17 views

Uploaded by Hussaini Hamisu

save

- Using Combinations of Heuristics to Schedule Activities of Constrained Multiple Resource Projects
- Cost Optimization b
- programacion lineal N 1.pdf
- m1010 p2
- Software Project Scheduling using Genetic Algorithms
- Product Line and Scheduling at Intel_Kempf_Wagner Presentation
- 990408
- MSSO_A_New_Tool
- Lecture 4
- 2014_Strategic Mine Planning and Optimization
- Concept of Operations Scheduling - Wba - July -Dec 2016
- Minemax Scheduler White Paper
- math 1010 project
- The Tactical Berth Allocation Problem With Quay Crane Assignment and Tran
- A Review of Analytical Models, Approaches and Decision Support Tools in Project Monitoring and Control
- Master Data for SCM 2
- Baseline Scheduling Front and Back
- optimisation
- Interactive Fuzzy Goal Programming approach for Tri-Level Linear Programming Problems
- Chapter 9
- LINEAR PROGRAMMING.pptx
- Ch 09 Aggregate Planning and Mps
- OT I Quiz 2
- Thevenin
- Lecture 1 - Intro to Math Optimisation
- 1-s2.0-S088677981530225X-main
- fgoalatt
- 1532-0545-2006-07-01-0088
- Attributes_of_a_Scheduler
- AOR-01_E
- ZZ_1237386579_BB0408121Crude_Oil_Blending
- Leaching
- A Stochastic HMM-based Forecasting Model for Fuzzy Time Series
- MODEL PREDICTIVE CONTROL OF A CRUDE DISTILLATION UNIT.pdf
- Body Size and Breast Cancer Risk in Black Women and White Women- The Carolina Breast Cancer Study.
- A comparison of fuzzy forecasting and Markov modeling.pdf
- A Multivariate Heuristic Model for Fuzzy Time Series Forecasting
- 2013-4M3-Drying
- Fuzzy Time Series Forecasting With a Probabilistic Smoothing Hidden Markov Model
- Advanced Breast Cancer in Nigeria
- Modelling of Glucose Insulin
- Multiscale modelling of biorefineries
- A crude distillation unit model suitable for online applications.pdf
- Modeling Glucose-Insulin Kinetics and Development of Type 2 Diabe
- A Comparison of Fuzzy Forecasting and Markov Modeling
- A Chemical Process Simulation
- Control Engineering and Therapeutic Drug Delivery
- Model Based Controllers for Blood Glucose Regulation in Type i Diabetics Final Thesis
- A Collection Of Ten Numerical Problems In Chemical Engineering.pdf
- Ref no. 2010.3 Pasant Suwanapal.pdf
- Modeling glucose-insulin kinetics and development of type 2 diabe.pdf

**A simultaneous optimization approach for off-line blending
**

and scheduling of oil-refinery operations

Carlos A. Méndez a , Ignacio E. Grossmann a,∗ , Iiro Harjunkoski b , Pousga Kaboré b

a Department of Chemical Engineering, Carnegie Mellon University, Pittsburgh, USA

b ABB Corporate Research Center, Ladenburg, Germany

**Received 22 December 2004; received in revised form 19 September 2005; accepted 7 November 2005
**

Available online 6 January 2006

Abstract

This paper presents a novel MILP-based method that addresses the simultaneous optimization of the off-line blending and the short-term

scheduling problem in oil-refinery applications. Depending on the problem characteristics as well as the required flexibility in the solution, the

model can be based on either a discrete or a continuous time domain representation. In order to preserve the model’s linearity, an iterative

procedure is proposed to effectively deal with non-linear gasoline properties and variable recipes for different product grades. Thus, the solution of

a very complex MINLP formulation is replaced by a sequential MILP approximation. Instead of predefining fixed component concentrations for

products, preferred blend recipes can be forced to apply whenever it is possible. Also, different alternatives for coping with infeasible problems are

presented. Sufficient conditions for convergence for the proposed approach are presented as well as a comparison with NLP and MINLP solvers

to demonstrate that the method provides an effective integrated solution method for the blending and scheduling of large-scale problems. The new

method is illustrated with several real world problems requiring very low computational requirements.

© 2005 Elsevier Ltd. All rights reserved.

Keywords: Scheduling; Planning; Blending; Refinery operations; Mixed-integer programming

**1. Introduction have been addressed through linear and non-linear mathemat-
**

ical approaches that make use of binary variables (MILP and

The main objective in oil refining is to convert a wide vari- MINLP codes) to explicitly model the discrete decisions to

ety of crude oils into valuable final products such as gasoline, be made (Grossmann, Van den Heever, & Harjunkoski, 2002;

jet fuel and diesel. The short-term blending and scheduling are Shah, 1998). Short-term scheduling problems have been mainly

critical aspects in this large and complex process. The economic studied for batch plants. Extensive reviews can be found in

and operability benefits associated with obtaining better-quality Reklaitis (1992), Pinto and Grossmann (1998), Kallrath (2003)

and less expensive blends, and at the same time making a more and, Floudas and Lin (2004). Much less work has been devoted

effective use of the available resources over time, are numerous to continuous plants. Lee, Pinto, Grossmann, and Park (1996)

and significant. A wide variety of mathematical programming addressed the short-term scheduling problem for the crude-oil

techniques have been extensively used for long-term planning inventory management problem. Non-linearities of mixing tasks

as well as the short-term scheduling of refinery operations. were reformulated into linear inequalities with which the orig-

For planning problems, most of the computational tools have inal MINLP model was converted to a MILP formulation that

been based on successive linear programming models, such as can be solved to global optimality. According to the authors,

RPMS from Honeywell, Process Solutions (formerly Booner this linearization was possible because only mixing operations

& Moore, 1979) and PIMS from Aspen Technology (formerly were considered (see Quesada & Grossmann, 1995). However,

Bechtel Corp., 1993). On the other hand, scheduling problems it was later pointed out by Wenkai, Hui, Hua, and Tong (2002)

that the proposed reformulation linearization technology (RLF)

may lead to composition discrepancy (the amounts of individ-

∗ Corresponding author. Tel.: +1 412 268 2230; fax: +1 412 268 7139. ual crudes delivered from a tank to CDU are not proportional to

E-mail address: grossmann@cmu.edu (I.E. Grossmann). the crude composition in the tank). The objective function was

**0098-1354/$ – see front matter © 2005 Elsevier Ltd. All rights reserved.
**

doi:10.1016/j.compchemeng.2005.11.004

C.A. Méndez et al. / Computers and Chemical Engineering 30 (2006) 614–634 615

the minimization of the total operating cost, which comprises mentation of decision support systems for off-line multi-period

waiting time cost of each vessel in the sea, unloading cost for blending problems at Texaco. Commercial applications such as

crude vessels, inventory cost and changeover cost. Several exam- Aspen BlendTM and Aspen PIMS-MBOTM from AspenTech are

ples were solved to highlight the computational performance of also available for dealing with online and offline blending opti-

the proposed model. Moro, Zanin, and Pinto (1998) developed mization problems. Since these software packages are restricted

a mixed-integer non-linear programming planning model for to solving the blending problem, resource and temporal deci-

refinery production. The model assumes that a general refinery sions must be made a priori either manually or by using a special

is composed of a number of processing units producing a variety method.

of input/output streams with different properties, which can be To solve both sub-problems simultaneously, Glismann and

blended to satisfy different specifications of diesel oil demands. Gruhn (2001) proposed a two-level optimization approach

Each unit belonging to the refinery is defined as a continuous where a non-linear model is used for the recipe optimization

processing element that transforms the input streams into sev- whereas a mixed-integer linear model (MILP) is utilized for the

eral products. The general model of a typical unit is represented scheduling problem. The proposed decomposition technique for

by a set of variables such as feed flowrates, feed properties, the entire optimization problem is based on solving first the

operating variables, product flowrates and product properties. non-linear model aiming at generating the optimal solution of

The main objective is to maximize the total profit of the refin- the blending problem, which is then incorporated into the MILP

ery, taking into consideration sales revenue, feed costs and the scheduling model as fixed decisions for optimizing only resource

total operating cost. Wenkai et al. (2002) proposed a solution and temporal aspects. In this way, the solution of a large MINLP

algorithm that iteratively solves two mixed-integer linear pro- model is replaced by sequential NLP and MILP models. Jia

gramming (MIP) models and a non-linear programming (NLP) and Iearapetritou (2003) proposed a solution strategy based on

model, resulting in better quality, stability, and efficiency than decomposing the overall refinery problem in three subsystems:

solving the MINLP model directly. Kelly and Mann (2003a,b) (a) the crude-oil unloading and blending, (b) the production unit

highlight the importance of optimizing the scheduling of an operations, and (c) the product blending and lifting (see Fig. 1).

oil-refinery’s crude-oil feedstock from the receipt to the charg- The first sub-problem involves the crude oil unloading from

ing of the pipestills. The use of successive linear programming vessels, its transfer to storage tanks and the charging schedule

(SLP) was proposed for solving the quality issue in this problem. for each crude oil mixture to the distillation units. The second

More recently, Kelly (2004) analyzed the underlying mathemat- sub-problem consists of the production unit scheduling, which

ical modeling of complex non-linear formulations for planning includes both fractionation and reaction processes. Reactions

models of semi-continuous facilities where the optimal opera- sections alter the molecular structure of hydrocarbons, in gen-

tion of petroleum refineries and petrochemical plants was mainly eral to improve octane number, whereas fractionation sections

addressed. separate the reactor effluent into streams of different proper-

In addition, the off-line blending problem, also known as ties and values. Lastly, the third sub-problem is related to the

blend planning has been addressed through several optimization scheduling, blending, storage and lifting of final products. In

tools. The main purpose here is to find the best way of mix- order to solve each one of these sub-problems in the most effi-

ing different intermediate products from the refinery and some cient way, a set of mixed-integer linear models (MILPs) were

additives in order to minimize the blending cost subject to meet- developed, which take into account the main features and dif-

ing the quality and demand requirements of the final products. ficulties of each case. In particular, fixed product recipes were

The term quality refers to meeting given product specifications. assumed in the third sub-problem, which means that blending

Rigby, Lasdon, and Waren (1995) discussed successful imple- decisions were not incorporated into this model. The MILP for-

Fig. 1. Illustration of a standard refinery system.

