proposed. There is no inven- distribution centers rather than using a unitary transportation cost.
tory held at production plants and, at the end of each day, the pro-
duction output is delivered for distribution. Delivery is performed 3.3. Stochastic approaches for PDP in SCs
by using a 3PL (outsourced). Distribution centers are assumed to
have unlimited storage capacity, and there are no lead-time con- Uncertainty in dairy supply chains is modeled in the work of
siderations as products are delivered on the same day of produc- Guan and Philpott (2011), the authors present a multistage stochas-
tion. tic quadratic programming model to solve the problem of produc-
The work of Bilgen and Dogan (2015) formulates a mixed linear tion planning for a dairy company under an uncertain milk supply
model to address a production planning problem of a multi-stage and deterministic linear demand functions. A 12-month planning
process within the dairy industry. The objective is to determine horizon is considered, and monthly time-grid is used. Every month
when and how much of each intermediate and the final product represents a stage of the scenario tree, and for each of them, a set
should be produced in a 20-days planning horizon to maximize of independent random outcomes and their probabilities are given.
benefits, considering production, stocks, lost sales and setup costs. The model decides on monthly milk collection from multiple re-
In their work, the focus is production planning, and distribution gions and monthly production of a set of powder and perishable
aspects are not included. Production is thoroughly described, and products. Given the complexity of the mathematical formulation,
detailed, and executable production plans are obtained. Although a specific algorithm -Dynamic Outer-Approximation Sampling Al-
this research is illustrated with an application in the dairy indus- gorithm (DOASA) developed by Philpott and Guan (2008)- is used
try, it has the flexibility to apply to a wide variety of environments for convergence. They generate a production and sales policy us-
since it covers the main characteristics of many food production ing the model and evaluate it in a simulation experiment against a
systems. deterministic policy.
A bi-objective MILP model to face the problem of production In the work of Bilgen and Çelebi (2013), the authors modeled
and distribution in a supply chain of dairy products is presented the production and distribution scheduling of yogurt. Yogurt can
4 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

be produced in several production lines and then distributed from bust/resilient strategies under harvest time and yield disruptions
these lines to a set of distribution centers. Production, in their (modeled as scenarios).
work, is planned to use the minimum and maximum lot sizes and In the work of Sel et al. (2017), a model for planning and
emphasizing the packaging process. For a given demand day, pro- scheduling make-and-pack yogurt production is proposed. The
duction is carried out on a set of manufacturing lines. The quan- work considers a predefined production sequence and an uncer-
tities sent from the inventories of lines to distribution centers for tain lifetime of intermediate products. This lifetime is modeled us-
daily demand are the quantities produced on the same day. Uncer- ing a Weibull distribution which is included in the model through
tainty in operation times is tackled by running a simulation model chance constraints. Once the model is solved, simulation is used to
of plants operation based on the schedule obtained with the math- mimic the proposed schedule by the mathematical model and to
ematical model. If substantial differences are found, the scheduling determine the waste caused by quality decay.
model is run all over again. This process is repeated until a stop- An exhaustive analysis of how uncertainty is being addressed in
ping criterion is reached, and the output is the final schedule. the management of agricultural supply chains, from the perspec-
A MILP model for the design and planning of multi-period tive of modeling and solving methodology in the literature is pre-
and multi-product closed-loop supply chains (CLSC), with uncer- sented by Borodin et al. (2016). Also, literature gaps and possible
tain levels in the amount of raw material supplied and customer directions for future research are introduced.
demands, is presented by Zeballos et al. (2014). The objective is
to minimize the total cost of the network, determining the facili-
3.4. Contributions
ties that will be opened, the products and quantities to be manu-
factured, transported, stored and recycled to fulfill clients demand.
The research aims at making a twofold contribution to the plan-
Uncertainty is modeled as a multi-layered scenario tree since de-
ning and scheduling literature in the dairy industry:
mand and supply are considered dynamic parameters that change
- On the process modeling features: Two types of production
during each period (two are considered). As the number of sce-
campaigns are explicitly modeled, campaigns that are defined ei-
narios grows, the model turns to be impractical, to overcome this
ther by an assignment of working shifts for a period or by their
issue a scenario reduction technique based on the SCENRED algo-
length and size. Moreover, setups arising from starting new cam-
rithm is used in order to get a reasonably good approximation of
paigns are included. Inventory considerations, balances, and per-
the original problem. The results show the relevance of consider-
ishability on both production facilities and distribution centers are
ing the stochastic multi-stage approach, instead of deterministic,
addressed. Additionally, for some families of products, work-in-
when considering multiple-period problems with uncertainty. It is
progress inventories need to be modeled, e.g., cheese, which is
relevant to point out that as stated in the work of Beraldi and
modeled as a two-phase production. Distribution aspects such as
Bruni (2013), SCENRED considers the original set of scenarios and
lead times, shipping lots, and requirements over distribution re-
determines how many scenarios to delete or preserve in its back-
sources (trucks) are explicitly modeled to account for a better rep-
ward and forward strategies respectively. Moreover, for each sur-
resentation of logistics activities. Finally, this work considers a 30-
viving scenario, a new probability is computed, adding to its orig-
day planning horizon and focuses on production and distribution
inal probability value, the probabilities of every deleted scenario
flows rather than detailed production scheduling.
closer to it. Despite being a successful approach, deleting an ar-
- On the uncertainty modeling features: As reviewed above, sev-
bitrary number of scenarios eliminates essential information and
eral proposals directly or indirectly include uncertainty in their
only preserves the variability described by the surviving scenario
planning models. The approach in this work explicitly models un-
set. To avoid this compromise, the authors in the aforementioned
certainty both in supply and demand. A two-stage stochastic pro-
work by Beraldi and Bruni (2013) propose a clustering scenario re-
gramming approach is adopted, and the effort to solve it is justi-
duction algorithm.
fied by computing the value of the stochastic solution on a real-
In the work of Keyvanshokooh et al. (2016), a two-stage
world case study. Independent demand distributions (represented
stochastic programming formulation is developed for a closed-loop
by individual scenario sets and their probabilities) are introduced
supply chain network design problem in the presence of uncer-
for every family of products and raw milk supply. To cope with
tainty in transport costs and customer demand and return quan-
the scenario explosion due to the size of the scenario tree (every
tities. In the first stage, strategic decisions are determined, such
leaf represents a possible demand outcome for each product and
as locations and capacities of the facilities, as well as distribution
a possible raw milk availability), a scenario reduction methodology
stock levels, and in the second stage, operational decisions, includ-
based on clustering is introduced.
ing production amounts and shipments between the entities of the
The features above -to the best of our knowledge- have not
network, are determined to maximize expected benefits. They as-
been previously described all together in the literature, as shown
sume the availability of historical data for the distribution of trans-
in Table 1. Table 1 classifies the literature and this work accord-
port costs for previous sales and model this uncertainty with an
ing to a set of dimensions grouped into process modeling features
algorithm for scenarios generation and reduction. Then, in each
(production and distribution processes) and uncertainty modeling
transportation cost scenario, polyhedral uncertainty sets for de-
features (how uncertainty is included in the modeling formula-
mands and returns are defined. An accelerated stochastic Benders
tion). This work attempts to fill a gap in the literature, allowing de-
decomposition algorithm is introduced as a solution approach.
cision makers and supply chain managers operate more efficiently
In a recent study Behzadi et al. (2017) introduced a two-stage
in real-world scenarios.
stochastic programming model with profit/cost optimization objec-
tives for supply chain operations is developed. It includes an ex-
ponential distribution to model probabilistic perishability. In the 4. Modeling strategy
first stage, the optimization model selects risk management strate-
gies from a set of robust/resilient strategies. In the second stage, In this section, the structure of the mathematical formulation
the model makes sourcing, inventory, distribution, and sales deci- for the planning problem being solved in this work is introduced.
sions based on the actual realization of disruption scenarios. The Here, its most relevant aspects are highlighted without resorting to
model maximizes the expected profit while satisfying constraints the full constraint set that defines them. First, the general model-
in supply, demand, and transport capacities. The model is applied ing strategy is introduced. Second, detailed aspects of this strategy
to a kiwifruit supply chain to investigate the effectiveness of ro- are discussed.
Table 1
A comparison table for the reviewed literature.

Finite
Distri-
bution
Production (Num-
Cam- ber of Solution

A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966
Planning Product / Application Supply chain Product paigns / Vehi- ap-
Paper Horizon Shelf-life Families industry Processes Stages Shifts cles) Lead Times SC Echelons Setups Uncertainty proach

Bilgen and Dogan (2015) 30 days OF + C P + PF Milk PRD-STR RM-WIP-FP – – – MP – MILP

Amorim et al. (2012) 10 days OF + C P Perishable PRD-STR O-FP – – – MP+DC (SD) – MILP
products
√ √
Touil et al. (2016) 5 days OF + C P Dairy PRD-STR-DST O-FP – – MP+DC (SD) – Bi-
objective
MILP + FZ

Kopanos et al. (2011) 7 days – P + PF Yogurt PRD O-FP – – – MP – MILP
√ √
Kopanos et al. (2012) 8 days – P + PF Dairy PRD-STR-DST O-FP – – MP+DC (SD) – MILP

van Elzakker et al. (2014a) 52 weeks – P FMCG PRD-STR-DST O-FP – – MP+WH+DC –

(NEMC) MILP + SKUA

van Elzakker et al. (2014b) 52 weeks OF + C P FMCG PRD-STR-DST O-FP – – MP+WH+DC –

(NEMC) MILP + SKUA

Keyvanshokooh et al. (2016) 16 periods – P PRD-STR-DST O-FP – – – MP+WH+DC – (TC+D + R) HRSP
Generalized
CLSC

Guan and Philpott (2011) 12 months – P Dairy MC-PRD-STR RM-FP – – – MP – (MS) Multi
Stage SP
√ √
Behzadi et al. (2017) 12 months C P SRC-STR-DST FP – – S+D – (S + M) Two-
Agribusiness Stage
SC SP

Zeballos et al. (2014) 10 years – P PRD-STR-DST FP-RS – – – ME – (S + D) Multi
Generalized Stage
CLSC SP + SCD
√ √ √
Bilgen and Çelebi (2013) 5 days OF + C P Yogurt PRD-STR-DST FP (SH) – – MP+DC (SD) (OT) HRSP
√ √
Sel et al. (2017) 24 H C P Yogurt PRD WIP-FP – – – MP (SD) (IM) MILP +
CC + SM
√ √ √ √
This research 30 days C P + PF Dairy PRD-STR-DST RM-WIP-FP MP+WH+DC (S + D) Two-

(SH+CA) stage
SP
Notes:.

Shelf-Life: OF-Objective Function, C–Constraint.

Product/Family: P-Product, PF-Product Family.

Supply Chain Processes: SRC-Sourcing, MC-Milk Collection, PRD-Production, STR-Storage, DST-Distribution.

Product Stages: RM-Raw Material, WIP-Intermediate Product, FP-Final Product, O-FP-One Stage Final Product, FP+RS-Final Product + Recycling Stage.

Production Campaigns / Shifts: NEMC-Enforced but Not Explicitly Modeled Production Campaign, SH-Shifts, CA-Production Campaign.

SC Echelons: MP-Manufacturing Plants, WH-Warehouse, DC-Distribution Center, S-Supply, d-Demand, ME-Multi-echelon.

Setups: SD-Sequence-Dependent Setups.

