E

conomic and operability benefits associated with better crude

oil blend scheduling are numerous and significant. The various
crude oils that arrive at a refinery to be processed into the dif-
ferent refined products must be carefully handled and mixed before
they are charged to the atmospheric and vacuum distillation units or
pipestill. Many details are involved in optimizing scheduling of a
refinerys crude oil feedstocks from receipt to charging of the pipestills.
Every refinery that charges a mix of crude oils to a pipestill has a
crude oil blend scheduling optimization opportunity. Producing
and updating a schedule of when, where, what and how much crude
oil to blend can be difficult. Although crude oils are often planned,
purchased, procured and have a delivery schedule set long before
they arrive at the refinery, details of scheduling crude oil off-load-
ing, storing, blending and charging to meet pipestill feed quantity and
quality specifications must always be prepared based on current
information and very short-term anticipated requirements. How-
ever, the rewards for performing better crude oil blend scheduling opti-
mization can be substantial depending on complexity and uncer-
tainty of the particular crude oil blendshop** operation.
Crude oil blendshop scheduling has historically been carried out
using fit-for-purpose spreadsheets and simulators with most deci-
sions made manually. The user attempts to create a feasible schedule
that meets flowrate and inventory bounds, operating practices and
quality targets over a near-term horizon, typically 1 to 10 days (up to
30). Actual schedule length depends on the reliability or certainty
of the crude oil delivery quantities and timing, and the pipestill pro-
duction mode runs. Farther into the future, the input data become
more uncertain and the scheduling work is typically cut off.
The goal of production scheduling optimization is to automate
many of these manual decisions by taking advantage of recent
advances in computer power and mathematical programming codes
and solving techniques. The main advantages of this approach are that
many thousands of scheduling scenarios can be evaluated as part of
the optimization in comparison to perhaps only one schedule found
by a user, a substantial reduction in time required to generate better
schedules and the ability to incrementally rerun the optimization
when different what-if scenarios are required (i.e., evaluating dis-
tressed crude oil cargos).
The focus of this article is four-fold. First, we delineate the busi-
ness problem of crude oil blend scheduling using a simple but reveal-
ing motivating example. Second, we highlight the hard and soft ben-
efit areas of improved blend scheduling to whet the appetite for
the impending details. Third, we provide a description of the new
scheduling approach for crude oil blending including the theory,
explicit problem formulation and related aspects such as how to seg-
regate crude oils when there are not enough tanks for dedicated stor-
age. And fourth, we discuss key elements of the scheduling solution
to enlighten the reader on the nuances and the challenges of solv-
ing large combinatorial problems.
Before further discussion, it is important to highlight the differ-
ence between production planning and scheduling and to discuss
the underlying need for continuous improvement of the schedul-
ing function. There will always be a planning activity and a schedul-
ing activity. Together, they form a hierarchical decision-making
framework that is very much a part of the organizational structure of
every corporation.
Planning is forecast driven and typically aggregates resources such
as equipment, materials and time to model and solve the breadth of
the problem. Planning generates simplified activities or tasks con-
sistent with these aggregations. Scheduling is order driven and uses the
decomposed equipment, materials and time to model and solve the
depth of the problem. It should be appreciated that a great number
of planning decisions are made long before any scheduling decisions
are generated. This implies that good scheduling can only result
from good planning.
The subtlety between forecasts and orders must also be appreci-
ated. At the planning stage firm and reliable customer orders are
rarely available over any significant planning time horizon except
Crude oil blend scheduling
optimization: an application with
multimillion dollar benefits
The abilit y t o schedule t he crude oil blendshop more ef f ect ively
provides subst ant ial downst ream benef it s
J. D. KELLY* and J. L. MANN, Honeywell Industry Solutions, Toronto, Ontario, Canada
June 2003, pgs 4753 and July 2003, pgs 7279
Used wit h permission.
www.HydrocarbonProcessing.com
Reprint ed f rom:
PROCESS/ PLANT OPTI M I ZATI ON
SPECIALREPORT
HYDROCARBON PROCESSING JUNE/JULY 2003
* Aut hor t o w hom al l correspondence shoul d be addressed.