sequence constraints.616 C. This second issue is also known as the on the notion of event points. tion and the properties of the components used in the blend. blend headers and product storage assumptions are: (a) fixed recipes for different product grades tanks. which allows the model to generate optimal. (12) Preferred product recipes. Specifically. typically 7–10 days. (c) the volume fraction of components included in each product. (10) A set of correlations. 1992). takes into account two major issues. Most of these features are tive function includes the total product value. Its importance comes from (11) Minimum and maximum component concentrations in the fact that gasoline can yield 60–70% of total refinery’s profit. which represents an additional difficulty for standard function to provide effective solutions for all circumstances. for predicting the refinery operations. and a specific objective objective is to produce on-spec blends at minimum cost. (f) the optimal timing decisions for production and storage The gasoline short-term blending and scheduling problem tasks.b). (2) A set of intermediate products from the refinery (compo- ing the majority of the problem features. inventory cost and penalties for deviation from preferred as discrete and continuous decisions in an optimization frame. which is generally agreed as being the most values of properties of each blend. in order of these non-linear terms can sometimes be linearized (see to reduce the inherent problem difficulty. and (c) all product properties are assumed to be lin- ear. Depending on the product property. Product qualities are usually predicted variables are principally related to assignment decisions of tasks through complex correlations that depend on the concentra- to event points. as pointed nents). The objec- different logistic and operating rules. Taking into account the major (8) A set of final products with predefined minimum and max- weaknesses of the available mathematical approaches. The first one is related to aspects of production logistics. (5) Component supplies with known flowrates. recipes. as well as the limited number of available resources that The goal is to determine: can be used to reach the production goals. However. On the other hand. The objective is to maximize the production profit while tiple production demands with different due dates. (6) Properties or qualities for components. On the other hand. demand constraints. tanks. Some of the common of component storage tanks. or near. Problem statement (e) the pumping rates for components and products. process topology corresponds to a multi-stage system composed ficient or unrealistic for real world cases. trilinear and exponential terms. Non-linear property specifications based on variable and preferred product recipes (a) the assignment of blenders to final products. the second issue is the production infeasible solutions can also be incorporated into the objective quality. Binary most of the markets. 2003. out by several authors solving logistics and quality aspects for (3) A set of dedicated storage tanks with minimum and max- large-scale problems is not possible in a reasonable time with imum capacity restrictions. more general mixed-integer non-linear (1) A predefined scheduling horizon. Some short-term blending and scheduling problem. The general on special assumptions that generally make the solution inef. 2. we assume that we are given the following are predefined. current mixed-integer non-linear programming (MINLP) codes (4) Initial stocks for components. The major issue here is related to non. important and complex subproblem. / Computers and Chemical Engineering 30 (2006) 614–634 mulation was based on a continuous time representation and scheduling problems. and global optimization techniques (Jia and Ierapetritou. non-linear correlations may matical programming approaches are currently available to the include linear. the raw material part of typical scheduling problems and are usually modeled cost. goal of this work is to develop a novel iterative mixed-integer (9) A set of equivalent blend headers working in parallel that linear programming (MILP) formulation for the simultaneous can be allocated to each final product. which mainly involves mul. From the above review it can be seen that a variety of mathe. are effectively handled through the proposed iterative linear (b) the inventory levels of components and products in storage procedure. etc. programming (MINLP) formulations are capable of consider. inventory satisfying the process. most of them rely Adams & Sherali. capacity constraints. Kelly and Mann. bilinear. (d) the total volume of each product. final products. Continuous variables are defined to represent flowrates product specifications are stringent and constantly changing in as well as starting and ending times of processing tasks. Additional terms involving slack variables for handling work. pumping constraints for products and components. operations and maintenance constraints. mostly non-linear. linear and non-convex constraints with which the computational (7) Minimum and maximum flowrates between component performance strongly depends on the initial values and bounds tanks and blend headers. optimal solutions with modest computational effort. ber of product demands and quality specifications for each final product. 1990. The . number. the complexity arises from the large num. The mathematical formulation off-line blending problem and takes into account variable prod- proposed to solve each sub-problem involves material balance uct recipes and property specifications such as minimum octane constraints. Méndez et al. gasoline short-term blending and scheduling problem of oil. assign. maximum sulfur and aromatic content. or to some specific aspect of each sub-problem. assigned to the model variables. where function. as well as final product demands and quality specifications.A. However. 2003a. The main ment constraints. (b) component and product flowrates are known items: and constant. On the other hand. Oral & Kettani. the major imum quality specifications.

depending on the problem characteristics. as well as Before describing the proposed MILP formulation. The solution of the scheduling problem defines the way in in time slot t.t P through the and Fp. The continu- I and V P define the amount of component and ous variables Vi. Finally.p. Additional continuous and discrete variables can be included into the mathematical model to tackle • A multi-period optimization model is used that is able to deal particular problem characteristics and operating constraints. respectively. (2003a.t I . for product p in time slot t and pri. the discrete variable Ap.k vIi. computed by using a volumetric average as shown in Eq. • The production logistics and quality specifications are solved 4.t is linked problem defines how the available components are blended or I to the volumetric flowrate variables Fi. • Fixed or variable product recipes are specified.t P = Fi.t 3.t . k. Proposed optimization approach p.A. k.p. etc.p.p. the MILP computational effort can be significantly increased. non-linear equality (2): The key decision variables involved in a standard problem are the following. • Linear approximations are used together with an iterative pro. during the time slot t whereas Fp. A comprehensive used.p.p. Méndez et al.k. particularly when more than 20 time-periods or time-slots are cedure to get better predictions of all product properties. PRp. The continuous variable Fi. When some features such as inter-temporal trans- val with known duration and position for discrete time.t defines which products are allocated to rized as follows: blenders in each time slot t. tinuous time formulations are focused on a part of the details • Discrete or continuous time domain representations can be around real-world problem logic constraints. t (1) In order to describe the main model variables. as well as any other logistic or production rule found in the A significant number of gasoline properties can be directly problem. Illustration of the meaning of the principal model variables.e. even defined in the model. ∀p.tP denotes the volumetric flow of product p being blended during each time slot t. Fig.t mixed together to produce on-spec products with minimum cost. where vIi. those naturally non-linear such as the octane number. C. and fer of logic are addressed (i.p. present in this section the main features of the iterative scheme • Binary variables are used to represent assignment decisions.k is the value of the property able equipment.k. run-lengths.t Fp.t product being stored at each time point t. for predicting product properties for the blending problem.t is the volume fraction of component i in product p lems. with multiple product demands with different due dates and It is worth mentioning that both the discrete and the con- quality specifications.). we minimum and maximum limits on component concentration.t defines the exact value of the property k which the products are processed with respect to time and avail. which has traditionally been treated as two separate prob. (1): PRp.t I defines the vIi. ∀p.b). t (2) Fig. mixing-delays. 2. The main features of the proposed approach can be summa. The volume fraction variable vIi. The term enumeration of most of them can be found in Kelly and Mann “time slot” is used in this paper to represent a time inter. On the other hand. unknown duration and position for continuous time. the solution of the blending k for component i. . / Computers and Chemical Engineering 30 (2006) 614–634 617 next section describes a simultaneous optimization approach for volumetric flow of component i being transferred to product p the blending and scheduling operations involved in this problem.t = pri. 2 illustrates i a simple example of a gasoline blending and scheduling prob- lem. Off-line blending problem simultaneously.

yielding Eq. t (3) i We want to point out that there are some complex blending Then.t is substituted by To illustrate the use of the iterative procedure and the pro- their respective minimum and maximum property values (prmin p. i Thus.k Fp. 2004. the rect this discrepancy.k.k.t is no longer required and the considered. 3. as shown in Eq. k.t ≤ prmax p. . even if only linear product properties are considered. (6): For this specific mixture of components the correction factor bias is equal to 0.k Fi. This difference arises because all instance.k Fi.t is a model variable that is not directly computed through the proposed linear approach and with the intention of maintaining the model’s linearity.t P . is only In this way. (7). the blend of two components. which are known problem data: the values of the linear volumetric average. which can be easily Fig. ∀p. P P ∀p. Given that sgp.k sgi Fi.tP : I i pri. The final product property is a non-linear function model remains linear.t can be substituted by an approximated specific gravity sgravp. In order to pre.p. (3). this correction factor depends on the product. The linear approximation comprising the p.k.t ≤ prmin ≤ prmax ∀p. t (7) mathematical model for the integrated blending and scheduling sgp. and it is iteratively calculated by using the procedure described below.t Fp.t . constraint (4) can be replaced problem that is not addressed in this paper. ∀p. t (8) sgravp. k. 3 shows a comparison between prmax p.k Fp. 2003a. A and B. which is composed of a vol- umetric average followed by a correction factor ‘bias’.t problem comprises a set of constraints with bilinear terms.b).k Fp. by constraint (5).k Fp. This linearization is valid only if prop. the variable vIi.p.5 and 88. k.t P PRp.A. This is a higher-degree multi-linear product property specifications. As can be seen. posed linear approximation.t + biasp. For example. Méndez et al.k.t = .t .t .t prp. k.p.t prmin P P volumetric average together with the correction factor bias will always predict the exact value of the property if the same compo- ≤ I pri. As shown in Fig. equality (2) can be incorporated into Eq. it is worth mentioning that some product properties such as oxygen and sulfur content are blended gravi- metrically. of component concentrations. the values problem. respectively.t + biasp.t Fp. taking advantage of minimum and maximum calculated endogenously.k. other gasoline properties can be approxi. t (4) be the requirement to blend to a distillate diesel sulfur quality i specification. the correction factor bias is calculated and non-linear part of the equation can be removed and replaced by used to yield a better property prediction in the next iteration. k. Kelly & Mann.p.t PRp. sgi and sgp. / Computers and Chemical Engineering 30 (2006) 614–634 Taking into account that volumetric flowrate variables are computed through the iterative procedure described below required to control inventory levels in tanks and volume fraction I variables are needed to predict product properties.t . 3. respectively. However. the exact value of sgp.k Fi. the general i pri.t = pri.t p.k sgi Vi. Fig.t = I pri. we need the diesel yield as well as the diesel specific gravity in each component blending stream to be Subsequently. In order to cor- non-linear terms.24. non-linear product properties can be approximated through the linear equation (6).t .k Fp. which means that component and product specific gravities are also taken into account for the prediction.p. (5) non-linear product property such as the motor octane number. the original equality (1) can be blending is as follows: expressed in an alternative way by multiplying it by Fp. decisions (especially in crude-oil blending) that may require to (4): explicitly define yield variables for streams within streams (see Kelly. if the correlation for predicting a particular product non-linear terms involved in the exact motor octane correlation property is based on a linear volumetric average plus additional are not included in the linear volumetric average.k vIi. a correction factor biasp.618 C. i In this example.k.p. in which the variable PRp. t P I P pri.k ). the non-linear origi- nal correlation and the proposed linear approximation for a real p. Furthermore. On the other hand.t ≤ min P ≤ prmax P p. such as the case of the octane number. ∀p. there may P PRp. For 88.k .t Fp. t (6) nent concentration is utilized in the next iteration. Therefore.t Fp. ∀p.p. prop- erty and time slot.k. In this case. A non-linear property and the proposed linear approximation. if 40% of erties computed volumetrically are considered in the blending component A is blended with 60% of component B.t define the specific gravity of component i and product p in time slot t. Therefore.t Fp.k Fp.t . the proposed linear approximation for gravimetric serve the linearity of the model. of the volumetric average and the real non-linear correlation are mated by adding minor changes to the previous equation. k. as shown in Eq.74.