Uncertainty: TC-Transportation Cost, d-Demand, R-Returns, MS-Milk Supply, S-Supply, M-Market Price, OT-Operation Times, IM-Intermediate Mixture Lifetime.

Solution Approach: MILP-Mixed Integer Linear Programming, SP-Stochastic Programming, HDRSP-Hybrid Robust Stochastic Programming, SM-Simulation, CC–Chance Constraint, FZ-Fuzzy Goal Programming, SCD-SCENRED, SKUA-
SKU Decomposition Algorithm.

5
6 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

ning Model) is then solved by defining the production quantity for


each product. In the same model, every replenishment (products
volume mix and required trucks) to distribution centers are set.
Inventory levels considering understock and overstock for every
product are set, and specific export quantities are defined. All these
decisions must fulfill production and export levels previously set in
the Family Aggregated Model. A complete planning model would
decide on production campaigns, individual products production,
and distribution altogether. This decomposition, which relies on
the definition of product families reduces the size of the underly-
ing decision models. Consequently, computational effort is reduced
as well. Numerical experiments are later shown in Section 8.
Sets and indices
T Planning horizon. t ∈ T = {1..hl }
F Product Families. f ∈ F
S Scenarios. s ∈ S
Fig. 1. First and second stage decisions. FT Family types according to their cooling requirements. i ∈
F T = { F resh , Dry }
MP Manufacturing Plants. m ∈ MP
This work introduces a two-stage stochastic programming ap- CT Production campaign types. CT =
proach for dairy supply chain planning. The rationale to define { LengthBased , Shi f tBased }
which decisions are taken in the first stage and which in the sec- L Distribution centers. l ∈ L.
ond is the following. First stage decisions need to be known by
every actor in the chain before the beginning of the planning hori-
First-Stage parameters (scenario independent)
zon. Second stage decisions are introduced as a consequence of
hl Planning horizon length.
supply and demand variability. Fig. 1 summarizes the structure of
ftyf The type (Fresh or Dry) of family f.
the stochastic programming model.
ctym Campaign type for production plant m. ctym ∈ CT
On the one hand, Production, Shipments, and Export levels are
fprm fpr is the family produced by manufacturing plant
the first stage or here-and-now decisions. This is because these
m. fprm ∈ F
decisions imply several compromises with internal and external
fyf family f production yield for ua nit of processed raw
stakeholders, e.g. for exports, space in ships must be pre-scheduled
milk.
through the freight forwarder and also the reception by the client
rsc Raw milk third supplier cost.
must be synchronized.
roc Raw milk over stock cost per volume unit.
These compromises must be established before executing the
elf Export lot size in metric tons of family f.
production and distribution plan. In fact, changes are not desir-
able during execution. On the other hand, DC and Raw Milk in-
τl Supply lead time for distribution center l. It equals
the estimated transportation time from factory ware-
ventory related decisions are modeled as a second stage or wait-
house to distribution center.
and-see decisions as uncertainty are modeled in DC final products
i0f, l Family f Initial inventory at location l.
demand and raw milk supply. The rationale behind defining Fac-
tory Inventory as a first stage decision is a fact that it depends tlmin , tlmax Minimum and maximum truckload capacity, respec-
only on changes in production and distribution decisions and both tively.
are second stage decisions. r0 Raw milk initial inventory.
We emphasize that in many cases, and specifically in this type rmax Maximum raw milk inventory.
dmmax Maximum campaign length (in days) for plant m.
of SC a production and distribution plan that minimizes the cost
min , cmax
cm Minimum and maximum daily production capacity
of absorbing the risk in supply and demand is needed. The varia- m
tions that are incurred are absorbed through external suppliers or at manufacturing plant m.
through the cost of shortages on the domestic consumption side αm Setup time in periods for production plant m.
through distribution channels (DCs). This is not just a regionalism, ostm Remaining days to finish an ongoing setup task at
the literature emphasizes that proposing a production and distri- manufacturing plant m.
bution plan for a short period of time (15 or 30 days) and modi- wpm, t This parameter accounts for work-in-progress inven-
fying it in the face of fluctuations is counterproductive, Li and Dis- tory from the previous planning horizon to be fin-
ney (2017); Schuh et al. (2019). ished on day t : t ≤ σ m
Production processes are modeled accounting both for manu- elf min , elf max Minimum and maximum number of lots of family f
facturing plants campaign definitions and capacities, as well as for to be exported, respectively.
the families of dairy products processed by them. As products from ism Maximum portion of the total capacity that can be
the same family have a common production process (except for left idle during a production campaign at manufac-
packaging, flavoring, etc.), the idea is to solve the complete produc- turing plant m with [0,1) value.
tion and distribution model considering only product families (the ω ff w Factory warehouse shelf-life of products from family
demand is aggregated). This allows to clearly define the levels of f.
production per manufacturing plant/production line for each prod- ωdcf
Distribution center warehouse shelf-life of products
uct family, and also the quantities to be exported from each family. from family f.
These values are set considering a complete distribution plan and rrf Revenue from reduced price selling of family f over
respecting all inventory and perishability constraints (Family Ag- stock (distressed sales).
gregated Model). rf Revenue from selling one ton of family f in any DC
From this solution, production levels (campaigns) and export of the Supply Chain.
targets by family are kept. A new decision model (Detailed Plan- ref Revenue from exporting a batch of family f.
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 7

imax
i,l
Maximum inventory at location l for fresh and dry IDs, f, l, t Inventory of family f in distribution center l on day t
family types. under scenario s.
max
zm For shift-based scheduled manufacturing plants, this
Binary variables
parameter takes the value of a maximum number
R1m, t When R1m,t = 1, it implies that in manufacturing plant
of shifts. While, for length-based schedules manufac-
m the production of the previous time period (t − 1) is
turing plants, takes value 1.
maintained in the current time period (t). Otherwise, it
scm The production capacity of manufacturing plant m
takes value 0
during one work shift.
R2m, t When R2m,t = 1, it implies that in manufacturing plant
βm Deterioration coefficient for products manufactured
m there was a change in production between one time
at plant m, e.g., if x is the quantity produced, (1 −
period (t − 1) and the next time period (t) (start or end
βm ).x is the usable production after quality testing.
of campaign). Otherwise, it takes value 0.
σm σm + 1 is the process time in periods for family pro-
Ym, t When Ym,t = 1, it means that a campaign starts on t at
duced in manufacturing plant m. If σ m > 1, work-in-
manufacturing plant m. Otherwise, it takes value 0.
process inventory modeling is required. When σm =
0, production is finished on the same day it started. Integer variables
iwip0m Work-in-progress inventory from the previous plan- TRi, l, t The number of trucks assigned for transporting i type
ning horizon at manufacturing plant m. family products from factory warehouse to DC l. Depart-
tci, l Transportation cost from manufacturing complex to ing in period t and arriving at l in period (t + τi ).
distribution center l. Ef, t family f daily export on day t measured in 25 tons lots.
sco Setups cost. Zm, t Campaign indicator variable. For length-based campaigns
if Zm,t = 1, there is an ongoing campaign in manufactur-
Second-Stage parameters (scenario dependent)
ing plant m at day t and production is allowed. For shift-
dps, f, l, t Family demand for distribution center l on day t under
based campaigns if Zm,t = x, there is an ongoing cam-
scenario s.
paign in manufacturing plant m with x production shifts
ρs The probability of scenario s.
at day t, so the corresponding production is allowed.
dris, t Raw milk daily input on day t under scenario s. max
Zm,t ∈ 0..zm
First-Stage decision variables (Scenario independent)
EXI Export income. 4.1. Campaigns definition
ENB Expected Net Benefit.
TCOST Total costs. When in manufacturing, lower and upper bounds are imposed
SN Number of setups due to starting a new campaign. on a contiguous production run, or production quantities within
RMt Raw milk consumption in period t. this run has to be in multiples of a batch size, we need to per-
IFf, t Family f factory inventory on day t. form campaign planning. A campaign is defined in the work of
FPf, t Family f production on day t. Suerie (2005) as “the production amount of a specific product type
Vi, l, t Fresh and dry (F T = { F resh , Dry }) quantity of one contiguous production run which can and generally will
shipped to distribution center l on day t. span several planning periods”. In that work, the author explores
DVf, l, t Family f quantity shipped to distribution center l on different production campaign types defined by their features and
day t. their possible formulations. As mentioned, campaign modeling in-
Am, t Length of an ongoing campaign in plant m in period cludes lower and upper bounds for production quantities. These
t. Its maximal value is the maximum campaign length quantities may be subject to minimal and maximal run campaign
allowed for plant m : ct ym = lenghtBased , Am,t ∈ lengths or may have to be multiples of predefined batch sizes. On
0.. dmmax the modeling side, campaigns differ on how time is represented, in
MOm, t Manufacturing plant m output at day t. either continuous or discrete time formulations. In between these
IWIPm, t Inventory of products in the process of completion at representation strategies non-uniform time slots and event points
manufacturing plant m on day t. arise to capture campaign start times and durations. In our work,
Qm, t The quantity of families that begin to be produced at we make focus on production campaigns arising in the dairy in-
manufacturing plant m on day t. dustry. In these industries, the output of a plant is conditioned by
Z1m, t Component of the campaign indicator variable at man- its capacity to process milk into a dairy product and then pack-
ufacturing plant m on day t. ing it into a final product. Usually, in dairy industries, the packing
Z2m, t Component of the campaign indicator variable at man- stage does not diminish the nominal daily production capacity of
ufacturing plant m on day t. the plant, even though this stage requires specific scheduling mod-
Auxm, t Represents the product (R1m,t .Zm,t−1 ). els such as (Kopanos et al. (2011)-Kopanos et al. (2012)) . In the
context of this hypothesis, which arises from the case study in-
Second-Stage decision variables spiring this work, the model assumes that: production lines within
SAs, f, l, t Family f sales quantity at distribution center l on day t a manufacturing plant have a daily production capacity which is
under scenario s. bounded by the number of working shifts during the day or by the
SOs, f, l, t Family f stock out quantity at distribution center l on length/size of the production campaign in course. We emphasize
day t under scenario s. that every manufacturing plant requires a given type of campaign
OSs, f, l, t Family f over stock quantity at distribution center l at planning, and, this is captured in our model through the parameter
day t under scenario s. ctym , campaign type for manufacturing plant m.
RCs Raw milk over stock costs under scenario s. To model both length-based campaigns and shift-based cam-
RSs, t The volume of raw milk acquired on day t from a third paigns, the campaign indicator variable Zm,t ∈ 0..zm max is used. For

supplier under scenario s. manufacturing plant m under a length-based campaign planning