* * The t erm bl endshop i s used t o descri be t he net work of equi pment and pi pi ng f ound i n a ref i nery t o
support handl i ng and bl endi ng of crude oi l s or ref i ned product s such as gasol i nes and di st i l l at es.
for wholesale agreements or contracts. Forecast or best-guess aggre-
gate demands and capabilities are used to optimize projections of
plant operations. Thus, the plans generated are used to set direc-
tions and not production orders. Scheduling on the other hand is
primarily based on orders. Orders are much more concrete both in
quantity and time and have the highest reliability for the immediate
future. Scheduling generates detailed tasks and activities to meet the
immediate orders and scheduling is typically updated whenever sig-
nificant changes to the order or plant capabilities occur.
There is always a requirement to continuously improve schedul-
ing utility. The underlying driving force for this is related to the
notion of innovation in industry. Three known innovations are
outlined by Norman et al.
1
The first is the manufacture of replace-
able or interchangeable parts that comprise the bill-of-materials of any
given product. The second is to produce many products within a
single facility and is sometimes referred to as product diversifica-
tion. And the third innovation, the one being implemented today
in industry, is production for final demand or demand-driven pro-
duction (DDP). This is the prevailing concept that we should only
produce product that satisfies actual product demand (demand
orders). Speculative or provisional production must be invento-
ried and, hence, can be considered to be inefficient and potentially
risky because a real customers purchase order is not secured. Many
popular production paradigms such as just-in-time (JIT), Kan-
ban, tach time, single minute change of a die (SMED), theory-of-
constraints (TOC), etc., are all examples of striving toward the
goal of DDP. The scheduling optimization solution presented in this
article is a critical step in the on-going struggle to improve effi-
ciency and profitability of an oil refinery or petrochemical plant
with respect to the DDP innovation.
Blendshop example. Marine-access blend-
shops are usually characterized by having a set of
storage or receiving tanks and a set of feed or
charging tanks with either a continuous- or batch-
type blend header in the middle. Pipeline-access
blendshops often only have receiving tanks
because settling of the unloaded crude oil for free
water removal after a marine-vessel has unloaded
is not required.
Fig. 1 shows a small blendshop problem with
one pipeline, two receiving tanks, one transfer-
line (or batch-type blend header), two charg-
ing tanks and one pipestill. Tank inventory
capacities and names are shown inside the tank
objects and flowrate capacities of the semicon-
tinuous equipment are shown above the
pipeline, transferline and pipeline objects. The bold arrows in the
figure indicate the flow or movement variables for the blendshop
scheduling problem. Connections between equipment are typi-
cally material based in the sense that only certain crude oils or mix-
tures are allowed to flow between a source and destination. These
material based connections permit use of crude oil segregations or
pooling by directing certain crude oils to be stored into specific
tanks. Segregations are useful when controllable equipment or
movements can be used to prepare a blend recipe or formulation to
be specified for charging the pipestill.
2
Segregations are useful to
reduce problem complexity by reducing the number of decisions to
be made in terms of where crude oils should be stored. In our
example we do not impose any batch recipe such as a 50:50 blend
for flow out of each segregation. That is, we do not impose a 50%
volume fraction from the light crude oil pool (TK1) and a 50%
fraction from the heavy crude oil pool (TK2). We have omitted
this detail so as not to detract from the main focus of the article.
Table 1 provides the information on crude oil receipts for four
different types of crude oils that are segregated into light and heavy
crude oil pools (i.e., crude oils #3 and #4 are light and #1 and #2
are heavy). The start and end times are in hours from the start-of-
schedule which is set at the zero hour; the end-of-schedule is on the
240th hr (10 days). Crude oil mixture liftings from the charging
tanks to the pipestill are continuously set at 5 Kbbl/hr and the
pipestills fuels production mode is unchanged over the scheduling
horizon. This defines the crude oil mixture demand schedule.