/ Computers and Chemical Engineering 30 (2006) 614–634 619 it was observed that the proposed linear approximation tends ear volumetric average. However. if preferred recipes are not defined. lem or a successive MILP model for the simultaneous blending linear correlation and the linear volumetric average are both used and scheduling problem. the solution of this problem is updated the best alternative for blending because they satisfy all prod. only tories and product demands and for this reason. Preferred recipes are imation. biasp. On the other until all product recipes meet the product specifications.A.k. products and time intervals as the difference between the of a non-linear property does not become active at the solution. The first step is to find an initial recipe cally. the use of them cations in a specific time interval are fixed. all hand. we provide sufficient to calculate the correction factor ‘bias’ (see Fig. the teed to converge to a local solution of the blending problem. The value predicted by the non. and the product recipes for those products meeting all specifi- uct specifications with minimum cost. The specific gravity of each product to predict a very close value of the real property if component and time slot is also computed. The proposed method can be conceptually blend. Méndez et al. For non-linear properties this parameter will converge to a for all products. component inven. 4. they should not those that are feasible will be fixed. for volumetric properties and the parameter sgravp. while optimizing all temporal and resource constraints in the ated.t . Proposed iterative approach for simultaneous blending and scheduling. are used for the same product in different time intervals. If different recipes strongly depends on the scheduling decisions. When that is not the case. the LP model is concentrations are not significantly changed in the next iteration solved that includes linear approximation with the parameter as shown in Fig. correction factor will be calculated for all non-linear proper. i. they provide the component volume fractions used in each scheduling problem. In Appendix A.t . Subsequently. The proposed iterative procedure to solve simultaneously the for gravimetric properties. These conditions are rigorous for the case that a specification ties. 3. Given that conditions under which the successive LP method is guaran- we are dealing with a multi-period optimization problem. which can then be employed as fixed parameters in more interpreted as a successive LP method for the blending prob- realistic non-linear correlations. value predicted by the original non-linear equation and the lin. If preferred product recipes are known they non-zero value that reflects the difference with the linear approx- can be proposed as initial product recipes. 4. 3). The main objective of this iterative pro- generating initial recipes is to solve the LP model including cedure is to progressively find feasible recipes for all products only linear product properties.e. The parameter bias will be equal to blending and scheduling problem using only linear equations zero for all linear properties that can be computed volumetri- is shown in Fig. the proof requires the assumption that Fig. This process is repeated be treated as fixed mixtures in a blending tool. After that. C. Once initial recipes are gener. one possibility for product recipes are fixed. .k.

Integrated blending and scheduling model nB t maximum number of blenders that can be working in parallel in time slot t A central aspect of any scheduling model is related to timing + decisions. Méndez et al. Since the best choice of the time rcpi. / Computers and Chemical Engineering 30 (2006) 614–634 the non-linear properties are a weak function of the composi. which of course is a strong assumption. inip initial inventory of product p lpmin minimum time slot duration when it is allocated to product p 5.t shortage of component i that is used for product p in I set of intermediates to be blended K set of properties for intermediates and products time slot t according to the preferred product recipe R + KNL set of properties that are predicted with non-linear cor. which is particularly relevant for industrial inii initial inventory of component i applications. sgravp. prmax maximum value of property k for product p p. However.k value of property k for component i capable of generating more flexible solutions in terms of tim. Before present.p penalty cost for a deviation from the minimum speci- place at certain time points.620 C.d demand of product p to be satisfied at due date d which they confirm the solutions obtained with the proposed et predefined ending time of time slot t (discrete time rep- method.p preferred concentration of component i in product p representation strongly depends on the problem characteristics according to product recipe and the desired solution quality. The dis.p penalty cost for a deviation from the maximum speci- using a discrete time grid is that mass balance and inventory fication for property k in product p constraints are easier to handle but at the same time the solution pltySH penalty cost for purchasing component i from third- i loses flexibility.p.t correction factor of the value of property k of product will be shown later in the paper. in the section of computational results d demand due date we present a comparison with NLP and MINLP methods in ddp. only few iterations are needed p in time slot t to obtain a very good solution for the blending and scheduling ci cost of component i problem.t deviation from the minimum specification of property P set of final products T set of time slots k for product p in time slot t Td set of time slots postulated for the sub-interval ending S+ Dk. fication for property k in product p aries of a set of predefined time slots. as it biasp. Thus.A. inventory and mass balance constraints are generally more ratemax p maximum flowrate of product p difficult to model since they have to be checked at any time dur- ing the scheduling horizon in order to ensure that a feasible ratemin p minimum flowrate of product p solution will be generated. or resentation) in some cases failing to converge depending on the starting point fi constant flowrate of component i used. In contrast.p maximum concentration of component i in product p formulation for each type of representation assuming a common rcpmin i.t specific gravity of product p in time slot t ing the proposed mathematical models the nomenclature is as st predefined starting time of time slot t (discrete time follows: representation) spi penalty cost for inventory of component i Indices spp penalty cost for inventory of product p d due dates of product demands Vimax maximum storage capacity of component i i intermediates or components Vimin minimum storage capacity of component i k properties or qualities Vpmax maximum storage capacity of product p p final products or gasoline grades Vpmin minimum storage capacity of product p t time slots Variables Sets Ap.k ing decisions. we developed a mathematical rcpmax i.p minimum concentration of component i in product p time grid for all resources working in parallel. Di. unless smaller time intervals are used. The main advantage of − pltySk.k. Furthermore. Parameters tions. prmin p. although with higher CPU time requirements. providing h time horizon very good solutions.k minimum value of property k for product p Also. the proposed method is fast and reliable. Mathematical formulations can be based on either pltyR i.p. pltyR i.p.p. but at either considerable higher computational cost. continuous time representations are pri. which party may significantly decrease the computational performance of pp price of product p the method.p penalty cost for excess of component i in product p − a discrete or continuous time domain representation.p penalty cost for shortage of component i in product p + crete time representation only allows processing tasks to take pltySk.t binary variable denoting that product p is blended in D set of product due dates time slot t R− Di.t deviation from the maximum specification of property at due date d (continuous time) k for product p in time slot t .t excess of component i that is used for product p in time relations slot t according to the preferred product recipe S− Dk. which correspond to the bound.

blending tasks. inventory limits have only to be checked at the end of .t = Fp. A discrete time domain representation is used where the and rcpmax i. Minimum/maximum component concentration P Vp.t amount of component i stored at the beginning of time equations. Then. The proposed i. Product composition constraint tion) I Fi.t = 0). unless they are the only possibility to produce a particular prod- 2.d amount of product p stored at due date d 6. the volumetric flowrate can be computed time slot. 3. Constraint (12) specifies that minimum and maximum vol- 5. Equivalent blenders working in parallel are available for dif.t volume fraction of component i in product p at time t P is replaced by and Fp. Material balance equation for components Given that a discrete time representation allows the blend- Ap. preferred recipes can be proposed as an ferent product grades. since a discrete time representation Model constraints and variables are introduced below. ratemin ∀p.t P . t P I max P (11) model has the following features: It should be noted that a fixed recipe for a particular product p can also be taken into consideration by fixing the values of rcpmin i.t the final products p to be processed in time slot t.5. Due to the fact that a constant product flowrate 6. Parameters rcpmin max In this section. Moreover.1. 6. the volumetric flowrate will be also 6. constraint (11) ensures that product composition will always satisfy the predefined com- ponent specifications. slot t V Pp. (10) is deleted from the model vIi. as expressed by constraint (10): P Fp.t ≤ rcpi. ∀p. p. nB t specifies the maximum number of units that can be working in parallel during time slot t: 6.p. Assignment constraint equal to zero: p (et − st )Ap.t . Constant flowrate of products is assumed during the allocated is assumed in this work.t exact value of the property k for product p in time t I Fi. Variable product recipes are considered and product proper- ties are predicted by linear correlations. C. initial solution of the proposed iterative procedure.p 1. in order to make the i i.p Fp. the time slot duration is a known parameter computed through the predefined starting st and ending times et of each time slot t.p Fp.p define the the entire scheduling horizon is divided into a finite number of minimum/maximum concentration of component i for product consecutive time slots that are common for all units and can be p.t ≤ Fi.t amount of component i being transferred to product p Every final product or gasoline grade p is a blend of different during time slot t components i.e. t (10) Si. by multiplying the upper and lower flowrates by the time slot duration whenever product p is allocated to a particular time slot t (Ap.p.p. It should be noted that if product p is not processed during time slot t (Ap. ∀t (9) ing tasks to start and finish at the boundaries of the time slot p allocated. i. Méndez et al. upper and lower bounds can be forced on the component concen- 6. Constant flowrate of components is assumed during the entire umetric flowrates must be satisfied when product p is blended scheduling horizon. we present an MILP model that assumes that i. Fp.t I Vi. t P max Constraint (9) defines through the binary variables Ap.t = 1). A particular product demand can be satisfied by one or more the generation of infeasible solutions will be avoided.t amount of component i stored at the end of time slot t P has been included in all model model easier to understand. / Computers and Chemical Engineering 30 (2006) 614–634 621 Et ending time of time slot t (continuous time representa.t ≤ Fp.p and rcpi. 6.t amount of component i to be purchased in time slot t i St starting time of time slot t (continuous time represen. uct.t .2.t V Ii.t F I . respectively: allocated to different products. However. rcpmin ∀i.t amount of product p being blended during time slot t PRp. However.t amount of product p stored at the end of time slot t In order to satisfy product qualities and/or market conditions.t ≤ ratep (et − st )Ap.k. is used.p to the predefined concentration of component i for scheduling horizon is divided into a set of consecutive time product p. Note that a significant reduction in the number of continuous tation) variables can be obtained if Eq.p. time slots whenever they are allocated to this product and finished before the product due date.t ≤ nB t . As a better option. Discrete time representation tration for specific gasoline grades. the use of fixed recipes should be avoided slots.3. In this way. Minimum/maximum volumetric ﬂowrates for products 4. during time slot t.4.A. Given that a set of (12) equivalent blenders are available to produce different gasoline grades simultaneously.p.