ROs, t Raw milk over stock which is thrown away on day t fashion zmmax is 1, whereas for shift-based campaigns zmax equals
m
under scenario s. the number of shifts available per day. In this way, for length-
RIs, t Raw milk inventory on day t under scenario s. based campaigns if Zm,t = 1, there is an ongoing campaign in man-
8 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

ufacturing plant m at day t and production is allowed. For shift- 4.3. Variability in demand and supply
based campaigns if Zm,t = x, there is an ongoing campaign with x
production shifts at day t, so the corresponding production is al- In this work, demand forecasts are assumed to be done at the
lowed. family level. The benefit of this assumption is two-fold; on the one
hand, it provides a higher level of accuracy for the forecast as it is
4.1.1. Length-based campaign definition explained in Ross (2004). On the other hand, it allows a better and
Manufacturing plants, indexed by m, which have a maximum more tractable description of demand scenarios for each product
campaign length, require the use of an accumulator variable Am, t family.
representing the number of production days of the on-going cam- For any family of dairy products and every planning horizon,
paign at the end of day t. This variable is updated each day as in the probability distribution of its monthly demand is expressed
Eq. (1), with the value of Zm, t is always nonnegative and cannot as a scenario set with its non-negative probabilities that sum 1.
surpass the max campaign length value (dm max ). Therefore, a cam- There are four families of products, and, fli is family l demand un-
paign restricted by length is modeled by defining the remaining der demand scenario i and P(fli ) is its probability. Raw milk sup-
production quantity, as in (1),(2). ply variability is described in the same way fr, m , is raw milk ex-
pected supply under supply scenario m. Family demand and raw
Am,t = Am,t−1 + Zm,t , ∀m ∈ MP, t ∈ T (1)
milk supply distributions are assumed to be independent. Under
this premises, a given scenario is composed by the combination
0 ≤ Am,t ≤ dm
max
, ∀m ∈ MP, t ∈ T (2) of an independent demand scenario for every family and for raw
milk supply [f1i , f2j , f3k , f4l , frm ] and its probability is given by the
Eqs. (25) to (28) formalize length-based campaigns.
product P(f1i ).P(f2j ).P(f3k ).P(f4l ).P(frm ).
A simplifying, yet realistic assumption is that there is a monthly
4.2. Perishability
forecast of demand and supply. This forecast is later disaggregated
either uniformly or using a disaggregation rule into daily demand
Perishability of refrigerated products, such as cheese and yo-
and supply parameters. Further disaggregation is required for DC
gurt, is assessed by modeling their shelf-life along the supply
demand and for characterizing product-specific demand, as the
chain. This work uses an indirect modeling approach similar to
forecasting function is performed aggregating DC and family de-
the approaches proposed by Ekşioğlu and Jin (2006) and later ex-
mand dimensions, this avoids the denominated bullwhip effect.
tended by van Elzakker et al. (2014b). This indirect approach limits
The disaggregation is represented in the model with parameters
the inventory levels to the amount of product to be dispatched in
dris, t , s ∈ S, t ∈ T, for raw milk supply on day t under scenario s,
the next ω periods, with being ω the shelf-life of the product.
and dps, f, l, t , s ∈ S, f ∈ F, l ∈ L, t ∈ T, for demand plan for family f on
Here, the maximum number of days a product can remain
day t at DC l under scenario s.
within the chain before being delivered to a retailer is used and
In Section 5 of this work the deterministic equivalents of the
is denoted as ω f or ωdc
fw
, accounting for the time a product fam-
f stochastic Family Aggregated and Detailed Planning Models (in-
ily can remain in the factory warehouse and in the DC warehouse cluding randomness in demand and supply), are given and dis-
respectively. cussed. A deterministic equivalent is a closed form mathematical
In the model, constraints guarantee that a product in the fac- formulation of a generic stochastic optimization problem as de-
tory warehouse has been shipped to a DC before its maximum fined in the work of Shapiro et al. (2009). These models rely on
ω ff w . Also, they ensure that it has been delivered to a retailer the valid hypothesis that there is a finite set of scenarios and their
before achieving its maximum ωdc f
. In the following constraints, probabilities are known, and an expected value function is being
DVf, l, t represents a shipment of family f products from factory optimized.
warehouse to DC l, dispatched on day t. Inventory at factory ware-
house and DC is represented by variable IFf, t and the variable 5. Model formulation
IDs, f, l, t , respectively.
 In this section, a hierarchical stochastic planning approach is
IFf,t ≤ DV f,l,t  , ∀ f ∈ F, t ∈ T presented in order to develop the production-distribution plan
l∈L (3) for a 1-month planning horizon. In the proposed hierarchical ap-
t  : t < t  ≤ t + ω ff w proach, firstly, an aggregated planning model reflecting the sys-
 tems physical constraints that provide an initial plan over the plan-
IDs, f,l,t ≤ DV f,l,t  , ∀s ∈ S, f ∈ F , l ∈ L, t ∈ T ning horizon is introduced. Secondly, specific product manufactur-
t  :t <t  ≤t +ωdc
(4)
f ing is planned using the output of the previous model to solve
a Detailed Planning model. The Aggregated model, introduced in
When t + ω f > hl, or t + ωdc
fw
f
> hl, an average of previous demand subSection 5.1, specifies the campaign definitions, levels of pro-
shipments is used. Additionally, as previously stated, every family- duction per manufacturing plants for each product family, and also
based constraint is refined in the Detailed Planning model, includ- quantities to be exported from each family. SubSection 5.2 intro-
ing shelf-life constraints which are individually considered for ev- duce the Detailed Planning model which refines the production-
ery product. The formulation of shelf-life constraints for factory distribution plan considering production quantities per product,
warehouse is given in Eq. (42) - (43), while for any DC is given specifies the export batch size of each product and the replenish-
in Eq. (44) - (45). ment to distribution centers.
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 9

5.1. Family aggregated model

5.1.1. Raw milk supply consumption


The following constraints model raw milk inventory flow within the industrial complex. In some scenarios the purchase of raw milk to
a third-party supplier (RSs, t ) or its disposal (ROs, t ) due to overstock may arise. For every scenario, dris, t is raw milk daily input (parameter)
and RMt (independent of scenarios) is the variable modeling the raw milk consumption of all plants in the complex.
   
RIs,t = r0 + dr is,t1 − RMt1 + RSs,t1 − ROs,t1
t 1∈T :t 1≤t t 1∈T :t 1≤t t 1∈T :t 1≤t t 1∈T :t 1≤t (5)
∀s ∈ S, t ∈ T
Raw milk storage size limitation constraint.
RIs,t ≤ r max , ∀s ∈ S, t ∈ T (6)
If Qm, t is the production quantity at day t in manufacturing plant m, and parameter f y( f prm ) is the amount of product of family f obtained
with one unit of raw milk, then, total raw milk consumption (RMt ) at day t in the manufacturing complex is calculated as follows.

 
RMt = Qm,t
f y( f prm )
, ∀t ∈ T (7)
m∈MP

5.1.2. General production constraints


Family f production of all plants in the complex is equal to the manufacturing output of plants producing f.

F Pf,t = MOm,t , ∀ f ∈ F, t ∈ T (8)
m∈MP: f prm = f

The value of MOm, t is modeled in constraints (9) and (10) where manufacturing plant output on day t is equal to the production that
started on day t − σm . Where σ m is the process time in periods for family produced in manufacturing plant m. This is due to the fact that
for some manufacturing plants and families of products, production taking place on day t is readily available on day t + σm . Parameter β m
accounts for the deterioration coefficient. Eq. (10) accounts for production that is readily available before σ m in the timeline, here, wpm, t
is a parameter that represents an amount of production that took place in the previous planning horizon.
MOm,t = (1 − βm ).Qm,t−σm , ∀m ∈ MP, t ∈ T : t > σm (9)

MOm,t = (1 − βm ).w pm,t , ∀m ∈ MP, t ∈ T : t ≤ σm (10)

5.1.3. Work-in-progress (WIP) inventory constraints


Manufacturing products with σ m > 0 generate WIP inventory which is depleted by the volume of finished products in period t rep-
resented by the variable MOm, t . Parameter iwip0m represents inventory from the previous planning horizon at manufacturing plant m.
 
IW IPm,t = iwip0m + Qm,t1 − MOm,t1 , ∀m ∈ MP, t ∈ T : σm > 0 (11)
t 1∈T :t 1≤t t 1∈T :t 1≤t

5.1.4. Production campaigns


To model production campaigns, it is crucial to determine when production campaigns start or end. Instead of modeling both starts and
ends straightforward, this work introduces an indirect formulation that accounts only for changes on the production status. A campaign
indicator variable represents this production status. This allows defining the start or end by examining changes in the campaign indicator
variable. This requires an elaborate set of constraints, that was built focusing both on adequate problem representation and computational
efficiency. It is described followingly.
max is composed as shown in constraint (12), where Z1
First, the campaign indicator variable Zm,t ∈ 0..zm m, t and Z2m, t are continuous
variables with the same domain as Zm, t and are limited from above as in constraints (13) and (14). Constraint (15) enables only one of the
variables that compose Zm, t in a time period t, trough the binary variables R1m, t and R2m, t . In this way, when R1m,t = 1, production level
(and campaign) of the previous time period is maintained in the current time period (constraint (14)), whereas when R2m,t = 1, it means
that there was a change in production between one time period and the next time period (which implies start or end of campaign). Con-
straint (16) ensures that no campaign is started in the first period of the planning horizon because configuration tasks must be performed.

Zm,t = Z 1m,t + Z 2m,t , ∀m ∈ MP, t ∈ T (12)

Z 1m,t ≤ zm
max
.R1m,t , ∀m ∈ MP, t ∈ T (13)

Z 2m,t ≤ zm
max
.R2m,t , ∀m ∈ MP, t ∈ T (14)

R1m,t + R2m,t = 1, ∀m ∈ MP, t ∈ T (15)

R1m,1 = 1, ∀ m ∈ MP (16)
Second, for every campaign type, it is mandatory that if Zm,t = l, then Zm,t+1 = l, unless a new campaign starts or an ongoing campaign
ends in t + 1, this is only possible when R1m,t = 1 and it’s ensured by constraints (17) and (18).
Z 1m,t ≤ Zm,t−1 , ∀m ∈ MP, t ∈ T : t > 1 (17)
10 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

Z 1m,t ≥ R1m,t . Zm,t−1 , ∀m ∈ MP, t ∈ T : t > 1 (18)

Constraint (18) involves a nonlinearity in its right-hand side which can be easily linearized as it is the product of a binary variable and a
continuous variable as in constraints (19) - (22). The variable Auxm, t represents the product of variables (R1m,t .Zm,t−1 ).

Auxm,t ≤ Zm,t−1 , ∀m ∈ MP, t ∈ T : t > 1 (19)

Auxm,t ≥ Zm,t−1 − zm
max
.(1 − R1m,t ), ∀m ∈ MP, t ∈ T : t > 1 (20)

Auxm,t ≤ Z 1m,t , ∀m ∈ MP, t ∈ T (21)

Auxm,t ≤ R1m,t .zm


max
, ∀m ∈ MP, t ∈ T (22)

5.1.5. Length-based campaigns


The level of production capacity during a production campaign of a length-based plant is set in constraints (23) and (24).
Qm,t
min
cm
≤ Zm,t , ∀m ∈ MP, t ∈ T : ct ym = lengthBased (23)

Qm,t
max
cm
≥ Zm,t , ∀m ∈ MP, t ∈ T : ct ym = lengthBased (24)

In this type of campaigns, a variable Am, t accounts for accumulated production days at manufacturing plant m on day t. This accumula-
max represents maximal value of A
tion (constraints (25) and (26)) continues until the current campaign ends. Parameter dm m, t . If a campaign
ends on day t, then the mandatory accumulation of production on day t + 1 (i.e. Am,t+1 ) is relaxed (constraint (26)). This allows to set the
accumulator variable in a new campaign to zero in a posterior campaign. In these equations the Boolean variable R2m, t (previously de-
fined), is used to model the start or end of a length-based campaign. Constraints (27) and (28) define upper and lower bounds for Am, t .
max
dm
The quotient min , represents the maximum number of campaigns that might take place for manufacturing plant m within the current
cm
planning horizon.