Fig. 1 also describes various operating rules that must be respected
for the blendshop for it to operate as a crude oil blendshop (more
details on these and others are given in Part 2). For this blendshop we
only allow one flow out of or in to the pipeline and transferline at a
time. Also, the tanks must be in standing-gage operation where there
can be only flow in or out at a time but not both. A generalization of
standing gage is the mixing-delay restriction. It imposes a time
delay after the last flow into a tank has finished before a movement
out is allowed. The receiving tanks have a 9-hr delay and the charg-
ing tanks a 3-hr delay.
The last logic constraint for this example is that all flows are semi-
continuous or disjunctive. This means that a flowrate must either
be zero or lie between lower and upper bounds.
Table 2 displays a subset of the assay information for the four
crude oils being stored and delivery into the blendshop. Only three
cuts are included: whole crude oil, kerosene and heavy gas oil. Blend-
ing of the cuts and assigned properties are based on volume or weight
depending on the property. Blending numbers or indices would be
SPECIALREPORT PROCESS/ PLANT OPTI M I ZATI ON
HYDROCARBON PROCESSING JUNE/JULY 2003
TABLE 1. Crude oil receipt or supply orders over
scheduling horizon (cycle dat a).
Crude St art t ime End t ime Valid
Oil # (hr) (hr) Flowrat e Durat ion Flow dest inat ion
1 82 91 20 Kbbl /hr 10 hr 200 bbl TK2
1 163 172 20 10 200 TK2
2 7 16 20 10 200 TK2
3 43 52 20 10 200 TK1
3 221 230 20 10 200 TK1
4 132 141 20 10 200 TK1
TK1
TK2
TK3
TK4
Receiving t anks
Charging t anks
Transf erline Pipeline
Pipest ill
One f l ow out
One f l ow i n
One f l ow out
One f l ow i n
220 Kbbl
100 Kbbl
20 Kbbl 16 Kbbl
Li ght crude oi l s
Heavy crude oi l s
Fuel s mode
5 Kbbl
Mi xi ng del ay = 9 hr,
st andi ng gauge
Mi xi ng del ay = 3 hr,
st andi ng gauge
Low er up-
t i me = 19
hours
Low er up-t i me
= 3 hr
Cut s
This blendshop problem involves one pipeline, t wo receiving t anks, t wo charging
t anks and one pipest ill.
FIG. 1.
used for those properties that blend nonlinearly such as Reid vapor
pressure (Rvp) and viscosity. Synergistic and antagonistic nonlinear
blending such as evidenced for octane are not considered further for
the crude oil blendshop problem.
Table 3 displays the minimum, target and maximum quality
specifications for the mix of crude oils required by the fuels operat-
ing mode. For this example we do not concern ourselves with bounds
on the quality variables though the scheduling optimizer can be con-
figured to respect these bounds over the scheduling horizon. The
target values are those typically found in the planning optimizer.
Table 4 completes the required information for the blendshop by
providing the opening tank inventories and crude oil compositions
in each tank.
The information provided for the example characterizes a simple
yet typical crude oil blendshop scheduling optimization problem.
It has been deliberately organized into two main data themes: model
and cycle. Model data are generally static and do not change within
the scheduling time horizon. They define the material/flow/capac-
ity network, desired operating logic and crude oil assay information
and mixture quality specifications. Cycle data define the dynamic
data that can change every time the next schedule is made such as tank
opening inventories and compositions, supply, demand and main-
tenance orders, and any actual or logged movements. The problems
model and cycle data can be further segmented into what we call
the quantity, logic and quality aspects of the problem and will be dis-
cussed in Part 2.
Use of the word cycle is taken from the well-known hierarchical
planning and scheduling philosophy of Bitran and Hax
3
who advo-
cate a rolling-horizon framework or scheduling cycle to mitigate
uncertainty due to such effects as order reliability, measurement inac-
curacies and execution errors. For more recent details see S. C. Graves
in the Handbook of Applied Optimization, 2002.
4
Potential benefits. Of course it makes sense to pursue only
those aspects of a solution to a problem that have value. Before we
describe how we formulate and solve the crude oil blend schedul-
ing optimization problem it seems prudent to analyze why we would
want to solve it. This leads into the discussion of expected benefits.