Objective function (maximize net proﬁt) 6. d (18) t≤d d ≤d P Vp.k ρiFi. the MILP mathe. The best value of this parameter can be obtained through the proposed iterative procedure.t Fi. An example is Eq.11.t . con. non-convex bilinear equations would be generated in the model. − I spi Vi.t (20) Although this linearization is only valid for properties volu. t (14) Refinery operations typically require that multiple demands for the same gasoline grade be satisfied during the entire schedul- 6.t + biask.9. ∀p. the parameter fi specifies the constant production rate of component i and et defines the ending time of time slot t. Then.p. Continuous time representation biasp. If a fixed recipe is enforced.t .k Fp.k Fp.p. Given that constant component satisfied in advance through the predefined component concen- flowrates are assumed.10. property k max pp Fp. (20). crete time domain representation. a perfect coordination between the pro. production and demands of product p: inip + P Fp.k Fp.t I I Vi. t (15) t ≤t d≤t 6.t prp. Méndez et al.p.t = inii + fi et − I Fi.t P ≤ Vpmax . as expressed by constraint (13). t (17 ) sgravp. / Computers and Chemical Engineering 30 (2006) 614–634 each time slot.p. To maintain the model’s linearity. The model introduced in the previous section relies on a dis- matical model is able to effectively deal with the quality issue.p.t − I ci Fi. Component storage capacity Note that constraints (17). i t metrically computed. I i pri.t ≤ prmin ≤ prmax ∀p.d : stored in tank at the end of time slot t taking into account the initial inventory. product properties must be during the scheduling horizon.k.t = inip + P Fp.t I ≤ Vimax .6.7. Material balance equation for products ing horizon. Given that a In turn.k Fp. duction of components and final products is required to satisfy the storage constraints through the entire scheduling horizon: 6. (17 ) defines the proposed linear approximation discrete time representation is used.t − ddp.t − I P spp Vp.t 6.t ≤t ≤ prmax P p. the . (17) nent produced up to the end of time slot t minus the component i transferred to blenders up to the end of time slot t: p.k Fi. t (13) i p.8. t (16) max P pp Fp.t (19) t p i 6. (17 ) and (17 ) are only required Constraint (14) imposes lower/upper bounds Vimin and Vimax for those gasoline grades that can be produced using variable on the total amount of component i being stored in a storage tank recipes.d .t .k Fp.d .A. Otherwise. both parameters are known for those product properties that are gravimetrically predicted: in advance. ∀i. the amount including variable recipes and non-linear properties: of component i being stored in tank at the end of time slot t is p.k Fp.t ≥ ddp.p. the main objective of the scheduling problem is to maximize the net profit A minimum safety stock and a finite storage capacity is defined as the total product value minus the total component cost: assumed for final products: Vpmin ≤ Vp. t (17 ) where inii is the initial inventory of component i at time t = 0. ∀p.t is not directly computed and bounds are only imposed on each t p i p t property.622 C. the original equation (17) can be slightly modified as Eq. ∀i. ∀p. As an alternative option. Constraint (18) guarantees that a sufficient amount of product p will always be available to satisfy each product Constraint (15) computes the amount of product p being demand ddp. t P I P equal to the initial inventory of component i plus the compo. ∀p. uct specifications. ∀p. where penalties related to Assuming that properties are volumetrically computed.t ≤ min P ≤ prmax P p.t . k. Minimum/maximum product qualities The formulation can also accommodate alternative objective functions.t ≤ prmin P I pri. as described in Section 4 with the use of the parameter 7. trations. component and product inventories has been included in order straint (17) guarantees that the value of property k for product p to also reduce storage costs: in time slot t will always satisfy minimum and maximum prod. k.t . Eq. which would then become non-linear. k.t p.p.p.k Fi. Product storage capacity While satisfying all quality and logistic issues. pri. In this way. (17 ) to account for real-world product prop- erties. Multiple product demands Vimin ≤ Vi.t − P ci Fi.

I Vi.t ≤ ratemax p hAp. Ap. whenever product p is processed during time slot t. starting and ending times for the time to make the modeling of inventory constraints more difficult. constraint (12 ) becomes redundant and constraint (12 ) imposes posed continuous time domain representation. time slot duration: which means that four sub-intervals are predefined. The proposed model has the following features: P Fp. ∀p. where two time slots are defined for each one of the ≤ Fp. Equivalent blenders working in parallel are available for dif- at the end but also at the beginning of each time slot. ∀i. For that reason. t (12 ) first three sub-intervals whereas three are postulated to the last one. To make this ferent product grades possible.t ≤t 5.t <t Note that despite the fact that Et and St are model variables. Méndez et al.t ) ratemin min time slots can be postulated for the entire scheduling hori- zon. a set of time slots with unknown duration are postulated for each one. 7. t p. A particular product demand can be satisfied by one or more original equation (13) is replaced by constraints (13 ) and (13 ). the for each one. The length of each sentation must be updated in order to maintain the linearity of sub-interval will depend on the product due dates. the model as well as to account new problem features. time slots whenever they are allocated to this product and The same idea for computing the inventory of components is finished before product due date. 5 illustrates the main features of the pro. Minimum/maximum volumetric ﬂowrates for products Once the sub-intervals are defined. Constraints that are not will be extended from the first up to the second product due date. applied to these new constraint: 4. The main idea here is first to partition the entire time horizon part of the original constraints used for the discrete time repre- into a predefined number of sub-intervals. the way they were presented in the previous section. (17 ) and (19). I Fi. A continuous time domain representation is used where the 7.p. Therefore. Material balance equation for components scheduling horizon is divided into sub-intervals and a set of time slots with unknown duration and position are postulated To ensure that only feasible solutions are generated. C. ∀i. a new variable V Ii.t will be equal to one production events to happen at any time during the schedul. Fig. For instance. special attention must be paid to the lim.t . four minimum and maximum volumetric flowrates depending on the product demands with different due dates are to be satisfied. p. 7.t . amount of component stored in tank has to be checked not only 2. Then. t (12 ) 1. Variable starting and ending times of time slots zero and constraint (12 ) enforces the variable Fp.t . nine p (Et − St ) − ratep h(1 − Ap.t P to be equal are introduced as new model continuous variables that allow the to zero as well. On the other hand. When product p interval length as well as the grade of flexibility desired for is not allocated to time slot t.2. Final product properties are based on a volumetric average and a correction factor computed through the proposed iter. In this case. The second one as well as the new ones to be added. required to change must be included into the model in the same A similar idea is applied to the next sub-intervals. ing the plant performance without significantly increasing the model size.p.t P ≤ ratemax p (Et − St ). number of sub-intervals will be equal to the number of product (9)–(11). the binary variable Ap.1. However. due dates and the starting and ending time of each one will be known in advance.3. when a continuous time representation is used. In this the first sub-interval will start at the beginning of the scheduling section we describe the set of constraints that must be modified horizon and finish at the first product due date.t is included into the model and the 3. Component storage capacity An additional constraint (14 ) is required to impose Fig. A constant flowrate of product is assumed during the allo- aims at generating more flexible schedules capable of maximiz. / Computers and Chemical Engineering 30 (2006) 614–634 623 continuous time formulation that is presented in this section 6. t (13 ) ative process. The number Constraints (12 ) and (12 ) replace original constraint (12) of time slots for each sub-interval will depend on the sub.t = inii + fi Et − I Fi. slots are new continuous decision variables. both constraints remain linear because a constant production rate fi is assumed for the components. such as Eqs. A constant flowrate of components is assumed during the V i.A. (14)–(17). lower/upper bounds Vimin and Vimax on the total amount of com- .t is equal to the solution. ∀p. When the mathematical model is based on a continuous time ited storage capacity since continuous time representation tends domain representation. In this case. 5. (17 ). Proposed continuous time representation. ing horizon.t = inii + fi St − (13 ) I entire scheduling horizon. cated time slot.

d 7. Méndez et al. according Constraint (23) establishes a sequence between consecutive to the preferred product recipe.t i.t . In this way. ∀i.t .9. ∀t ∈ Td (24) due date d.7.p. ∀d. Constraint (29) penalizes the time slots t and t + 1: R+ and DR− in the objective to ensure that slack variables Di.t ) (pltyR (29) 7. recipe whenever it is possible: (22) is included into the model: − P rcpi.t + Di. ∀t (22) + p P rcpi. p. Treatment of infeasible solutions as fixed parameters.p.5. t constraints are required to establish which time slots can satisfy a specific product demand. a minimum product safety stock can be guaranteed at any time during the scheduling horizon.4.8. (16 ) P ∀p. p.A. Time slot duration These penalties can partially relax some hard problem specifi- cations that can generate infeasible solutions when real world Constraint (21) defines a minimum time slot duration when problems are addressed.p Di. t (28) R + R − 7. In order to allow more flexible solutions and avoid ations is a very complex and highly constrained problem. 7.t deviations from the preferred recipe are minimized: Et ≤ S(t+1) . Sub-interval bounds t p i The set Td comprises all time slots that are postulated for a 8.p. Penalty for minimum/maximum speciﬁcation deviation sub-interval related to a particular due date d. Penalty for preferred recipe deviation eliminating schedules using very short time slots. even after Et ≤ d. ∀p. ∀i.t + pltyi. In other words.(t+1) ≤ nBt Ap. t (21) constraints can be included in the model to try to use the desired To ensure that duration of a slot is zero if it is not used.p.2. It is generally used to model an existing operating condition. If desired product qualities cannot be fully achieved. (t.p.p. t (27) Et − St ≤ h Ap.t − ddp. in this section we present an constraints. d Ap.t define the excess and the shortage of component i that is used in product p in time slot t. As a result.p Fp. a time slot t + 1 can be only inventory variable: allocated to a product p whenever the previous time slot within the same sub-interval has been used: Vpmin ≤ V p. t + 1) ∈ Td (26) 7. a correct order and sequence between even feasible solutions may be difficult to generate in some postulated time slots must be established through the next set of circumstances.p.d = inip + (15 ) P P Fp.t ≤ Vimax . the following Et − St ≥ lpmin Ap.p. in continuous time representation models these decisions are treated as additional continuous variables to The short-term blending and scheduling of oil-refinery oper- be optimized.6. Set of time slot timing constraints p p Instead of defining starting and ending times of time slots 8. Time slot sequencing where Di.t . and at Constraint (24) defines that time slots pre-allocated to this sub. Eq.t and where Di. additional set of variables and equations that defines penalties that can be added to the objective function of the proposed model. more flexible schedules 7. (14 ) I ∀i. For that reason.p. which are usu- ally inefficient in practice: If a preferred combination of components is defined for a particular product through the parameter rcpi. interval must start after due date d − 1 whereas constraints (25) the following constraints can be used in order to minimize the . they can partially be violated for certain products.d .1.624 C.t − Di. ∀p.p Di. Time slot assigment t<d d ≤d Constraint (26) imposes an order for using the set of prede- Constraint (16 ) explicitly defines the lower bound on the new fined time slots. the same time.t .t . The main goal of this assumption is that neither additional variables nor new Vimin ≤ V i.d . ∀t (23) + R+ R− R− penalty = i. / Computers and Chemical Engineering 30 (2006) 614–634 ponent i being stored in tank at the beginning of time slot t: imposes that they must end before due date d.p Fp. product p is allocated to time slot t.t R ≤ Fi. Material balance equation for products can be obtained without increasing the complexity of inventory constraints: Constraint (15 ) computes the inventory of product p available right after satisfying the demand of product p taking place at St ≥ d − 1. but at the same time permits 8.t R ≥ Fi.p . ∀t ∈ Td (25) a product lifting is carried out: V p. This sub-interval begins at the previous due date d − 1 and finishes at due date d. where overlapping time slots.p.