Am,t ≤ Am,t−1 + Zm,t , ∀m ∈ MP, t ∈ T : ct ym = lengthBased ,t > 1 (25)

 max

Am,t ≥ Am,t−1 + Zm,t −
dm
min .R2m,t
cm
, ∀m ∈ MP, t ∈ T : ct ym = lengthBased , t > 1 (26)

Am,t ≥ Zm,t , ∀m ∈ MP, t ∈ T : ct ym = lengthBased (27)

max
Am,t ≤
dm
min ,
cm
∀m ∈ MP, t ∈ T : ct ym = lengthBased , t > 1 (28)

Accumulated production at the beginning of the horizon. The following equations (constraints (29) and (30)) model the remaining quantity
production for the special case of t = 1.

Am,1 ≤ Zm,1 , ∀m ∈ MP : ct ym = lengthBased (29)

max
Am,1 ≤
dm
min ,
cm
∀m ∈ MP : ct ym = lengthBased (30)

 max

Am,1 ≥ Zm,1 −
dm
min .R2m,1
cm
, ∀m ∈ MP : ct ym = lengthBased (31)

5.1.6. Shift scheduled based campaigns


The level of production capacity during a production campaign for a shift scheduled plant is set in constraints (32) and (33). It is set
according to the number of shifts defined by the production campaign indicator (Zm, t ). In these equations scm represents the production
capacity of manufacturing plant m on one work shift. The parameter ism in (0,1] is the maximum portion of the capacity of a shift which
can be idle.
Qm,t
scm
≤ Zm,t , ∀m ∈ MP, t ∈ T : ct ym = shiftBased (32)

Qm,t
scm .(1−ism )
≥ Zm,t , ∀m ∈ MP, t ∈ T : ct ym = shiftBased (33)
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 11

5.1.7. Campaigns setups


In order to model these features, the binary variable Ym, t is introduced. This variable takes value 1 when a new production campaign
starts at t, if and only if R2m, t > 0 and Zm,t−1 = 0, which is ensured by constraints (34) - (35) – (36). We emphasize the redundancy of
constraint (36) which was included to improve the computational performance of the family aggregated model.
If at any period t, a production campaign for a manufacturing plant start (Ym,t = 1), then on periods t − t 1 : t 1 ∈ 0..(αm − 1 ) setup
tasks may be required and therefore production is not allowed in these periods . In this way if Ym,t = 1 keeps Zm,t −t 1 = 0 (there is no
production) until finish the setup task (constraint (37)). Now for the special case that at the beginning of the horizon there is a setup task
in progress, this is reflected in the parameter ostm > 0, constraint (38) keeping campaign indicator variable to 0 until the task is finished.

Zm,t−1 ≤ zm
max
.(1 − Ym,t ), ∀m ∈ MP, t ∈ T : t > 1 (34)

R2m,t ≥ Ym,t , ∀m ∈ MP, t ∈ T (35)

R2m,t − Zm,t−1 ≤ Ym,t , ∀m ∈ MP, t ∈ T : t > 1 (36)

Zm,t −t 1 ≤ zm
max
.(1 − Ym,t ), ∀m ∈ MP, t ∈ T , t1 ∈ 0..(αm − 1 ) : αm > 0, (t − t1 ) ≥ 0 (37)

Zm,t ≤ 0, ∀m ∈ MP, t ∈ T : t ≤ ostm (38)

5.1.8. Factory inventory balances


The following constraints model the filling of factory inventory with finished products and the shipments and exports that deplete this
inventory.
    
IFf,t = i0 f,F W + F Pf,t1 − DV f,l,t1 − el f .E f,t1
t 1∈T :t 1≤t l ∈ L t 1∈T :t 1≤t
(39)
t1 ∈ T : t1 ≤ t
∀ f ∈ F, t ∈ T
5.1.9. Inventory at DCs
A shipment departed from factory inventory at period t, arrives to a distribution center l at period (t + τl ), where τ l is the lead time for
l. Inventory at distribution centers is scenario dependent accordingly to future demand realization (dps, f, l, t ). Possible inventory fluctuation
due to understock and overstock quantities are represented by scenario variables SOs, f, l, t , OSs, f, l, t , respectively.
   
IDs, f,l,t = i0 f,l + DV f,l,t1 − d ps, f,l,t1 + SOs, f,l,t1 − OSs, f,l,t1
t 1∈T :(t 1+τi )≤t t 1∈T :t 1≤t t 1∈T :t 1≤t t 1∈T :t 1≤t (40)
∀s ∈ S, f ∈ F , l ∈ L, t ∈ T
In any DC, fresh and dry warehouse size limitations may arise; this is modeled by constraint (41).

IDs, f,l,t ≤ imax
i,l
, ∀s ∈ S, i ∈ F T , l ∈ L (41)
f ∈F : f t y f =i

5.1.10. Shelf life constraints


Shelf life constraints at FW. Based on the concept discussed above in subSection 4.2, the following two constraints are introduced into
the planning model to enforce the shelf-life indirectly. This constraint ensures that a product family will be transported to the distribution
centers before the end of its warehouse shelf-life.
  
IFf,t ≤ DV f,l,t1 , ∀ f ∈ F , t ∈ T : f t y f = F resh , t + ω ff w ≤ hl
 
t1 ∈ (t + 1 ).. t + ω f fw
(42)
l ∈ L

IFf,t ≤
 (DV f,l,t1 ) .t + ω f w − hl  + 
DV f,l,t2
  ω ff w f  
t1 ∈ t − ω ff w + 1 ..t t2 ∈ (t + 1 ).. t + ω ff w : t2 ≤ hl
(43)
l ∈ L l ∈ L
 
∀ f ∈ F , t ∈ T : f t y f = F resh , t + ω f fw
> hl

Shelf life constraints at DC: Similarly, two constraints are introduced that ensures that a product family will be transported to the retailers
before the end of its distribution center shelf-life.

IDs, f,l,t ≤  
d ps, f,l,t1
t 1∈(t +1 ).. t +ωdc (44)
f
 
∀s ∈ S, f ∈ F , l ∈ L, t ∈ T : f t y f = F resh , t + ωdc
f
≤ hl
  
(d ps, f,l,t1 ) . t + ωdc − hl + 
IDs, f,l,t ≤   ωdc f  
d ps, f,l,t2
f
t 1∈ t −ωdc +1 ..t t 2∈ (t +1 ).. t +ωdc :t 2≤hl (45)
f
  f

∀s ∈ S, f ∈ F , l ∈ L, t ∈ T : f t y f = Fresh , t + ωdc


f
> hl
12 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

5.1.11. Shipments consolidation


Shipments (Vi, l, t ) from the factory inventory to DCs l are consolidated into fresh and dry shipments with variables DVf, l, t according to
their refrigerated transportation requirements.

Vi,l,t = DV f,l,t , ∀i ∈ F T , l ∈ L, t ∈ T (46)
f ∈F : f t y f =i

5.1.12. Required number of trucks


The volume shipped from factory inventory to DC l needs to be loaded in trucks. The number of required trucks (TRi, l, t ) is calculated
in the following constraints. A shipment to a distribution center may require n trucks, only one of these n trucks is allowed to have less
than a truck load tlmax and the minimum amount of cargo it can transport is given by parameter tlmin .
Vi,l,t ≤ T Ri,l,t .t l max , ∀i ∈ F T , l ∈ L, t ∈ T (47)

 
Vi,l,t ≥ T Ri,l,t − 1 .t l max + t l min , ∀i ∈ F T , l ∈ L, t ∈ T (48)
Every shipment from a factory warehouse to a DC must be planned to arrive within the horizon. For that, Vi, l, t must be set to zero every
time t + τl > hl
Vi,l,t = 0, ∀i ∈ F T , l ∈ L, t ∈ (hl − τl + 1)..hl : τl > 0 (49)

5.1.13. Exports lots


Bounds for the number of lots to be exported for each family f on the horizon is given by constraint (50).

el min
f
≤ E f,t ≤ el max
f
, ∀f ∈ F (50)
t∈T

5.1.14. Objective function


The following variables aid the definition of the objectives for the aggregated planning stochastic model.
Total costs are composed of the sum of transport costs, setups costs and raw milk costs per scenario multiplied by their respective
probability. Raw milk costs per scenario are calculated for the cases of overstock and understock. In cases of understock are the costs
obtained from buying raw milk from a third party and for the case of overstock are production costs.
  
T COST = (T Ri,l,t .t ci,l ) + (Ym,t .sco) + ρs . (RSs,t .rsc + ROs,t .roc )
i ∈ FT m ∈ MP s∈S t∈T
(51)
l ∈ L t∈T
t ∈ T
The expected net benefit is obtained from income minus total costs. The incomes include from left to right, the export sales income, sales
income, and overstock incomes. Overstock incomes are produced from selling at a reduced price by the quantity produced in overstock for
each of families.
⎛ ⎞
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
   ⎜       ⎟
ENB = r e f .l s f .E f,t + ρs . ⎜ ⎜ (r f . d ps, f,l,t − SOs, f,l,t ) + OSs,l, f,t .r r f ⎟ − T COST
⎟ (52)
f ∈F s∈S ⎜ l ∈ L l∈L ⎟
⎜ ⎟
t∈T ⎜f ∈ F f ∈ F ⎟
⎝t ∈ T t ∈ T ⎠

5.2. Detailed planning model

Sets and indices


P Products, p ∈ P

First-age parameters (Scenario independent)


FPf, t Family Aggregated Planning Variable Values. Family f production on day t.
lsp Product p export lot size, expressed in metric tons.
id0p, l Product p initial inventory at any location l ∈ L∪ < {FW}.
Ef, t Family Aggregated Planning Variable Values. Family f daily export on day t measured in elf tons lots.
rp Revenue from reduced price selling of product p over stock (distressed sale of products).
rp Revenue from selling one ton of product p in any DC of the Supply Chain.
ref Revenue from exporting a batch of product p.

Second-Stage parameters
dpds, p, l, t1 Product p demand for distribution center l on day t under scenario s.

First-Stage decision variables (Scenario independent)


PDp, t Product p production on day t.
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 13

EDp, t Export level of product p on day t.


PSp, l, t Product p shipment quantity from Factory to DC l departing on day t.
IFDp, t Product p factory inventory on day t.
VDi, l, t Fresh and dry (i ∈ F T = { F resh , Dry }) quantity shipped to distribution center l on day t.

Second-Stage decision variables


IDDs, p, l, t Product p inventory at DC l under scenario s on day t.
SODs, p, l, t Product p stock out quantity at distribution center l on day t under scenario s.
OSDs, p, l, t Product p over stock quantity at distribution center l at day t under scenario s.

Integer variables
TRDi, l, t The number of trucks assigned for transporting i type family products from factory warehouse to DC l. Departing in period t and
arriving at l in period (t + τi ).

In the following constraints, for every product p, flyp accounts for product p family. Manufacturing of products from any family f is
constrained to the production level previously set in the Family Aggregated Model (FPf, t ).

F Pf,t = P D p,t , ∀ f ∈ F, t ∈ T (53)
p ∈ P: f = f l y p

Export levels for family f products (Ef, t ) defined in the Family Aggregated model define the export levels for every product in the Detailed
Planning model.

el f .E f,t = E D p,t .l s p , ∀ f ∈ F, t ∈ T (54)
p ∈ P: f = f l y p

Inventory balances in distribution centers. These restrictions are homologous to eq. (40) of the family aggregated model.
   