Three major types of disturbances arguably affect a refinery at
any time during its production or operation: crude oil mixture qual-
ity variability, ambient temperature changes and unreliable or faulty
equipment. Processing equipment malfunctions can cause serious
production outages and safety concerns, and are normally mitigated
by sound maintenance practices. Seasonal and diurnal ambient tem-
perature swings also disturb stability of operations and are mitigated
by providing increased cooling or heating capability and improving
controls.
The molecules from each crude oil receipt are eventually pro-
cessed at every unit within a refinery. Variation in crude oil mixture
quality charged to the pipestill is perhaps the single most influen-
tial disturbance to a refinery. It is the foremost reason why reducing
variability around the many quality targets by blending crude oils
is of tremendous importance.
Crude oil blendshop scheduling optimization is a relatively inex-
pensive and timely way to seriously improve performance of almost
any refinery. Five benefit areas are all aided by applying better blend
scheduling:
Reducing quantity and quality target variabilityAs men-
tioned, reducing quality target variance should be at the top of the list
for refinery improvements. Deviations from quality targets should be
minimized to charge the pipestill with a steady mixture of crude oil.
Steadier quality crude oil mixtures charging the pipestills will also
translate into steadier operations for downstream production units.
PROCESS/ PLANT OPTI M I ZATI ON SPECIALREPORT
HYDROCARBON PROCESSING JUNE/JULY 2003
TABLE 2. Crude oil assay inf ormat ion (model dat a).
Cut / propert y Crude oil #1 Crude oil #2 Crude oil #3 Crude oil #4
Whol e crude
oi l /speci f i c gravi t y 0.870 0.872 0.856 0.851
Kerosene/yi el d,
vol % 9.26 8.71 10.44 9.60
Kerosene/pour poi nt 27.0 42.0 31.0 27.0
Heavy gas oi l /yi el d,
vol % 9.16 9.75 10.17 9.51
Heavy gas oi l /speci f i c
gravi t y 0.864 0.878 0.859 0.849
Heavy gas oi l /sul f ur,
w t % 1.83 0.77 1.67 1.28
TABLE 3. Fuels product ion mode cut /propert y
specif icat ions (model dat a).
Cut / propert y M inimum Target M aximum
Whol e crude oi l /speci f i c gravi t y 0.863
Kerosene/yi el d vol ume
Kerosene/pour poi nt 30.833
Heavy gas oi l /yi el d vol ume
Heavy gas oi l /speci f i c gravi t y
Heavy gas oi l /sul f ur 1.508
TABLE 4. Crude oil opening invent ories and
composit ions (cycle dat a).
Tank I nvent ory Crude oil #1 Crude oil #2 Crude oil #3 Crude oil #4
TK1 100 Kbbl 0% 0 100 0
TK2 3 100% 0 0 0
TK3 50 100% 0 0 0
TK4 100 50% 0 50 0

Qualit y variabilit y was reduced wit h aut omat ed crude oil
blend scheduling.
FIG. 2.
It also makes good sense to run pipestills with a constant flowrate
for as long as possible. An example of an improved quality target or
key planning proxy is shown in Fig. 2.
Improving the ability to generate more than just feasible
schedulesFor those blendshops that are tightly resource con-
strained due to previous cost-cutting initiatives, it may be arduous to
generate a feasible schedule for the immediate future. For these blend-
shops it is valuable to have an automated scheduling application
generate in seconds what would take a human scheduler hours to
construct. Multiple better-than-feasible or what we call optimized
schedules may be presented that meet the production goals for selec-
tion by the scheduler. The effect of not being able to generate feasi-
ble schedules means that either the supply scenario must be changed
by distressing crude oil deliveries or demand must be altered by
decreasing or increasing the flowrate of crude oil mixtures charging
the pipestill. Unfortunately, both of these alternatives are undesir-
able for various reasons.
Rapid acceptance and inserting spot supply and demand
opportunitiesTypically, a refinery will be some mix of contract
(wholesale, nondiscretionary, strategic or base) versus spot (retail,
discretionary, tactical or incremental) crude oil purchases. Faster and
better ability for a refinery to assess whether a particular spot crude
oil purchase will result in a feasible operation the better the refinery
can capitalize on short-term market opportunities.