10 100.00 Min stock (Mbbl) 5.80 P12 0.04 0.00 3.00 100. S+ propmin P p.60 33.t + pltySk.00 50.50 100. Penalty for intermediate shortage basis of the example comprises nine intermediate products or components from the refinery. The 8.00 75.00 50.00 0.00 .90 91.10 117.p. both variables will be equal to zero.8692 0. In the proposed iterative approach.0 5.00 0.p. C.0 0.0 0.90 P3 0. with different specifications over an 8-day scheduling horizon.00 0. The Tables 1 and 2.00 100. The main goal penalty = (pltySH i Si. (17 ) whereas P8 and k P12 are computed in Eq.8187 0.t ≤ I pri.00 104.00 250.k Fi.00 0. initial.00 43.00 1. rate (Mbbl/day) 15.90 P7 35.10 0.00 100.t ) (34) is to maximize the total profit (see Eq.00 Prod.00 100.00 0.00 15. constant production rate.007 0.00 97.30 21.00 93.00 20. t P S I 9.70 351. (19)). (17 ) (gravimetrically predicted).00 0.00 52.50 86.40 P8 0. respectively.50 100.00 0.p.p.00 31.00 23.30 P5 16.00 22.90 36.00 0. Dedicated storage tanks with limited capacities for The penalty term (34) is directly proportional to the compo- components and products and three equivalent blend headers nent purchase cost: working in parallel are available in the refinery.k Fp.00 12. which can be blended in differ- A common source of infeasible solutions is the lack of a min.t + Dk.t defines the eight can be predicted by a linear volumetric average whereas amount of intermediate i needed in time slot t. ∀i.00 87.10 93.0063 0.00 79.p.00 100. + S+ − S− while the rest are linear.t ) (32) t p erties P1 and P2 are estimated through Eq. Fig. The performance of the proposed MILP-based approach was If property k for product p is between minimum and maximum tested with several real-world examples. The data are shown in specification values.30 P2 92. considering compo- t i nent cost. prop- penalty = (pltySk. intermediate products can be purchased at higher consideration for solving the blending problem. where the first cost from a third-party.50 100. ∀p.00 26.7460 0.60 0. Méndez et al.30 P11 0.00 20.0 5.00 100.00 100.p. in some way.k Fp.t .00 18. ent ways to satisfy multiple demands of three gasoline grades imum amount of intermediate required to satisfy either prede.00 Property P1 93.05 0.00 Initial stock (Mbbl) 48.20 0. fined component concentrations or certain market specifications.t ≤t tory constraints.057 P9 0.00 0.p.00 250.00 100. t (33) in Table 1.0 5. All the infor- relax minimum inventory constraints: mation about components such as cost.00 0. recipe limits and specifications are given in Table 2.90 0.00 0.00 15.p Dk.00 0.00 100.00 14.00 94.90 95.A.40 0.80 81.009 0.7069 0.98 65. Note that no inventory costs the stopping criterion of the iterative procedure should be that were considered. mum and maximum specification for property k. k.30 P10 0.k Fi.t + S i.3. P8 and P12 are non-linear.10 P6 94.00 22. requirements.00 0.00 30.00 20.00 20.00 88.p Dk.t − Dk. P2.0 5.30 34.30 93. which allows to the remainder is based on non-linear correlations.00 3. product values and different penalties for component It should be noted that for the case of infeasible specifications shortages and out-spec products.00 10. Table 1 Component data Component C1 C2 C3 C4 C5 C6 C7 C8 C9 Cost ($/bbl) 24.30 48.00 104.0002 0.20 93.p. properties P1.90 81.80 87.t .00 1. The continuous variable Si. Note that from the 12 product specifications corresponding objective penalty terms are shown in Eq.00 54.00 33.00 0.6540 0.90 91.t and Dk.t define a value els can be solved using the iterative procedure outlined in that. t (30) i − p.6731 0. rate.7460 0.00 1. Product data including price.00 6.30 4.1 61.0 Max stock (Mbbl) 100.00 24.10 0. minimum and maximum stocks and properties is shown I Vi.90 94. inven- p. Numerical results pri.30 95.t . Twelve key component and product properties are taken into In this case.10 95. represents the deviation from the mini.40 118. / Computers and Chemical Engineering 30 (2006) 614–634 625 deviation: the relative change in the bias parameter be less or equal than a specified tolerance. 4 for the simultaneous blending and scheduling operations.00 0.60 1.30 0.30 16.00 2. k. (32): shown in Table 2.6167 0.00 50.50 15.50 85.t ≥ propmax ∀p.1805 0.p.0 0.00 50.0 5.t = inii + prodi et − I Fi.00 100.30 0.10 91.30 63. (31) i S + S − The proposed discrete and continuous time MILP mod- where the continuous variables Dk.00 57.00 0.7339 P4 3.33 0.

temporal. component stocks and time are mined based on the product due dates.00 Rate (Mbbl/day) 5.00 25.00 P11 1.00 46.00 C6 0.00 60.00 25.00 10.00 150.00 0. which were null for all solutions generated.00 Product G1 (price ($/bbl) = 31. In this way only a pure blending problem is taken into consid- .00 60.00) MIN MAX MIN MAX MIN MAX Recipe (%) C1 0.72 0.00 10.00 0.00 C2 0.00 1. Preferred the discrete time representation is used.00 50.00 0.00 0. 9.00 23.00 23.00 25.00 Day 8 5. For the continuous time available as needed to produce 1 Mbbl of each product once. 4 and 6 have 1-day duration whereas intervals 2 and 5 have this particular problem.00 P10 18.00 C9 0.00 Specifications P1 95.00 71. lated for each one of the six subintervals defined by the product lyzing the strong interaction between blending and scheduling due dates.00 100.00 P12 2.00 85.775 0.00 20. one time slot with unknown duration is postu.00 50.00 P7 45.775 P4 20.00 42.00 48.00 0.00 71.626 C.00 0.00 0.00 25.00 45. For 3.00 10.00) G2 (price ($/bbl) = 31.00 42. where intervals 1.00 P5 46.00 0.00 0.008 P9 42.00) G2 (price ($/bbl) = 31. Example 1 (blending problem) ages were included in all cases.00 C4 0.00 50.00 C3 2.00 0.00 98.00 50.00 45.00 0. Méndez et al. ing and scheduling problems addressed in Examples 2–4.00 12.00 60.00 0. decisions.00 23. representation.00 0.A.00 45.015 0.00 150.00 5.00 Inventory (Mbbl) 5.00 0.00 10.00 50.00 50.00 0.00 5.00 45.00 24.00 0.00 18.015 0.00 P6 85.00 5. inventory and resource con- 2-day duration.00 P8 0.00 25.00 5.00 0. The main goal is to find the best or Examples 2–4 are solved using the proposed model with the ‘preferred’ recipe for each product that minimizes blend cost discrete and the continuous time domain representation.00 Day 4 5.00 24. assuming that enough resources.00 150.00 45.00 C8 0.00 50.00 0.00 0.00 0.00 P3 0.00 0.00 18.00 1.00 50.00 C5 0.00 30. the time discretization is deter.00 98.00 85.00 100.00) G3 (price ($/bbl) = 31.00) MIN MAX LIFT MIN MAX LIFT MIN MAX LIFT Requirement (Mbbl) Day 1 5.00 10.00 88.00 0.00 10.00 25.00 5.00 10. three products (G1–G3). When and simultaneously satisfies all quality specifications.00 5.00 5.00 5.72 0.72 0.70 2.00 10.00 50.00 10.70 2.00 71. Example 1 is only focused on the blending problem Example 1 deals with a single-period blending problem of and its solution is used as initial product recipes for the other.70 Four different examples were solved with the purpose of ana. slack variables for property deviations and intermediate short.00 C7 0.00 6.00 Day 3 5.00 25.00 22.00 88. In order to make a direct comparison with the straints coming from the scheduling problem are disregarded by continuous time formulation.00 46. In order to guarantee that feasible solutions are found.00) G3 (price ($/bbl) = 31.00 P2 85.00 90.00 22.00 22. / Computers and Chemical Engineering 30 (2006) 614–634 Table 2 Product data Product G1 (price ($/bbl) = 31.00 0.00 20. the scheduling horizon recipes are proposed as the initial blends for the integrated blend- is divided into six consecutive time intervals.00 Day 5 Day 7 5.00 100.00 0.00 45.775 0.1.