IDDs,p,l,t = id0 p,l + P S p,l,t1 − dpds,p,l,t1 + SODs,p,l,t1 − OSDs,p,l,t1
t 1∈T :(t 1+τi )≤t t 1∈T :t 1≤t t 1∈T :t 1≤t t 1∈T :t 1≤t (55)
∀s ∈ S, l ∈ L, p ∈ P, t ∈ T
Finished products inventory at the factory is determined by the fluctuations caused by product manufacturing, shipments, and exports.
    
IF D p,t = id0 p,F W + P D p,t1 − P Sl,p,t1 − E D p,t1 .l s p
t 1∈T :t 1≤t l∈L t 1∈T :t 1≤t
(56)
t1 ∈ T : t1 ≤ t
∀ p ∈ P, t ∈ T
Shelf-life considerations are included in the Detailed Planning model as well.
Factory shelf-life:

 
IF D p,t ≤   PS p,l,t1 , ∀ p ∈ P, t ∈ T : f t y f l y p = Fresh , t + ω ff lwy p ≤ hl
t1 ∈ (t + 1 ).. t + ω ff lwy (57)
p

l ∈ L

  
IF D p,t ≤
(PS p,l,t1 ) . t + ω f w − hl + P S p,l,t2
  ω fw
f ly p
f l y p  
t1 ∈ t1 − ω ff lwy + 1 ..t t2 ∈ (t + 1 ).. t + ω ff lwy : t2 ≤ hl
p p
(58)
l ∈ L l ∈ L
 
∀ p ∈ P, t ∈ T : f t y f l y p = F resh , t + ω ff lwy p > hl
Distribution Centers shelf-life:

IDDs,p,l,t ≤  
dpds,p,l,t1
t 1∈(t +1 ).. t +ωdc
f ly p
  (59)
∀s ∈ S, p ∈ P, l ∈ L, t ∈ T : f t y f l y p = Fresh , t + ωdc
f lyp
≤ hl


  
IDD ≤
(dpds,p,l,t1 ) . t + ωdc − hl +
s,p,l,t   ωdc
f ly p
f lyp   dpds,p,l,t2
t 1∈ t −ωdc
f ly
+1 ..t
p t2 ∈ (t + 1 ).. t + ωdc
f lyp
:
(60)
t2 ≤ hl
 
∀s ∈ S, p ∈ P, l ∈ L : t ∈ T : f t y f l y p = Fresh , t + ωdc
f lyp
> hl

Family type (Fresh or Dry) i shipments to distribution center l on day t from factory to Distribution Centers are consolidated into fresh
and dry shipments according to their refrigerated transportation requirements.

V Di,l,t = P Sl,p,t , ∀i ∈ F T , l ∈ L, t ∈ T (61)
p∈P: f t y p =i

The number of trucks required is calculated with the following equations, analogous to constraints (47) and (48) of the Family Aggregated
model.
V Di,l,t ≤ T RDi,l,t .t l max , ∀i ∈ F T , l ∈ L, t ∈ T (62)
14 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

 
V Di,l,t ≥ T RDi,l,t − 1 .t l max + t l min , ∀i ∈ F T , l ∈ L, t ∈ T (63)

4.2.1. Objective function


The detailed planning model maximizes the expected net benefit (DENB), and is obtained as follows.
⎛ ⎞
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
 ⎜       ⎟
⎟ −  (T Ri,l,t .t ci,l )
(re p .l s p .E D p,t ) + ρs . ⎜ ⎜ (r f . dpds,p,l,t − SOs,p,l,t ) + OSs,l,p,t .r r p ⎟ (64)
p∈P s∈S ⎜ l ∈ L l∈L ⎟ i ∈ FT
⎜ ⎟
t∈T ⎜p ∈ P p ∈ P ⎟ l ∈ L
⎝t ∈ T t ∈ T ⎠ t ∈ T

6. Solution strategy

In this section, a novel solution strategy based on a scenario reduction algorithm and a hierarchical-based decomposition of the plan-
ning decisions is presented. SubSection 6.1 introduces a scenario reduction methodology based in a well-known clustering technique.
Subsection 0 introduces the solution method developed to obtain a stochastic production and distribution plan for the dairy supply chain
given the reduced scenario set. While subSection 6.3 presented the measures to compare and evaluate the stochastic solution against
deterministic planning.

6.1. Scenario reduction algorithm

For any real-world application of the planning methodology introduced in this work, the number of scenarios grows significantly,
making the resulting mathematical models intractable. For example, the case study inspiring this work has four families of products, and
for each of them, three possible demand scenarios are considered. Also, three additional scenarios represent raw milk variability. In this
setting, |S|=35 =243, that is there are 243 different scenarios which make the Family Aggregated and Detailed Planning models extremely
difficult to solve. Other problem settings with bigger scenario sets present further challenges.
In the literature, to specifically deal with the number of scenarios, researchers have followed two different directions, Scenario Reduc-
tion Römisch (2009) and Monte Carlo Sampling-Based Methods Homem-de-Mello and Bayraksan (2014) also known as Sample Average
Approximation Methods (SAA). In this work, we follow the Scenario Reduction direction. The reason behind this design choice is straight-
forward, our modelś computational complexity prevents its use under a simulation-based optimization framework. This can be concluded
after examining Table 12, even for a small subset of scenarios, the solution time is considerable, and, SAA requires statistical experimen-
tation over Samples, which will impose a greater computation burden to obtain a good solution.
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 15

For scenario reduction we introduce as simple, yet fast and re- 6.2. Solution method
liable Scenario Reduction method based on unsupervised learn-
ing techniques. It is based on defining a set of scenario vec- The approximate method to obtain the stochastic production
tors; each scenario vector holds the demand and supply val- and distribution plan is shown in Fig. 2. First, stochastic scenar-
ues that define it. For example, if there were 4 product fam- ios are reduced in order to obtain a tractable scenario set. Then,
ilies (UHT,PM,Yogurt,Cheese), any scenario i, takes the form i= there is one solving phase for the Family Aggregated model and
[UHTi ,PMi ,Yogurti ,Cheesei ,RawMilki ] = [640 0,60 0,280 0,30 0,862.5] one for the Detailed Planning model. These phases applied sequen-
with p(si ) = (0.4∗0.25∗0.4∗0.3∗0.5). The values for this possible tially allow obtaining a production and distribution plan that con-
scenario were taken from Table 4, and it is one of the 243 pos- siders the manufacture of each specific product. There is one single
sible scenarios in the case study analyzed in this work. The idea goal, which is maximize expected benefits, which are a result of
behind the algorithm is to generate a data set W which is pop- the extent of the satisfied demand, distressed sales of products and
ulated with scenarios having an occurrence greater than epsilon exports, minimize setups, and, raw milk understock and overstock.
after N (big number, e.g. 10,0 0 0 trials). The smallest the epsilon The lasts three components of this objective function (setups and
the larger W becomes. E.g. if epsilon =1 and after 10,0 0 0 trials the raw milk costs) are only considered for the aggregated model since
expected occurrence of any scenario is less than 1, then this sce- the Detailed Planning model keeps the definition of campaigns ob-
nario is discarded. Every scenario vector considered populates W tained when solving the aggregated model. We emphasize that this
with f(si ) repetitions of itself, where f (si ) = p(si ) · N. This require- sequential solution procedure cannot obtain the global optimal so-
ment arises because the selected unsupervised learning technique lution, but, ensures an optimized hierarchical planning-based solu-
does not consider weights as a component of the input data set. tion, and is aligned with forecasting procedures recommendations
K-means clustering is applied to W. And, the resulting cluster cen- as exposed in Section 2.
troids are the scenarios considered in the mathematical formula-
tion of the planning model. The algorithm is summarized as fol- 6.3. The value of the stochastic solution
lows, table 2.
In this work, the mean value problem (MVP), is obtained by
defining a unique scenario vector s∗ which arises from comput-
Table 2
Scenario Reduction Algorithm. ing the expected value of the scenario set. This scenario s∗ has
a demand value for each of the product families and raw milk
Require: Vector set of Scenarios S, where si = [si1 , si2 , .., sim ] ∈ S and m is the
availability. EMVP is the expected value of using the MVP solu-
number of real components describing scenario si . For each scenario si , p(si )
is its probability of occurrence. tion (obtained by solving the planning model using scenario s∗ )
Require: f (si ) expected frequency of scenario si after N experiments, where against each of the scenarios. EMVP is computed considering opti-
f (si ) = p(si ) · N, e.g. N = 10 0 0 0 trials. mal second stage reactions for every scenario in set S and taking
K : Target number of scenarios (K > 1) the appropriate expectation considering the occurrence probabil-
ε : minimum allowed expected frequency (integer value).
W is a dataset with m fields, corresponding to the m number of components
ity of each scenario. If SS is the optimal objective function value
in a scenario vector. of the stochastic program, then the value of the stochastic solution
1 for(s = 1; s++; s<=card(S)){ VSS is VSS = SS-EMVP, for maximization problems, as in this case
if (f(s) >= ε ) { SS ≥ EMVP. This measures how useful it is to solve the stochastic
for(i = 1; i++; i<=f(s)){
problem instead of just solving the MVP.
W.add(s) // add scenario s, f(s) times
}
} 7. Case study analysis
}
2 Normalize W into W  , as follows: To validate our proposal, we used a case study based on a dairy
If Wx is a field of W , and xmin , xmax are its minimum and maximum values,
supply chain in Argentina. The supply chain is composed of one
respectively, for any element in Wx . Then, any value xs of Wx is normalized
to xs  using the following formula. manufacturing complex and 9 distribution centers; it serves a na-
xs  = (xs − x̄ )/Sx , where x̄ is the mean value for Wx and Sx its standard tionwide market and has some export product lines. The supply
deviation (∗ ) – This is called z-score normalization in statistics. chain produces and distributes a wide variety of dairy products. In
The next step is to apply the K-means clustering algorithm Aggarwal and
this study, the set of considered products was selected by using a
Reddy (2013). This will result in a set of K clusters Ck each with a centroid
Ck  , where, Ck  is a vector of m components (the number of components Pareto rule approach, including 20% of the products that represent
describing a scenario). more than 80% of the demand. The remainder of the demand was
3 De-normalize Ck  according to (∗ ). proportionally added to this set of critical products. There are 7
For each centroid, calculate its probability as p(Ck ) = card (Ck )/card (W ) UHT Milk products, 9 Powdered Milk products, 11 Yogurt products,
4 Return reduced scenario set sr = {Ck | k in 1..K } with probabilities p(Ck ).
and 15 Cheese products, for a total of 42 dairy products. This case
was solved using CPLEX 12.8 on a four-core Intel I7 machine with
32GB of RAM.
The target number of scenarios can be calculated using the av- Raw milk arrives on a daily basis and is distributed to four
erage within cluster distance (AWCD) metric which is calculated manufacturing plants that produce UHT Milk, Powdered Milk, Yo-
averaging the distance between the centroid and all examples of gurt, and Cheese. Raw milk is treated in a continuous heating,
a cluster. This distance is the Euclidian distance between the cen- cooling, and filtering facility before its consumption for product
troid Ck  and the elements (Scenarios) of Ck . The minimal value manufacturing. The plants operate through the definition of pro-
of AWCD is zero and it takes place when we choose a K equal duction campaigns. Campaign types and plant production capaci-
to the size of the original Scenario set. We leave the decision ties are shown in Table 3.
maker with the burden of deciding which value of K provides an Each plant has a factory warehouse where final products are
acceptable scenario distance, but for reference, we have included stored. Additionally, there are work-in-progress inventories for
a numerical analysis on the effect of K, on both the quality of powdered milk quality testing and cheese ripening. Powdered milk
the solution and the AWCD. In any case some authors have dis- is quality tested, and laboratory results are available one day af-
cussed the subject of the selection of K in clustering algorithms ter samples are taken. The deterioration coefficient is 0.1 or 10%
Pham et al. (2005). for both Powdered Milk production lines. Cheese manufacturing
16 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

Fig. 2. Solution strategy schematization.