Consistency of schedulesA common problem in production
scheduling is that usually only one skilled scheduler can schedule a
refinerys crude oil blendshop effectively. When the main scheduling
individual is sick or on vacation it is very difficult to backfill with
another appropriately trained scheduler. Hence, if this occurs, sched-
ules generated by the two individuals can be widely different to the
point where unfortunately schedules generated by the relief may
actually be infeasible. Using an automated scheduling tool alleviates
some of these issues since schedules are made to satisfy the same
business logic and reflecting the same constraints and limitations.
Production schedule visibility throughout the refinery
Finally, given the use of Internet technologies, it should be standard
now and in the future to disseminate the official schedules online
so that managers, operators and engineers can all view the same pro-
duction program for the next several days or weeks. Although this can
be easily accomplished with spreadsheets and simulators, it is not
always possible with these solutions to look out into the future many
days or weeks and show the longer-term schedules to those who can
take advantage of greater look-ahead.
If we take for the purpose of discussion a medium-sized 100,000
bpd refinery, or equivalently 35,000,000 bpy with a 350-day yr pro-
duction schedule, it is possible to make a list of some of the expected
benefits and their value. Here we only detail the tangible benefits.
However, a range of intangible benefits can translate into significant
value. Each benefit cited is incremental over what would be achiev-
able using spreadsheets or simulators.
Quality target variability improvement such as whole crude
oil sulfur: $2,000,000/yr. This number was estimated from the ben-
efits identified when a similar application for crude oil blend schedul-
ing was developed and applied by the first author to a sweet crude oil
processing 100,000 bpd refinery. The benefits were captured at the
planning feedstock selection activity level because the proxy constraint
on bulk crude oil sulfur was raised from 0.55 to 0.85 %wt sulfur over
a three-month period (Fig. 2). This resulted in a cheaper slate of crude
oils being purchased while still being able to meet the quality specifi-
cations for all of the finished products refined and blended.
Reduced chemical injection: $100,000/yr. Further and unex-
pected savings on corrosion control chemicals were also observed
for the refinery over a one-year period due to the fact that less inhibitor
needed to be added given the improved regulation of crude oil bulk
sulfur concentration.
Distressed sale of crude oils from the refinery: 5-incidents/yr
$1.30/bbl 50,000 bbl = $325,000/yr. Here we assume that five
times in one year the refinery needs to distress a 50,000 bbl batch of
crude oil at a loss, in terms of lost netback production (opportunity)
and loss in selling price of $1.30 per bbl. This means that the refin-
ery lost the opportunity to process the crude oil and make netback
$1.00/bbl and sold the crude oil batch at a loss of $0.30/bbl.
Pipeline penalty charge for changes in sequence or timing:
3 incidences/yr $25,000/incidence = $75,000/yr. A penalty of
$25,000 per incidence is realized for altering the start or end time of
a batch of crude oil before it can be received at the refinery.
Spot opportunity for crude oil trades: 5 opportunities/year
50,000 bbl $1.00 net margin/bbl = $250,000/yr. There are five
extra opportunities per year to run batches of 50,000 bbl at a net-
back or net margin of $1.00/bbl due to better crude oil blendshop
scheduling.
Reduced working capital (decommissioning of a crude oil
tank): 50,000 bbl cycle stock $20/bbl 10% cost of capital/yr =
$100,000/yr. This is a result of sustaining a long-term reduction of
safety or cycle stocks of reserve crude oil in the crude oil blendshop.
Total savings: $2,000,000 $100,000 $325,000 $75,000
$250,000 $100,000 = $2,850,000/yr! These hypothetical and
somewhat anecdotal benefit calculations translate into over two-
and-a-half million dollars worth of the potential savings to the refin-
ery profit and loss that would not have been achieved with manual
scheduling alone. It must be emphasized that this number only pro-
vides a benchmark or yardstick to Pareto, at least qualitatively, the
priorities in terms of choosing between other possible and compet-
ing capital investment projects at the refinery. From the perspective
of overall crude oil costs of the refinery over one year, this $2.85 mil-
lion in savings is less than 0.41% of the total feedstock cost (i.e.,
$2,850,000 / (35,000,000 bbl) / ($20/bbl) 100 = 0.407%).