0152 0. Convergence to preferred recipe for product G1 (iterative procedure).418 35. / Computers and Chemical Engineering 30 (2006) 614–634 627 Table 3 Iterative blending problem for product G1 Quality Min.00 P8 0.00 60.30) cost ($/bbl) = 29. 6.00 1.00 96.00 60.00 88. In Example 2 preferred product recipes found in Example posed procedure.2. initial product recipes were generated taking into account only linear product properties.470 88. cost ($/bbl) = 29.5488 1. the continuous time repre- terms of component concentration is presented in detail in Fig. since the maximum MILP-based procedure.70 eration.210).00 P10 14. best product recipes and ‘bias’ factors for all products continuous variables and its solution was found in 0.0378 0. Spec. Example 2 (blending and scheduling with limited performed to update the initial recipes with the purpose of sat. component concentrations for each product tion tools. Initial recipe (blend Iteration 1 (blend Iteration 2 (blend Max Spec. .00 60. Despite using linear approximations.775 P4 20.898 97. However.8928 P2 85.00 P6 85.893 97. Component costs and properties.00 1. The recipe evolution for product G1 in cost and product value ($ 1. which gradually representations obtained the same profit in terms of component increased the blend cost.5684 2.80 50. ties tend to improve when the number of iterations is increased. case. In addition to the exact values for each prop- considered for solving Example 1.974 16.5684 1.409 50.455 35.1223 1.13 s.923 42.458 22.99) Value Value Approx Value Approx P1 95.00 60. Preferred recipes for products G2 and G3 were found by executing just one iteration of the pro.611. In order to gener. Méndez et al.00 P12 1.438 88.00 P11 0. The proposed LP-based predicted by the proposed linear functions are also presented in iterative procedure was used to find preferred recipes for all Table 3.7324 0. variable recipe limits Blend cost and product properties associated to each recipe are and stringent product specifications are the central features to be shown in Table 3.7325 0.00 50. although the discrete and continuous time were updated by the LP model in each iteration.00 P5 46.0493 88.A. Then the iterative procedure was 9.833 71.0150 0. production) isfying all product specifications.891 97.417 88.97) cost ($/bbl) = 29. It should be noted that predictions of non-linear proper- required products.460 91. C. As reported in Table 13 in Section 11 on com.780 P7 45.0150 0. sentation is able to find a schedule that operates the blenders at Fig.00 46.797 91. iteration the same solution generated by non-linear optimiza- ate feasible recipes.7737 97. the approximations scheduling decisions are made a priori.00 34.8964 1. 6. the problem involves 81 constraints and 127 Finally.4335 P3 0. where it is assumed that all erty predicted by non-linear correlations.0150 0. whereas an additional iteration was needed to 1 were used as the initial solution for the proposed iterative satisfy all specifications for product G1.256 15.015 P9 28.72 0.974 22.005 18. specification for property P8 was violated both in the initial the proposed MILP model was capable of finding in just one recipe as in the first iteration (see Table 3). putational results.7418 0.00 97. In this are reported in Table 4.5687 1.

00 0.00 40.00 T6 7.00 0. Both models tation.00 T2 1.00 60.00 G3 T1 0.00 0.02 T6 7. As shown in Table 13.00 22.00 45.00 0.00 30.00 5.00 90.00 1.00 40.00 2.00 8.00 4.00 50.00 38.3734 P5 50.00 0.259 25.00 45.00 25.00 25.00 50.00 15.99) G2 (blend cost ($/bbl) = 25.00 T3 3.893 (bias = 1.794 1.00 9.00 0.00 50.00 90.00 0.00 45.00 4.00 12.00 55.00 40.00 8.00 T6 7.2214 (bias = 1.4561 1.527) 98.00 63.0184 97.00 50. 9 In this way the amount to be produced in each time interval Table 5 Table 6 Product schedule (Example 2—discrete time representation) Product schedule (Example 2—continuous time representation) Product Period Start End Prod Lift Inventory Product Period Start End Prod Lift Inventory G1 T1 0.00 C3 2.00 3.00 2.00 50.00 5.98) Recipe (%) C1 22.00 5.628 C.00 55.00 T5 5.00 7. which can significantly reduce the total operating cost.00 7.00 5.958 C9 9.00 0.923 21.00 0.00 8.00 25.00 90.00 T4 4. binary variables.4594 (bias = −1.00 0.3310 (bias = −1.00 0.00 5.00 0.00 0.00 50.00 0.0861) P3 0. / Computers and Chemical Engineering 30 (2006) 614–634 Table 4 Preferred product recipes Product G1 (blend cost ($/bbl) = 29. and 841 continuous variables.00 50.00 40. 9 binary variables.02 T3 3.33 0.00 75.00 5. and 757 continuous variables.02 T5 5.00 0.00 .005 17.00 10.00 23.00 45. were solved in 0.26 s.00 T5 5.00 T2 1.00 0.02 T4 4.0150 P7 60.28) G3 (blend cost ($/bbl) = 24.00 90.00 0.409 41.33 15.7324 0.00 4.00 25.00 0.00 10.00 P12 1.00 T4 4.00 1.2363 18.00 4.00 1.00 40.00 10.00 16. 7.00 1.6966 P10 16.00 T2 1.00 40.00 T4 4.00 0.00 7.00 T3 3.833 54.02 G2 T1 0.00 1.00 16. Gantt 9.5208) P2 88.00 0.00 23.00 G2 T1 0.00 38.00 T5 5.00 0.00 0.4122 (bias = 1.00 68.00 63.00 0.2597 full capacity for 2.00 G3 T1 0.00 23.00 0.00 5.00 0.A. respectively.5932 54.00 12.00 4.00 25.00 5.992 C7 5.198 C8 0.0072 P9 22. minimum and maximum requirements for each time interval.00 8.00 90.00 50.00 25.00 3.02 T2 1.00 90.00 0.00 T6 7.00 1.00 0.997 Quality P1 97.00 T5 5.00 0.438 (bias = −0.00 1.00 0. Example 3 (blending and scheduling with ﬂexible charts and inventory evolution of components for both discrete production) and continuous time representations are shown in Fig.00 40.00 C6 10.00 P11 1.00 90.00 5.00 T2 1.02 T6 7.00 40.780 97.00 5.636 C5 25.00 T2 1.00 T3 3.00 45.00 0.00 T4 4.00 45.5475 P6 91.00 1. Méndez et al.00 40.00 68.00 0.00 5.00 T4 4.00 25.00 50.3.00 75.659) 88.3410 42.00 40.00 4.00 35.0439) 88.7289 P4 35.00 10.00 40.00 40.00 25.02 G1 T1 0.67 days less than the discrete time represen. The continuous time formulation comprises 832 constraints.00 64.00 50.00 90.02 10.0150 0.00 T3 3. Product schedules based on a discrete and continuous time rep- resentation are reported in Tables 5 and 6.00 8.00 0.00 2.00 22.33 0.00 T6 7.00 23.00 8.6536 21. the discrete time formulation involves 679 This example evaluates in Example 2 the effect of predefining constraints.00 T3 3.847 25.7305 0.2465 P8 0.372 C4 4.00 35.00 7.00 0.0079 0.00 T5 5.00 3.02 0.5611) 98.00 C2 20.947 24.00 10.5687 1.

00 50.00 22.00 150.00 50.00 1. Gantt charts and evolution of component stocks (Example 2): (a) discrete time.00 8.00 0.00 0.00 5.00 50. and 841 continuous variables and its solution T4 4.00 40.00 70.00 12. T5 5. (d) continuous time.00 35. Preferred product recipes are Product schedule (Example 3—discrete time representation) used for all products and one iteration is only executed.26 s (see Table 13).00 8.00 40.00 1.00 0.00 23.e. from 400. which repre- sents increasing the total profit to $ 2.00 7.00 limited production) T3 3.00 50. this example deals with a modified version of the T6 7.00 0.00 60.00 10.00 T4 4.448.60 days T2 1.00 75.00 0.050. 18 T3 3.00 3. Example 4 (full re-blending and re-scheduling with T2 1.00 25.00 30. (1) properties P1 and P2 are decreased by one for components imum production rates.00 72.00 50.00 95.00 7.00 757 continuous variables and its solution was found in 0.00 G3 T1 0.00 0. which is almost Table 7 a 52% increase (see Table 12).00 T5 5.00 9.00 4. This (3) component cost is increased to 27 and 23 $/bbl for C1 and modification allows the model to increase the total production C2 and (4) production rates for C1 and C2 are reduced to 13 by almost 36%.00 3.00 0.00 Finally.02 less of total operating time to reach the same production level T3 3.4.00 1. Product schedules based on a discrete and continuous time representa.02 lution of component stock for Example 3. The amount of product to be lifted at C1–C3 and C6. T2 1.00 40.00 25.00 0.00 38.A. / Computers and Chemical Engineering 30 (2006) 614–634 629 Fig.00 122.00 150.00 we note that the continuous time representation needs 2.00 4.02 as the discrete time model.00 45.00 112. C. T5 5.02 to 542.00 112.00 112.00 4.00 T6 7.05 $/bbl. Méndez et al.00 100.00 70. 18 binary variables. and G2 T1 0. respectively. 7.00 0. specific due dates is still a hard constraint to be satisfied.00 0.00 0.02 formulation comprises 679 constraints.00 binary variables. Fig.00 0.23 s. 8 shows Gantt-charts and evo.00 8.00 0.00 10.00 40.00 0. (c) discrete time.00 was generated in 0.00 10.00 The continuous time formulation comprises 832 constraints. Product Period Start End Prod Lift Inventory tion are shown in Tables 7 and 8.02 0. The discrete time T6 7.00 7. (b) continuous time.00 13. i.00 5. becomes a model variable only restricted by minimum and max.00 60.00 45.02 T4 4. (2) the price of G3 is increased to 31.02 Mbbl. In this example G1 T1 0.00 3.00 0.00 5.00 10.00 original Example 2 where the following changes are introduced: .

80 40.00 4.00 T5 6.00 0.00 7.00 50.00 3.00 0.00 0.00 4.00 12.00 1.00 10.00 3. (d) continuous time.00 150.00 38.00 10. infeasible because of the modifications introduced.00 T4 4.00 8.00 T4 4.00 90.00 60.00 15.00 0.00 45.00 8.00 0. Gantt charts and evolution of component stocks (Example 3): (a) discrete time.00 9.00 5.00 4.00 .00 4.00 140.00 40.00 50.00 T6 7.00 50.00 T3 3.00 0.02 75.00 30. However.00 35.00 0.00 0.00 50.00 40. they these changes in the blending and scheduling decisions.00 0.00 12.00 113.00 4.A.00 T4 4.00 T2 2.00 10.80 40. The main goal here is to analyze the effect of Example 1 are proposed as the initial solution.00 50.00 20.00 T3 3.00 4. Detailed have to be updated because some preferred recipes become product schedules for discrete and continuous time representa.00 1.00 37.00 0.00 25.00 65.00 100.00 15.02 T6 7.00 4.00 5.00 68. Only one tions for Example 4 are shown in Tables 9 and 10.00 0. / Computers and Chemical Engineering 30 (2006) 614–634 Fig.00 32.80 3.00 50.80 3.00 140.00 38. Méndez et al.00 50.00 140.00 10.00 10.00 1.00 7. and 31 Mbbl/day respectively.00 T2 1.00 0.00 35.00 10.00 G3 T1 0.80 4.00 T2 2.00 1.00 60. (c) discrete time.00 47.80 7.00 40.00 32.00 140.00 50.00 22.00 10.00 T6 7.00 50.00 75.00 T4 4.00 8.00 T5 5.00 T6 7.00 T3 3.00 10.00 10.00 T5 5.00 23.00 130.00 T4 4.80 7.00 T2 1.00 T4 4.00 35.00 5.00 0.00 T5 5.00 T3 3. All other data remain as in the Regarding the blending decisions.00 50.00 T2 1.00 0.00 1.00 3.00 68.00 T3 3.00 10.00 0.00 T6 7.00 40.00 G1 T1 0.00 T5 6.00 8.00 5.00 40.00 T5 6. 8. respectively.00 44.00 45.00 150.00 60.80 3.00 8. (b) continuous time.00 10.00 0.80 7.00 4.00 100.00 5.00 5.00 4.00 140. preferred recipes found in original example.00 T3 3.00 0.00 35.00 40.00 0.00 45.00 90. iteration is required to modify the infeasible recipes related to Table 8 Table 9 Product schedule (Example 3—continuous time representation) Product schedule (Example 4—discrete time representation) Product Period Start End Prod Lift Inventory Product Period Start End Prod Lift Inventory G1 T1 0.00 7.00 60.00 64.00 23.00 8.00 G3 T1 0.00 10.00 25.00 10.00 T6 7.00 9.00 10.00 1.00 G2 T1 0.630 C.00 35.00 0.00 38.00 T2 2.00 40.00 G2 T1 0.