Table 3
Production structure.

Manufacturing Plant Campaign Type Manufacturing Capacity Product yield per ton of raw milk

Powdered Milk Length-based Campaign Two production lines which can 0.12 t
operate for up to 15 days without
major cleaning tasks. One line has a
daily production output rate of 102 t
of powdered milk and the other of 18
t
UHT Milk Shift-based Campaign Up to three daily shifts with up to 0.975 t
110 t of UHT per shift
Yogurt Shift-Based Campaign Up to three daily shifts with up to 0.9 t
150 t of yogurt per shift
Cheese Shift-based Campaign Up to three daily shifts with up to 0.11 t
16.66 t of cheese per shift

has a ripening phase. In the manufacturing complex, only soft-type Yogurt: 5 MU; Cheese: 12 MU;the margins for a ton of exported
cheese is produced, and it has an average time of the total process, powdered milk is 5 MU. Sales revenues for products produced in
including ripening, of 4 days. overstock is 0 for products of Yogurt and Cheese families, 1 MU
The product is shipped to DCs by truck transportation. There for UHT and 3.75 MU for Powdered Milk. The purchase price of
are two types of shipments: dry (not refrigerated) for products of raw milk from a third supplier is 2 MU. For every setup task, the
the UHT and Powdered Milk families; and fresh (refrigerated), for cost is 20 MU.
products of the Yogurt and Cheese families, which are made using
refrigerated trucks. A shipment may require one or more trucks
7.1. The case
to be dispatched to a DC. If a shipment requires n trucks to be
dispatched, only one of them can be sent with less than a truck-
In this study, historical data of dairy products demand was an-
load. The minimum load capacity needed to dispatch a truck is 4
alyzed. As a result of this analysis, three DC demand scenarios
t, and the maximum capacity is 8 t Powdered Milk products can
were built as the basis for the definition of a case discussed in
also be exported in lots of 25 t, and only a maximum of 120 lots
this work. The case consists of: a demand forecast for a 30-day
can be exported. In such a case, an amount of it is set apart; and
horizon; initial inventories in both FWs and DCs; data from the
an exportation subprocess begins, which is not explicitly modeled
previous planning period such as work-in-progress inventories and
in this work.
their status (powdered milk to be approved, cheese to be ripened),
Each of the nine distribution centers holds separate fresh and
and products being shipped. Three supply-demand scenarios (S1,
dry inventories of all 42 dairy products considered in this supply
S2, and S3), represent the discrete probability distribution func-
chain. Fig. 3 shows a summary of the distribution structure. In the
tions for each product family and raw milk supply. In Table 4, the
figure, each label shows in ascending order the name of the DC,
scenarios (si ), measured in t (metric tons), and its occurrence prob-
maximum fresh and dry inventories, measured in tons and fleet
abilities are given.
cost/truck for fresh and dry cargo in monetary units MU. Also, lead
The joint probability distribution of these individual sets of sce-
times measured in days are presented for each DC. Costs and in-
narios generates a total of 243 (35 ) possible scenarios (each sce-
comes are measured in MU, for confidentiality issues.
nario gives a value for every product and supply), and their occur-
Figure 4, Fig. 4a, Fig. 4b, Fig. 4c
rence probability. This number of scenarios is large enough to in-
Shelf-life for Yogurt products is 5 days at FW and 7 days at DC;
crease the resolution times of the proposed models to intractable
for Cheese, it is 10 and 14 days respectively. Sales margins after
levels. Therefore, the scenario reduction algorithm presented pre-
production costs are UHT: 3.425 MU; Powdered Milk: 12.5 MU;
viously is applied with K = 9 and ε = 0.0 0 01. With those pa-
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 17

Fig. 3. Dairy Supply Chain - Distribution Side.

Table 4
Product families and Supply scenarios.

UHT PM Yogurt Cheese Raw Milk

i si p(si ) si p(si ) si p(si ) si p(si ) si p(si )

1 6400 0.4 600 0.25 2800 0.4 300 0.3 862.5 0.5
2 8000 0.35 1000 0.35 4000 0.3 500 0.4 1725 0.15
3 12,000 0.25 1200 0.4 8000 0.3 600 0.3 3450 0.35

Table 5
Reduced Scenario Set. Demand values are given in t (metric tons).

Scenario UHT PM Yogurt Cheese Raw Milk Cluster Size Prob.

1 8352 981 8000 300 1062 584 0.0584


2 8365 979 8000 470 3450 1050 0.1050
3 7147 1107 3314 470 3450 1377 0.1378
4 8361 600 3313 470 1061 1136 0.1136
5 8362 600 3315 470 3450 613 0.0613
6 7146 1107 3315 543 1062 1791 0.1792
7 12,000 1106 3314 470 1899 1312 0.1313
8 7146 1107 3316 300 1062 768 0.0768
9 8359 980 8000 543 1062 1365 0.1366

rameters W, the dataset to cluster has 9996 rows (each row is a cental improvement on the economic returns obtained by solving
scenario). After clustering, we obtain a reduced set of scenarios the stochastic problem is 21.1%.
vectors that optimally represent the total of possible scenarios as Moreover, the “mean value” planning results are less stable than
shown in Table 5. These scenarios are the clusters of a huge 5- the stochastic planning results as can be seen in Table 7. The SP
dimensional vector set W (as defined in Section 6.1) arising from solution requires only 5 production campaigns while the EMVP so-
the 243-sized original scenario set. Each dimension represents the lution requires 8. This means that the SP planning approach mini-
value of either demand family or available supply. mizes setups even further.
The results in Table 8 show how the mean value plan max-
imizes production and raw milk utilization, but this maximiza-
7.2. Case planning results
tion negatively impacts performance as it cannot address variations
arising from the different possible scenarios. This situation reflects
The planning methodology of this work is applied to the case
the fact that an inadequate production level will increment raw
study in this section and results are shown. For a better appreci-
milk purchases at premium costs. Similarly, there is a fine balance
ation of the features of this work, every planning result from this
between distressed sales of products and allowing some stockout
proposal (SP – Stochastic Problem) is compared to the values ob-
days.
tained from the EMVP, in which the mean value problem faces the
In Table 9, the evolution of the decomposition strategy for the
occurrence of the stochastic scenarios.
case is shown. In the table, using the stochastic solution allowed
In Table 6, the main production and distribution metrics for the
to reducing the number of setups from 4 to 2, which is a 50% re-
whole planning horizon (30 days) are given. The stochastic pro-
duction.
gramming outperforms planning for the mean in most every met-
The computational effort required to solve the stochastic plan-
ric shown when they are economically quantified. The value of
ning models instead of the mean value planning models is signifi-
the stochastic solution is 34,670.1 MU which means that the per-
18 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

Table 6
Aggregate results for the case study in t (metric tons) and MU (Monetary units).

SP EMVP % deviation from EMVP

Sales [t] 12,815 11,149 +14.9


Lost Sales [t] 1715.2 3381 −49.3
Distressed Sales of Products [t] 3694.1 1410.2 +162.0
Raw Material Losses [t] 22,951 20,334 +12.9
Raw Material Purchase [t] 1078.4 4471.9 −75.9
Exports [t] 975 2500 −61.0
Production [t] 16,561 14,460 +14.5
Product shipped to DC [t] 15,556 11,607 +34.0
Sales Income [MU] 62,241 53,541 +16.2
Distressed Sales of Products [MU] 3516 752 +367.5
Raw Material Losses Cost [MU] 22,951 20,334 +12.9
Raw Material Purchase Cost [MU] 2156.90 8943.7 −75.9
Exports Income [MU] 4875 12,500 −61.0
Cost for shipped to DC [MU] 10,814 8798.5 +22.9
Setups Cost [MU] 40 80 −50.0
Expected Net Benefits [MU] 34,670.1 28,636.8 +21.1

Table 7
SP and EMVP production campaigns.

Plant N° of Campaigns Campaigns Duration N° Shifts Daily Campaign Production

SP Results UHT 1 30 Days 3 Shifts 330 [t]


PM-1 – – – –
PM-2 2 19 Days – 102 [t] in all Campaigns
Yogurt 1 30 Days 1 Shift 149 [t]
Cheese 1 30 Days 1 Shift 15.4 [t]
EMVP Results UHT 1 30 Days 2 Shifts 220 [t]
PM-1 3 28 Days – 18 [t] in all Campaigns
PM-2 3 28 Days – 102 [t] in all Campaigns
Yogurt 1 30 Days 1 Shift 150 [t]
Cheese – – – –

Table 8
Detailed manufacturing and distribution planning results.

SP EMVP

Production [t] UHT 9881.6 6600


Powdered Milk 1938 3360
Yogurt 4465.3 4499.8
Cheese 276.4 –
N° of PM Export Lots 39 100
Processed Raw Milk [t] 33,759 397,769
Required trucks Fresh 598 554
Dry 1348 910
UHT 9.7 20.2
DCs Expected Stockout days Powdered Milk 5.5 16.1
Yogurt 10 13.1
Cheese 6.7 29.7
UHT 3.5 7.9
DCs Expected Overstock days Powdered Milk 3.8 4.8
Yogurt 6.4 18.1
Cheese 2.3 1.2
Expected number of days with raw milk Overstock 10.5 12.1
Expected number of days with raw milk bought at premium price 2.2 9
Expected Sales (MU) 62,241 53,541
Expected Raw Milk Costs (MU) 25,107 29,277
Expected Net Benefits (MU) 34,670.1 28,636.8

Table 9
Aggregated and Detailed planning phases for SP and EMVP.

SP EMVP

Family Aggregated Phase Max Expected Net Benefits Max Expected Net Benefits
Expected Net Benefit 34,670.7 28,637
Number of Setups 2 4
Solution Time (s) 103 43.8
Integer Gap 0.01 0.01
Detailed Planning Phase Max Expected Net Benefits Max Expected Net Benefits
Expected Net Benefits 34,607.7 28,342.4
Solution Time (s) 97 26.9
Integer Gap (%) 0.1 0.1
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 19

Table 10
ANOVA results for the Family Aggregated Model.

Source Sum SQ. D.F. Mean SQ. F Prob>F

DCs 8,413,385 2 4,206,692.5 374.91 8.55E-39


Scenarios 3,576,527 2 1,788,263.5 159.37 3.60E-27
DCs∗ Scenarios 3,637,650.4 4 909,412.6 81.05 6.65E-26
Error 807,887.7 72 11,220.7
Total 16,435,450 80

Table 11
ANOVA results for the Detailed Planning Model.