Formulating problem logistics and quality details.
Modeling the crude oil blendshop is the cornerstone of being able to
capture the potential benefits outlined. Although the modeling must
ultimately reside as a collection of complex mathematical expres-
sions relating variables and constraints to offer some level of opti-
mization, we supply a qualitative description of the model only. We
do this so as not to detract from the general understanding of the
overall problem and reasons for solving it. Some of the more spe-
cific details around the mathematical modeling of the crude oil
blendshop scheduling optimization problem can be found in Lee
et al.,
5
Shah
6
and Jia et al.
7
At the core of our formulation is the hierarchical

decomposi-
tion of the problem into logistics and quality subproblems. The logis-
tics subproblem is very similar to the supply chain logistics problem
except that our logistics problem has less spatial scope. It considers only
the crude oil blendshop (inside the production chain) and not the
entire supply chain but has more of an in-depth operational view of
the crude oil handling and blending. The logistics subproblem only
SPECIALREPORT PROCESS/ PLANT OPTI M I ZATI ON
HYDROCARBON PROCESSING JUNE/JULY 2003
Use of t he w ord hi erarchi cal means w e decompose t he probl em not al ong t he spat i al or t empo-
ral di mensi ons but al ong t he deci si on-maki ng di mensi on such as pl anni ng t hen schedul i ng t hen
execut i on.
considers the quantity and logic variables and constraints of the prob-
lem and ignores the other quality variables and constraints. The qual-
ity subproblem is solved after the logistics subproblem whereby the
logic variables are fixed from the logistics solution and the quantity
and quality variables are adjusted to respect both the quantity and
quality bounds and constraints. The quality optimizer is very simi-
lar to commercially available oil refinery and petrochemical plan-
ning software formulations which are used to select the crude oils, for
example, that will be processed at the refinery.
The reason for the logistics and quality subproblem decomposition
is three-fold. First, commercially available optimization software, opti-
mization theory and computer horsepower have not progressed to
the point where we could solve a full-blown simultaneous quantity, logic
and quality crude oil blendshop scheduling problem in reasonable
time. Second, theory tells us that if you cannot find a feasible solu-
tion to the logistics subproblem then you will not be able to find a
feasible solution to the quality subproblem. Hence, the decomposition
provides a very useful problem-solving aid because if the logistics sub-
problem is infeasible for whatever reason (i.e., bad input data, overly
aggressive production plans, etc.) then there is no use spending time
solving the quality subproblem until something has changed in terms
of the quantity and logic aspects of the problem. However, even if the
logistics subproblem is feasible there is no guarantee that the quality
subproblem is feasible. Third, theory also tells us that if the logistics sub-
problem is globally optimal (i.e., the very best solution has been found)
and the quality subproblem is feasible then we have found the global
optimum for the overall problem. Hence, our decomposition pro-
vides a very powerful structure given that it is easier to check for qual-
ity feasibility than quality global optimality after the best logistics
solution has been found. Unfortunately if the quality subproblem is
infeasible then there must be a mechanism to send back special con-
straints to the logistics subproblem to force it away from those regions
of the search space that are known to cause quality infeasibilities.
We find that this break down of the overall problem into two
subproblems is in fact very intuitive for the scheduling users who
are using spreadsheets, especially the aspects of the quantity and
quality. The logic details are known by the users but rarely included
in formulating their spreadsheets due to the discrete nature (i.e.,
requires some search mechanism and/or trial-and-error). Logic aspects
are usually resolved ad hoc after a quantity-quality solution has been
circumscribed, but only for the immediate near term of the schedule.
There could also be further decomposition within each of the
logistics or quality subproblems. For example, in the logistics sub-
problem it is possible to decompose the scheduling into assignment
and sequencing stages. The assignment stage can be solved relatively
easily to optimality by assigning orders or jobs to equipment and
ignoring the equipment job sequencing. The assignment decisions
are then fixed and the sequencing stage is solved. If a feasible sequenc-
ing solution can be found then the overall logistics subproblem is
feasible. If the lower-level sequencing is infeasible then extra con-
straints are added to the assignment stage problem to guide the
higher-level solution away from those assignments that are known to
cause problems for the sequencing.