00 16.00 68.00 10.269 C6 9.00 26. / Computers and Chemical Engineering 30 (2006) 614–634 631 Table 10 satisfied with a very small margin. For this reason every single problem needs no more than 1 s of CPU time with CPLEX 8.00 64.0079 0.00 while Table 13 provides the computational statistics on the four T4 4.831 25.00 10.223 1.00 T5 5.00 40.00 12.6966 21.00 50.7309 0.00 50.00 10. lations usually have a better computational performance when ing cost with which the profit is reduced from $ 2.00 140.2597 1. significantly reduce the operating time of the available equip- It should be noted that key properties such as P1 and P2 are ment.7305 0.00 32.0184 97.00 40.00 6.015 97.052 P2 87.00 4. T2 1.00 C3 16.00 not very large and involves a modest number of 0–1 variables.00 ments were relaxed (Example 3) and certain changes in compo- G3 T1 0.00 Different blending and scheduling problems were solved in T6 7.00 140.00 60.00 0. it was observed that discrete time formu- time.6546 P10 17.00 0.00 T6 7.00 the original Example 2 where minimum and maximum require- T6 7. which highlights the computational efficiency of the proposed models and the iterative MILP proce- products G2 and G3.7294 88.00 8.28) ($/bbl) = 26.00 0.00 24.00 1.6837 21.00 10.050 compared to continuous models.015 97.00 4.2362 16.3734 41. As shown.636 14.00 10. updated component concentrations increase the blend.640 Quality P1 97.00 60. In addition.00 the previous section in order to evaluate the efficiency of the pro- G2 T1 0.3626 . Example 1 dealt with a pure blending problem T2 1.448.00 3.00 5.00 0.0133 87.0074 P9 21.20 0.09 16.00 40. C.992 8.20 9.831 42.00 10. Examples 3 and 4 correspond to modified versions of T5 5.00 6.259 24.968 18.00 3.00 1.67) Recipe (%) C1 25.00 5.3408 40.234.00 0.5476 54.00 50. continuous to 1.8204 98.00 1.9938 1.1.7289 0.00 5.00 140.0072 0.00 113.869 C5 9.00 T3 3.00 0.00 T4 4.00 0.2473 68.20 0.00 35.571 54.00 8. formulations are able to generate more flexible schedules that nent costs were increased and octane numbers were reduced.00 5.00 26. This difference mainly arises because compo.00 P11 1.00 C2 23.00 40.00 35.00 3.00 decisions. Table 12 summarizes the results for Examples 2–4.A.00 5.00 35.00 24.00 nent properties and cost and product prices were incorporated T2 1.8588 88.195 C4 25.00 Table 13. Product Period Start End Prod Lift Inventory Computational requirements for this example are summarized in G1 T1 0. Original preferred and updated recipes dure.856 7.00 25.00 5.00 18.00 8. On the other hand.5936 54. T3 3.00 40.00 4.00 10.00 examples.00 (Example 4).00 24.5491 1.00 100.015 P7 60.6533 21.00 140.98) ($/bbl) = 26.00 60.00 60.00 5.00 0. Méndez et al.0235 97. which means that qual- Product schedule (Example 4—continuous time representation) ity giveaway is also minimized through the proposed method.00 whereas Examples 2–4 also accounted for optimal scheduling T3 3.00 38.00 10.372 3.00 0. the to satisfy all product specifications in all the examples.00 25.6305 P6 97.00 37.9799 P12 1.0081 0. As a new recipes satisfy all product specifications but at the same general characteristic.92) cost ($/bbl) = 24.00 posed method. Table 11 Updated product recipes (Example 4) Product G2 G3 Preferred (blend Updated (blend cost Preferred (blend Updated (blend cost cost ($/bbl) = 25.00 10. As can be seen. the size of the MILP problems is T5 5.00 16.00 45. Computational results T4 4.490.0455 P3 0.00 50.00 5.794 0.4562 1.00 75.00 25.7285 P4 41.1274 P8 0.00 60.947 24. a very small number of iterations were required for these products are compared in Table 11.6283 98.9724 P5 54.

13 1 MINOS 0 Execution errorc – 1 CONOPT 0 Infeasible solutiond – 1 MINOS 10 80. In order to examine the solution of the scheduling and blend. Local that the proposed method is significantly faster than the NLP solutions are obtained with MINOS. vars. CPU time Iterations were solved considering different initial values for these key constraints problem variables.21 1. be solved in some cases.251 0. First. 832 0.251 0.6 1 MINOS Maximum allowed Execution errorc – 1 CONOPT Maximum allowed Execution errorc – 2 Proposed approach – 1611.9 2 BARON Minimum allowed 1611. Also. d Infeasible solutions suggest that convergence problems may arise from non-convexities of non-linear functions.4 2 DICOPT/CONOPT 10 1611.26 2 DICOPT/MINOS 0 Execution errorc – 2 DICOPT/CONOPT 0 1611.3 2 BARON 10 1611. An additional problem was the execution a errors that arose from evaluation errors in the non-linear func- Seconds on Pentium IV PC with CPLEX 8.26a 1 important features. 832 0.26a 1 4 (discrete) 9. in several cases the non- 3 (continuous) 18. Although these problems can partially Examples 1 and 2. Consideration of the original non-linear cor. 841. However. and DICOPT. Usually. the required changes may compromise relations gives rise to NLP and MINLP models. These errors are generated because the original non-linear correlations may be not defined for the entire domain of the vari- ables or some possible combinations of feasible values. examples Example Binary vars.13a. b For specified recipe. stock production (M$) Comp.21 ∼1000 2 DICOPT/MINOS Maximum allowed Execution errorc – 2 DICOPT/CONOPT Maximum allowed Execution errorc – 2 BARON Maximum allowed Execution errorc – a Seconds on Pentium IV PC with GAMS 21.61 22352 11562.21 1. 757. 757.23a 1 linear models failed to converge due to the non-convexities in 4 (continuous) 9. / Computers and Chemical Engineering 30 (2006) 614–634 Table 12 Summary of results Example Blend value (M$) Comp. Table 14 presents the computational results for number to a real power.2. which were the optimality of the solution. inventory build (M$) Total profit (M$) Profit per barrel ($/bbl) 2 12400. from Table 14 it is clear that the Table 14 Comparison with non-linear codes Example Solver Initial component concentration (%) Objective function CPU timea 1 Proposed approach – 80.05 4.21 ∼1000 2 DICOPT/MINOS 10 1611. and particularly. Thus.21 4. c Execution errors arise from evaluation errors in the non-linear functions.21 1. non-linear solvers have difficulty recovering after attempting an ing problem addressed in this paper using directly the non-linear undefined operation such as dividing by zero or raising a negative correlations.251 0. it is worth mentioning that both the lin- 2 (continuous) 9. 841.26a 1 ear and the non-linear models were able to find the same optimal 3 (discrete) 18.23a 1 solution for these examples.26a 1 the non-linear model.b 2 The analysis of the results reported in Table 14 reveals some 2 (discrete) 9.61 22352 7997. 679 0. 1 –. 757.49 3.21 ∼1000 2 DICOPT/MINOS Minimum allowed Execution errorc – 2 DICOPT/CONOPT Minimum allowed 1611.2. tions. 127.6 1611. Since solvers in Example 2. it can be seen in Table 14 solved by several non-linear general-purpose optimizers.25 Table 13 the component concentrations in each product may have a signif- Model size and computational requirements icant influence on the non-linear model performance. 832 0.52 4 11785 23504 12953. 679 0. b All scheduling decisions are predefined.1 in GAMS 21.5 2 BARON 0 1611.632 C.4 1 CONOPT 10 80. compared to the MINLP whereas global solutions can be obtained with BARON.251 0. 81 0.03 3 16802. cont. .3 1 MINOS Minimum allowedb Execution errorc – 1 CONOPT Minimum allowed 80.44 2448. 679 0.A.21 1. 841. CONOPT. solvers in Example 1.21 0. Méndez et al.49 1234.