Source Sum SQ. D.F. Mean SQ. F Prob>F

Products 39,426.1 2 19,713 2.37 0.1032


DC 18,647.6 2 9323.8 1.12 0.3335
Scenarios 39,637.4 2 19,818.7 2.38 0.102
Products∗ DC 7607.2 4 1901.8 0.23 0.9212
Products∗ Scenarios 15,517.8 4 3879.5 0.47 0.7601
DC∗ Scenarios 9502.9 4 2375.7 0.29 0.8861
Products∗ DC∗ Scenarios 3644.5 8 455.6 0.05 0.9999
Error 449,224.4 54 8319
Total 583,207.9 80

The Family Aggregated model was solved in 1371s. (seconds) in


the worst case, and the Detailed Planning model took 521.9 s. to
be solved in the worst case. Every computational experiment was
performed using CPLEX 12.8 on a four-core Intel I7 machine with
32GB of RAM. Each instance was solved three times as CPLEX was
set with the non-default opportunistic parallel MIP setting.
The nomenclature used to indicate each instance in Table 12 is
the following, from left to right, products per family, distribution
centers and the number of scenarios. In this way, the instance
named “P3DC4S8” corresponds to the instance that considers 3
products per family, 4 distribution centers, and 8 scenarios.
Analysis of variance (ANOVA) is used to study the effects of
the different factors on computational performance. Following,
Table 10 and
Table 11 show the result of applying ANOVA to the solution
times of the family aggregated model and to the solution times
of the Detailed Planning model respectively.
From results obtained from previous tables, it is concluded that:
- Considering a cutoff value of 0.05, in the solution stage cor-
Fig. 4a. - Fig.4c the main effects of each factor are plotted, showing the average
results for each significant factor for both the aggregated and detailed planning
responding to the Family Aggregated model, there is evidence sup-
stages. (a) Main effects plot for time (Scenarios). (b) Main effects plot for time porting that the number of distribution centers, the number of sce-
(DCs). (c) Main effects plot for time (Products). narios, and their interaction have a significant effect on compu-
tational time. The factor product per family was not evaluated in
Table 10 because the family aggregated model does not consider
cant, but general solution times are inside acceptable bounds. The products.
solution time for the mean value plan slightly above 1 min., while, Considering a cutoff value of 0.1 in the solution stage of the
obtaining the stochastic solution takes more than 3 min. This ef- Detailed Planning model, one can observe that the number of sce-
fort is valuable as it has a strong impact (21.1%) on improving the narios and the number of products per family, individually have a
expected benefits of the supply chain. significant impact on the solution time. Fig. 4a-4c, show the main
effects plot for each factor of the experiment.

8. Computational time analysis


8. Evaluating the quality of the solutions obtained through the
Numerical experiments are performed to assess the computa- solution strategy
tional effort required to apply the proposed formulation. The main
idea is to investigate the factors of the model playing a significant The computing time required to solve large instances of the
role in solution time. To answer this question, a full factorial de- proposed dairy SC production and distribution plan prevents the
sign experiment with three factors and three levels is designed. use of large scenario sets without previously applying the reduc-
The factors defined in this experiment and their levels are number tion strategy. Nevertheless, for the smaller instance P3DC4, of the
of products per family (3, 6 and 10), number of distribution centers numerical study presented in Section 8, we could obtain almost
(4, 8 and 12) and the number of scenarios (8, 16 and 24). A total of the optimal solution after 10 h of computation for a new in-
27 (33 ) instances were considered and solved, solution times, types stance called P3DC4243 (with the full scenario set). In this in-
and number of variables and the number of constraints are given stance, the expected net benefits objective function gave a value
in Table 12. of 45,022,1621 with an integrality gap of just 0,03%. We use this
20 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

Table 12
Instances solved for the computational time study.

Family Aggregated Detailed Plan


Instance
Float Variables Integer Variables Constraints Solved Time (s) Float Variables Integer Variables Constraints Solved Time (s)

P3DC4S8 15,933 960 15,258 9.7 36,975 360 30,098 3.3


P3DC4S16 30,117 960 25,362 35.5 71,543 360 58,522 10.9
P3DC4S24 44,301 960 35,466 60.8 106,111 360 86,946 19.0
P6DC4S8 15,933 960 15,258 9.7 73,695 720 59,222 9.2
P6DC4S16 30,117 960 25,362 35.4 142,823 720 116,062 31.2
P6DC4S24 44,301 960 35,466 60.4 211,951 720 172,902 68.4
P10DC4S8 15,933 960 15,258 9.7 122,655 1200 98,054 20.0
P10DC4S16 30,117 960 25,362 35.5 237,863 1200 192,782 76.6
P10DC4S24 44,301 960 35,466 60.5 353,071 1200 287,510 146.5
P3DC8S8 30,093 1200 26,546 38.0 73,215 360 59,234 8.4
P3DC8S16 57,717 1200 46,250 143.7 142,343 360 116,074 27.8
P3DC8S24 85,341 1200 65,954 356.8 211,471 360 172,914 51.5
P6DC8S8 30,093 1200 26,546 38.3 145,935 720 116,774 22.1
P6DC8S16 57,717 1200 46,250 141.8 284,183 720 230,446 85.8
P6DC8S24 85,341 1200 65,954 364.1 422,431 720 344,118 161.4
P10DC8S8 30,093 1200 26,546 39.7 242,895 1200 193,494 71.3
P10DC8S16 57,717 1200 46,250 147.7 473,303 1200 382,942 161.2
P10DC8S24 85,341 1200 65,954 350.6 703,711 1200 572,390 357.7
P3DC12S8 44,253 1440 37,836 74.0 109,455 360 88,370 11.2
P3DC12S16 85,317 1440 67,140 942.0 213,143 360 173,626 31.6
P3DC12S24 126,381 1440 96,444 1337.3 316,831 360 258,882 89.5
P6DC12S8 44,253 1440 37,836 73.0 218,175 720 174,326 33.7
P6DC12S16 85,317 1440 67,140 942.1 425,543 720 344,830 92.5
P6DC12S24 126,381 1440 96,444 1331.7 632,911 720 515,334 271.6
P10DC12S8 44,253 1440 37,836 80.0 363,135 1200 288,934 60.2
P10DC12S16 85,317 1440 67,140 1371.3 708,743 1200 573,102 275.4
P10DC12S24 126,381 1440 96,444 865.7 1,054,351 1200 857,270 521.9

Table 13
Optimality gap value as a function of K.

K AWCD (Average within cluster distance) Reduced scenario solution vs. real-World Scenario Set Opt solution - full scenario set optimality gap (%)

2 4.079 15,027.265 45,022.162 199.603


4 2.61 31,777.031 45,022.162 41.681
6 1.634 43,990.338 45,022.162 2.346
8 1.341 43,948.921 45,022.162 2.442
10 1.167 43,949.446 45,022.162 2.441
12 1.044 43,949.205 45,022.162 2.441
14 0.912 44,431.800 45,022.162 1.329
16 0.788 44,439.311 45,022.162 1.312
18 0.723 44,690.291 45,022.162 0.743
20 0.652 44,196.041 45,022.162 1.869
22 0.57 44,213.692 45,022.162 1.829
24 0.543 44,317.407 45,022.162 1.590
36 0.345 44,637.595 45,022.162 0.862
44 0.285 44,744.489 45,022.162 0.621

optimal solution as a reference to evaluate the quality of our ap- Additionally, we solve (SP1) for P3DC4SK, and this time S=SK,
proach, as follows. is the set of k reduced scenarios, and obtain Fk = (xk , yk ).
Given that any 2-stage stochastic problem with a finite set of Then, SP1 is solved for P3DC4243, where the constraint x =
scenarios S can be represented as SP1, we use SP1 structure to eval- xk , is added. We emphazise that SP1’ has S = S243 as its sce-
uate the quality of our approach. If, we call (SP1), the following nario set. Therefore, we evaluate the performance of the first-
S stage reduced scenario set solution against the full set of scenar-
optimization problem. min cT x + ρs qTs ys #(SP1 ) O (F ∗ )−V RSSk
s=1 ios with the following optimality gap formula, V RSS_k
We em-

phasize that O(F ) is greater or equal than VRSk , as it was calcu-
Ax = b
lated as the best possible answer for the whole set of scenarios.
S
Ts x + Ws · ys = hs V RSSk = min cT x + ρs qTs ys #(SP1 )
s=1

Ax = b
x ∈ Rn1 −p1 × Z+p1
x = x∗k
ys ∈ Rn2 −p2 ; s ∈ S
Ts x + Ws · ys = hs

S
And, F∗ =(x∗ ,y∗ ) with value O(F∗ ) =c T x + ρs qTs ys , the solu-
x ∈ Rn1 −p1 × Z+p1
s=1
tion of P3DC4243, which is an instance without scenario reduction,
with S = S243, where S243 is the set of 243 scenarios. ys ∈ Rn2 −p2 ; s ∈ S
A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966 21