Linearization strategies for a class of zero- assumes fixed values of δp.j + µj δj − ρj = 0 (A. prob- short-term scheduling problem in oil-refinery. where in the above KLIN and KNL represent the linear and non- I /F P ) represents linear properties. sufficient con- ditions for the convergence of the successive LP procedure to a Proof. respectively.A. W.k (Fi.1)–(A. Conclusions Case 1.k < prmax p. one mixed integer programming problems. PRp. it trivially follows that a solution to this problem is equiv- that can be used either on discrete or continuous time formula. convergence of the successive LP procedure to a local minimum of problem (P).k = max 0.k = 0 and (A. / Computers and Chemical Engineering 30 (2006) 614–634 633 proposed successive linear approach is fast and robust. i ∈ I2 (A.k < prmax p. The method is lem (P) reduces to the problem given by (A.k at the local minimum of problem (P). j Acknowledgements s. δp. ∀k ∈ KLIN (A.4) i ∈ I1 i ∈ I2 i Since δj can be set equal to gj (x* ) at the optimum solution x* . Several examples representative of real world problems Case 2.p (A.4) is redundant. µi .p ≤ prmax P p.8) j The authors would like to thank ABB Corporate Research and the Center for Advanced Process Decision-making (CAPD) ai.6) guaranteed if the non-linear properties are not a strong function of the compositions. and PRp. H.5) FpP Note that from (A.3) condition of the Karush–Kuhn–Tucker condition yields: i cj + λi ai. p i p When the successive LP procedure is used (A. P P ∀p.k Fi. ∀p. i ∈ I2 (A.k ) = 0. j xj ≥ 0 (A. The equations in (1)–(6) with a cost objective function and i ∈ I1 i ∈ I2 assuming only upper bound for the specifications can be written as follows: ∗ ∂gj + µj gj (x ) + x j − ρj = 0 (A. ai. An integrated MILP-based approach has been proposed to simultaneously optimize the gasoline off-line blending and the Proof.2) ai.p I /F P . but faster and more reliably.3) and (A. alent to the solution of problem (P). Also. Méndez et al. Operations Research.9) is replaced by s. On the convergence of the successive LP method for blending operations where we assume the function gi (x) > 0 represents (A. Since in this case δp. . The Karush–Kuhn–Tucker conditions of (PC) yield the fol- For simplicity we consider for the blending problem the case lowing stationary condition: of a single period problem and hence we drop the subscript t for cj + λi ai. recipes through a successive LP or MILP iterative procedure Hence. I Fi. i ∈ I1 (A. 11. and very The two following cases provide a sufficient condition to the useful for addressing real-world cases.p ≤ prmax p. note that the successive LP method p Adams. ∀p. FpP = I Fi. able to deal with non-linear product properties and variable which is the LP solved at the first iteration of the procedure. For this case the stationary I pri.k = Fi.6). as form as: well as numerical comparisons with NLP and MINLP solvers (PC): showing that the proposed method converged to the same solu.p p References the evaluation of the non-linear property at the volume frac- tion νi. Also.k Fp .12) j i where δi is treated as a constant.7) tions.p this implies that (A. (1990). respectively.13) I pri.j time. tions. ∀p (A.1) where λi . ∀k ∈ KNL (A.j xj + δi xi ≤ di . FpP ≥ 0 (A. The non-linear property PRp.p .k Fi.k . ρi are the multipliers of (A.11) and (A.t. & Sherali.k Fp + δp.5) case 2 means that convergence to a local solution of the non-linear programming problem (P) can be I Fi.9) at Carnegie Mellon for financial support of this work.j xj + gi (x)xi ≤ di .k )/(δvi.11) (P): ∂xj max αp FpP − I βi Fi. For convenience we represent problem (P) in compact local solution of the blending problem have been presented.k at the local were presented to illustrate the flexibility and efficiency of the minimum of problem (P) and (δPRp. min cj xj (A.k Fp .j + µi ai. C.j + µi ai. The non-linear property PRp.p .10)..8)–(A.j xj ≤ bi . 38.k at each successive iteration.10) Appendix A.t.12) are identical if (δgj )/(δxj ) = 0. 217. ∀k ∈ KNL (A.5). proposed models and solution technique.

Institute. I. 433.. S109. & Tong. L. 64.. (1998). I. H.. & Pinto. / Computers and Chemical Engineering 30 (2006) 614–634 Bechtel Corp. (2003). D. Hua. M. & Lin. Texaco blending systems—From Omega to StarBlend. E. A planning model approaches for scheduling of chemical processes: a review. C. Floudas. scheduling. linear programming model for refinery short-term scheduling of crude Booner & Moore.. 43–50. (2004). (1979). 825. operations. & Kettani. & Mann. L. AIChE Symposium Series No. S. Méndez et al. E. Lasdon. (1992). S.. 627. F. Houston. G. L. 28. A. X. (2001). Scheduling crude-oil An application with multi-million dollar benefits: Part II. neering Progress. (2003a). oil unloading with inventory management. B. I. 6723. AIChE Symposium Series No 320 (pp..0. Crude-oil blend scheduling optimization: Wenkai. Planning and scheduling in the process industry. J. p. J. & Waren. 1219. TX.. Houston. Mixed-integer linear programming for process networks with multicomponents flows. L. 75). C. M. Z. Grossmann. J. Crude-oil blend scheduling optimization: Shah. Global optimization of bilinear Jia. Discrete for the scheduling of chemical processes. B.. Short-term scheduling and recipe opti. Chemical Engi. 11. . Computers and Chemical Engineering. L. Van den Heever. I. 326 (pp. I. Pinto. C. 22. Rigby. Z. 42. 41. 98. Chemistry Research.. A. Computers and Chemical Engineering. E. & Mann. Glismann.and multisite planning and scheduling: Current sta- An application with multi-million dollar benefits: Part I. J. G. A. Assignment and sequencing models Grossmann. Hydrocarbon tus and future challenges. Mixed-integer manual. (1992). Research.. S1039. D. V.. 2109. Industrial Engineering and Chemistry Processing. The evolution of Kelly. and Chemical Engineering. Continuous-time versus discrete-time Moro.. (1998). Springer-Verlag. (1996). & Grossmann.... J. Oral. 72–79. Hui. K. Interfaces. Computers and Chemical gasoline blending and distribution scheduling. & Park. E.. Annals of Operation Research. optimization methods and their role in the integration of planning and 81. N. & Harjunkoski. Pinto. TX. I. 40. 25.634 C. (1995). User’s Lee. S. cubic mixed-integer problems. 19. 94. Reklaitis. Processing. (2003). (2002). PIMS (Process Industry Modeling System). O. A linearization procedure for quadratic and mization of blending processes. (1993).A. A system description. 150). & Iearapetritou. Industrial Engineering and Engineering. J. Chemistry Research. (1995). (2004). J. Hydrocarbon unloading. (2003b). Antalya. L. & Gruhn. Zanin. Single. 47–53. & Grossmann. J.. M. M. J. ISBN 3-540-00222-7. (2002). 1630. Industrial Engineering and chemical Modeling System). A.. Kelly. version 6. Quesada. Technical Report. 35. Computers for refinery diesel production. Kelly. RPMS (Refinery and Petro.. Formulating production planning models.-W. Turkey: NATO Advanced Study In Advanced planning and scheduling solutions in process industry.. M. D. Management Science. Operations Research. storage and processing. (1998). Overview of scheduling and planning of batch process Kallrath. D. C. 25.

- Using Combinations of Heuristics to Schedule Activities of Constrained Multiple Resource ProjectsUploaded byJohn McIntosh
- Cost Optimization bUploaded byocmain
- programacion lineal N 1.pdfUploaded byAlbino Quispe M
- m1010 p2Uploaded byapi-242082851
- Software Project Scheduling using Genetic AlgorithmsUploaded byjfrchicanog
- Product Line and Scheduling at Intel_Kempf_Wagner PresentationUploaded byparul_15_12
- 990408Uploaded bysaisenth
- MSSO_A_New_ToolUploaded byBrandy Gonzalez
- Lecture 4Uploaded bynoorhilla
- 2014_Strategic Mine Planning and OptimizationUploaded byluisparedes
- Concept of Operations Scheduling - Wba - July -Dec 2016Uploaded bymutiso
- Minemax Scheduler White PaperUploaded byWelsinsin Kevin Sin
- math 1010 projectUploaded byapi-240285476
- The Tactical Berth Allocation Problem With Quay Crane Assignment and TranUploaded byEdwina Seviera
- A Review of Analytical Models, Approaches and Decision Support Tools in Project Monitoring and ControlUploaded bystriker190
- Master Data for SCM 2Uploaded byManivannanc
- Baseline Scheduling Front and BackUploaded bysarwar_dba
- optimisationUploaded byShubhamAgarwal
- Interactive Fuzzy Goal Programming approach for Tri-Level Linear Programming ProblemsUploaded byAnonymous 7VPPkWS8O
- Chapter 9Uploaded byTalha Bin Saeed
- LINEAR PROGRAMMING.pptxUploaded byKwin Hazel Satiada
- Ch 09 Aggregate Planning and MpsUploaded byAmrik Singh
- OT I Quiz 2Uploaded byRobertBellarmine
- TheveninUploaded byMateus Corato Zanarella
- Lecture 1 - Intro to Math OptimisationUploaded byKuknesvary Puniamurthy
- 1-s2.0-S088677981530225X-mainUploaded byMarkusMakuAldo
- fgoalattUploaded byiman145
- 1532-0545-2006-07-01-0088Uploaded byNP
- Attributes_of_a_SchedulerUploaded byDadeida Jimenez
- AOR-01_EUploaded byDevKumar

- ZZ_1237386579_BB0408121Crude_Oil_BlendingUploaded byHussaini Hamisu
- LeachingUploaded byHussaini Hamisu
- A Stochastic HMM-based Forecasting Model for Fuzzy Time SeriesUploaded byHussaini Hamisu
- MODEL PREDICTIVE CONTROL OF A CRUDE DISTILLATION UNIT.pdfUploaded byHussaini Hamisu
- Body Size and Breast Cancer Risk in Black Women and White Women- The Carolina Breast Cancer Study.Uploaded byHussaini Hamisu
- A comparison of fuzzy forecasting and Markov modeling.pdfUploaded byHussaini Hamisu
- A Multivariate Heuristic Model for Fuzzy Time Series ForecastingUploaded byHussaini Hamisu
- 2013-4M3-DryingUploaded byHussaini Hamisu
- Fuzzy Time Series Forecasting With a Probabilistic Smoothing Hidden Markov ModelUploaded byHussaini Hamisu
- Advanced Breast Cancer in NigeriaUploaded byHussaini Hamisu
- Modelling of Glucose InsulinUploaded byHussaini Hamisu
- Multiscale modelling of biorefineriesUploaded byHussaini Hamisu
- A crude distillation unit model suitable for online applications.pdfUploaded byHussaini Hamisu
- Modeling Glucose-Insulin Kinetics and Development of Type 2 DiabeUploaded byHussaini Hamisu
- A Comparison of Fuzzy Forecasting and Markov ModelingUploaded byHussaini Hamisu
- A Chemical Process SimulationUploaded byHussaini Hamisu
- Control Engineering and Therapeutic Drug DeliveryUploaded byHussaini Hamisu
- Model Based Controllers for Blood Glucose Regulation in Type i Diabetics Final ThesisUploaded byHussaini Hamisu
- A Collection Of Ten Numerical Problems In Chemical Engineering.pdfUploaded byHussaini Hamisu
- Ref no. 2010.3 Pasant Suwanapal.pdfUploaded byHussaini Hamisu
- Modeling glucose-insulin kinetics and development of type 2 diabe.pdfUploaded byHussaini Hamisu