In Table 13, we evaluate the decrease in the value of the optimality milk storage) and distribution center inventories. Distribution re-
gap as K grows. The table shows us that even for a small value of source considerations are made with an emphasis on the use of re-
k, the obtained solution can cope with the uncertainty entailed by frigerated and non-refrigerated trucks and lot-sizes consolidation.
the full scenario set with acceptable performance. All the previous features facilitate the conversion of the resulting
plan into detailed operations.
9. Conclusions To apply the results presented in this work in a dairy supply
chain, special attention must be paid to the required data, which
Planning the supply chain in dairy industries faces the chal- is provided by the operation planners, as well as by advanced man-
lenge of highly varying product demand and raw material availabil- agement systems such as ERP systems, digital twins, among others.
ity. Managing this variability in a systematic planning methodology The model must be integrated with existing systems and provide
may yield significant benefits as compared to traditional determin- it with the required data, both resources and personnel, as well as
istic planning. In this work, we presented a hierarchical stochas- the uncertain data set and its representation through scenarios.
tic approach for integrated production and distribution planning For future work on Dairy SCs optimization, one possible re-
that was successfully applied to a real problem case. However, the search direction is detailed manufacturing scheduling. Coordinat-
model is complicated to solve and therefore, the proposed solu- ing detailed scheduling with planning targets is a delicate pro-
tion strategy has two dimensions. On the one hand, it requires a cess which requires both specific information systems and robust
scenario reduction process, and on the other, a decomposition ap- scheduling models.
proach based on the aggregation of products in families. Scenario
reduction is based on well-known clustering techniques that en- Declaration of competing interests
sure that the cluster based reduced set of new scenarios is the
most representative of the vast original set of scenarios. With re- The authors declare that they have no known competing finan-
gards to the decomposition approach, it is widely accepted that cial interests or personal relationships that could have appeared to
supply chain forecasting requires aggregation, in this work extend influence the work reported in this paper.
this good practice to the planning methodology.
We compare the deterministic model with the two-stage CRediT authorship contribution statement
stochastic programming model version by the measures of ex-
pected mean value problem (EMVP) and value of the stochastic so- Armando Guarnaschelli: Software, Validation, Methodology, In-
lution (VSS). The computational study showed that the VSS might vestigation, Writing - review & editing. Héctor Enrique Salomone:
rise to 21.1% above the deterministic solution. This indicates the Software, Validation, Methodology, Investigation, Writing - review
importance of considering uncertainty in raw milk supply and de- & editing. Carlos A. Méndez: Investigation, Project administration,
mand for dairy products. Also, larger instances are tested to study Funding acquisition, Supervision.
the scalability of the solution approach. We observed that the in-
stances for Family Aggregated Model become much harder to solve Acknowledgments
with a larger number of distribution centers. These same instances,
in the Detailed Planning phase became harder to solve when the Financial support for this work by Consejo Nacional de Investi-
number of scenarios, number of products and number of distribu- gaciones Científicas y Técnicas (CONICET), Argentina and by Ponti-
tion centers increased. ficia Universidad Católica de Valparaíso, Chile.
The decisions taken before the realization of scenarios, such
as campaign policies, export levels, cargo loads and trucks have Supplementary materials
proved to be determinant of total expected net benefits. Another
important aspect is the fluctuating availability of raw milk and Supplementary material associated with this article can be
how it impacts second-stage decisions, those taken after scenario found, in the online version, at doi:10.1016/j.compchemeng.2020.
realization. Raw milk losses (due to perishability) and purchases at 106966.
premium prices (due to unavailability) have a significant impact on
supply chain performance. For the case study, the expected quan- References
tity of raw milk purchased at premium prices in the solution of
the "mean value" problem almost doubles its counterpart in solv- Aggarwal, C.C., Reddy, C.K., 2013. Data Clustering: Algorithms and Applications.
ing the stochastic problem. This emphasizes the need of explicitly Chapman & Hall/CRC 1st ed..
Akkerman, R., Farahani, P., Grunow, M., 2010. Quality, safety and sustainability in
considering the variability of raw material supply in planning for- food distribution: a review of quantitative operations management approaches
mulations. and challenges. OR Spectr 32, 863–904. doi:10.10 07/s0 0291- 010- 0223- 2.
Besides demand and supply variability considerations, this work Amorim, P., Günther, H.-.O., Almada-Lobo, B., 2012. Multi-objective integrated pro-
duction and distribution planning of perishable products. Int. J. Prod. Econ. 138,
innovates on how it effectively addresses the delicate balance of 89–101. doi:10.1016/j.ijpe.2012.03.005.
supply, production and distribution planning functions. Raw milk Banaszewska, A., Cruijssen, F., van der Vorst, J.G.A.J., Claassen, G.D.H., Kampman, J.L.,
has very narrow storage flexibility since it perishes in a short pe- 2013. A comprehensive dairy valorization model. J. Dairy Sci. 96, 761–779.
doi:10.3168/jds.2012-5641.
riod. This flexibility is determined by the available inventory capac- Behzadi, G., O’Sullivan, M.J., Olsen, T.L., Scrimgeour, F., Zhang, A., 2017. Robust and
ity at both factory and DCs inventories and also by production con- resilient strategies for managing supply disruptions in an agribusiness supply
straints. In the traditional planning approach, raw milk assignment chain. Int. J. Prod. Econ. 191, 207–220. doi:10.1016/j.ijpe.2017.06.018.
Beraldi, P., Bruni, M.E., 2013. A clustering approach for scenario tree reduction: an
is decided considering demand and production capacity, without
application to a stochastic programming portfolio optimization problem. Top 22,
considering projected demand and inventories on every layer, dis- 1–16. doi:10.1007/s11750- 013- 0305- 9.
tribution resources, lead times, and constraints. In this work, we Bilgen, B., Çelebi, Y., 2013. Integrated production scheduling and distribution plan-
ning in dairy supply chain by hybrid modelling. Ann. Oper. Res. 211, 55–82.
include all these additional considerations, yielding a better plan-
doi:10.1007/s10479- 013- 1415- 3.
ning solution. Bilgen, B., Dogan, K., 2015. Multistage Production Planning in the Dairy Industry:
The methodology proposed achieves an executable monthly a Mixed-Integer Programming Approach. Ind. Eng. Chem. Res. 54, 11709–11719.
production and distribution plan. It can define campaign policies doi:10.1021/acs.iecr.5b02247.
Borodin, V., Bourtembourg, J., Hnaien, F., Labadie, N., 2016. Handling uncertainty in
of two types: shift-based and length-based. Perishability issues are agricultural supply chain management: a state of the art. Eur. J. Oper. Res. 254,
addressed both at factory inventories (finished products and raw 348–359. doi:10.1016/j.ejor.2016.03.057.
22 A. Guarnaschelli, H.E. Salomone and C.A. Méndez / Computers and Chemical Engineering 140 (2020) 106966

Brunaud, B., Laínez-Aguirre, J.M., Pinto, J.M., Grossmann, I.E., 2019. Inventory poli- Römisch, W., (2009). Scenario reduction techniques in stochastic programming bt
cies and safety stock optimization for supply chain planning. AIChE J 65, 99–112. - stochastic algorithms: foundations and applications, in: Watanabe, O., Zeug-
doi:10.1002/aic.16421. mann, T. (Eds.), . Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 1–14.
Chen, L., Lee, H.L., 2017. In: Ha, A.Y., Tang, C.S (Eds.). Springer International Publish- Ross, D.F., 2004. Distribution Planning and Control : Managing in the Era of Supply
ing, Cham, pp. 3–25. doi:10.1007/978- 3- 319- 32441- 8_1. Chain Management. Springer, US.
de Kok, T.G., Fransoo, J.C., Kok, T.G.De, 2003. Planning Supply Chain Operations: def- Salema, M.I.G., Barbosa-Povoa, A.P., Novais, A.Q., 2007. An optimization model for
inition and Comparison of Planning Concepts. in: Handbooks in Operations Re- the design of a capacitated multi-product reverse logistics network with uncer-
search and Management Science 597–675. doi:10.1016/S0927-0507(03)11012-2. tainty. Eur. J. Oper. Res. 179, 1063–1077. doi:10.1016/j.ejor.2005.05.032.
Ekşioğlu, S., Jin, M., 2006. Cross-Facility Production and Transportation Planning Santoso, T., Ahmed, S., Goetschalckx, M., Shapiro, A., 2005. A stochastic program-
Problem with Perishable Inventory. In: Gavrilova, M., Gervasi, O., Kumar, V., ming approach for supply chain network design under uncertainty. Eur. J. Oper.
Tan, C.J.K., Taniar, D, Laganá, A., Mun, Y., Choo, H (Eds.), Computational Science Res. 167, 96–115. doi:10.1016/j.ejor.2004.01.046.
and Its Applications - ICCSA 2006 SE - 75, Lecture Notes in Computer Science. Schuh, G., Prote, J.P., Luckert, M., Hünnekes, P., Schmidhuber, M., 2019. Effects of the
Springer, Berlin Heidelberg, pp. 708–717. doi:10.1007/11751595_75. update frequency of production plans on the logistical performance of produc-
Garcia-Herreros, P., Wassick, J.M., Grossmann, I.E., 2014. Design of resilient supply tion planning and control. Procedia CIRP 79, 421–426. doi:10.1016/j.procir.2019.
chains with risk of facility disruptions. Ind. Eng. Chem. Res. 53, 17240–17251. 02.115.
doi:10.1021/ie5004174. Sel, Ç., Bilgen, B., 2015. Quantitative models for supply chain management within
Geoffrion, A.M., Graves, G.W., 1974. Multicommodity distribution system design by dairy industry: a review and discussion. Eur. J. Ind. Eng. 9, 561–594. doi:10.
benders decomposition. Manage. Sci. 20, 822–844. doi:10.1287/mnsc.20.5.822. 1504/EJIE.2015.071772.
Guan, Z., Philpott, A.B., 2011. A multistage stochastic programming model for the Sel, Ç., Bilgen, B., Bloemhof-Ruwaard, J., 2017. Planning and scheduling of the make-
New Zealand dairy industry. Int. J. Prod. Econ. 134, 289–299. doi:10.1016/j.ijpe. and-pack dairy production under lifetime uncertainty. Appl. Math. Model. 51,
20 09.11.0 03. 129–144. doi:10.1016/j.apm.2017.06.002.
Homem-de-Mello, T., Bayraksan, G., 2014. Monte Carlo sampling-based methods for Shapiro, A., Dentcheva, D., Ruszczyński, A., 2009. Lectures on Stochastic Program-
stochastic optimization. Surv. Oper. Res. Manag. Sci. 19, 56–85. doi:10.1016/j. ming, MOS-SIAM Series on Optimization. Society for Industrial and Applied
sorms.2014.05.001, https://doi.org/https://doi.org/. Mathematics doi:10.1137/1.9780898718751, https://doi.org/doi:.
Keyvanshokooh, E., Ryan, S.M., Kabir, E., 2016. Hybrid robust and stochastic opti- Suerie, C., 2005. Campaign planning in time-indexed model formulations. Int. J.
mization for closed-loop supply chain network design using accelerated Benders Prod. Res. 43, 49–66. doi:10.1080/00207540412331285823.
decomposition. Eur. J. Oper. Res. 249, 76–92. doi:10.1016/j.ejor.2015.08.028. Touil, A., Echchtabi, A., Charkaoui, A., 2016. Possibilistic programming approach for
Kopanos, G.M., Puigjaner, L., Georgiadis, M.C., 2012. Simultaneous production and production and distribution problem in milk supply chain. In: Proc. 3Rd Ieee
logistics operations planning in semicontinuous food industries. Omega (West- Int. Conf. Logist. Oper. Manag., p. 06002. doi:10.1109/GOL.2016.7731714.
port) 40, 634–650. doi:10.1016/j.omega.2011.12.002. Tsiakis, P., Shah, N., Pantelides, C.C., 2001. Design of multi-echelon supply chain net-
Kopanos, G.M., Puigjaner, L., Georgiadis, M.C., 2011. Resource-constrained produc- works under demand uncertainty. Ind. Eng. Chem. Res. 40, 3585–3604. doi:10.
tion planning in semicontinuous food industries. Comput. Chem. Eng. 35, 2929– 1021/ie010 0 030.
2944. doi:10.1016/j.compchemeng.2011.04.012. van Elzakker, M.a.H., Zondervan, E., Raikar, N.B., Hoogland, H., Grossmann, I.E.,
Li, Q., Disney, S.M., 2017. Revisiting rescheduling: MRP nervousness and the 2014a. An SKU decomposition algorithm for the tactical planning in the FMCG
bullwhip effect. Int. J. Prod. Res. 55, 1992–2012. doi:10.1080/00207543.2016. industry. Comput. Chem. Eng. 62, 80–95. doi:10.1016/j.compchemeng.2013.11.
1261196. 008.
OECD/FAO, 2018. OECD-FAO Agricultural Outlook 2018-2027. OECD Publish- van Elzakker, M.a.H., Zondervan, E., Raikar, N.B., Hoogland, H., Grossmann, I.E.,
ing, Paris/Food and Agriculture Organization of the United Nations, Rome 2014b. Optimizing the tactical planning in the Fast Moving Consumer Goods
https://doi.org/https://doi.org/ doi:10.1787/agr_outlook- 2018- en. industry considering shelf-life restrictions. Comput. Chem. Eng. 66, 98–109.
Pham, D.T., Dimov, S.S., Nguyen, C.D., 2005. Selection of K in K-means cluster- doi:10.1016/j.compchemeng.2014.01.020.
ing. Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci. 219, 103–119. doi:10.1243/ You, F., Grossmann, I.E., 2008. Mixed-integer nonlinear programming models and
095440605x8298. algorithms for large-scale supply chain design with stochastic inventory man-
Philpott, A.B., Guan, Z., 2008. On the convergence of stochastic dual dynamic pro- agement. Ind. Eng. Chem. Res. 47, 7802–7817. doi:10.1021/ie800257x.
gramming and related methods. Oper. Res. Lett. 36, 450–455. doi:10.1016/j.orl. Zeballos, L.J., Méndez, C.A., Barbosa-Povoa, A.P., Novais, A.Q., 2014. Multi-period de-
2008.01.013. sign and planning of closed-loop supply chains with uncertain supply and de-
Prajogo, D., Olhager, J., 2012. Supply chain integration and performance: the effects mand. Comput. Chem. Eng. 66, 151–164. doi:10.1016/j.compchemeng.2014.02.
of long-term relationships, information technology and sharing, and logistics in- 027.
tegration. Int. J. Prod. Econ. 135, 514–522. doi:10.1016/j.ijpe.2011.09.001.

You might also like