8
A strong parallel to the logistics and quality decomposition is
found in discrete parts manufacturing refined by the Japanese. That
is the decomposition of JIT, Kanban, SMED, TOC, etc., with the sta-
tistical quality control philosophies of W. E. Deming and S. Taguchi.
A clear separation between the two is brought together in the end to
guide the manufacturing machine to produce quality products effi-
ciently, effectively and punctually.
Finally, formulating crude oil blend scheduling optimization is in
the class of production scheduling known as a closedshop.

Defini-
tion of a closedshop, and its counter-part an openshop, can be found
in the review paper by Graves.
9
In an openshop all production orders
are by customer request and no inventory is necessarily stocked. In
a closedshop all customer requests are serviced from inventory and a
production activity is generally a result of inventory replenishment
decisions. These definitions really state that closedshops involve
quantity variables and inventory balances whereas openshops typically
dont. Closedshops are generally associated with lot-sizing problems
(requiring a flow path or network to be defined) and almost always
are formulated using some form of scheduling horizon segmenta-
tion into time periods. Even when continuous-time closedshop for-
mulations are used (see Jia et al.)
7
the number of time-event points
is required as an input. Segmenting the time horizon is necessary to
perform inventory or material balances. Solving the industrial-scale
closedshop problem has been attempted by first determining lot,
batch or blend sizes and then making the assignment, sequencing
and timing decisions (or logic decisions) using an openshop frame-
work. However, this decomposition has been met with limited suc-
cess. A unique formulation of the logistics subproblem is to model and
solve the blendshop as a closedshop explicitly by including both
quantity and logic decisions simultaneously in one optimization.
To facilitate more specific information on the formulation we
must first talk about the problem variables. These can be classed
into continuous and combinatorial variables. Continuous vari-
ables are the quantity and quality variables, and the combinatorial
variables are the logic or discrete variables. There are also auxiliary
or intermediate variables such as startup (and shutdown or
switchover) and flow times yield variables that are used to sup-
port solving both the logistics and quality subproblems. Bounds
and constraints associated with these variables follow.
Quantity details (hydraulic capacities). There are essen-
tially three types of hydraulically related quantity bounds: flowrate,
flow and inventory. Each of these has continuous variables asso-
ciated with them in both the logistics and quality optimizer for-
mulations. All inventory variables are related to the flows through
the material balances on each piece of equipment.
Flowrate bounds are capacity bounds associated with a move-
ments process and transfer-type equipment such as pipestills,
headers, line segments, pumps, valves, etc. They specify how much
material can flow within a certain amount time through the piece
of equipment and are defined by an upper and lower bounds.
Flow bounds specify a quantity of material that can be trans-
ferred from one piece of equipment to another. They extend the
flowrate bound to fully describe a supply or demand order. Know-
ing the rate and the quantity determines the duration. Both flow
and flowrate bounds are associated with a connection between a
source and destination piece of equipment and ultimately relate to
the underlying limiting or shared transfer-type piece of equip-
ment that moves the material from the source to the destination.
Inventory bounds are capacity bounds for inventory-type
equipment such as spheres, tanks or drums. They specify how
much material can be stored in a piece of equipment and are
defined by an upper and lower bound.
PROCESS/ PLANT OPTI M I ZATI ON SPECIALREPORT
HYDROCARBON PROCESSING JUNE/JULY 2003

A bl endshop i s a cl assed as cl osedshop.

Logic details (operating rules). Fourteen different kinds
of logic constraints are typical of a crude oil blendshop operation.
This list is not exhaustive but is a very reasonable starting point. As
mentioned, to model these constraints we need to have logic variables
or combinatorial variables. These are also referred to as 0-1 or binary
variables and are associated specifically with a flow between source
and destination equipment. Zero indicates the flow is inactive and
one implies the flow is active and must be between its lower and
upper flow bounds. We also have two other logic variables to indi-
cate when a flow route has been started up (time it is made active)
or has been shut down (time it is been made inactive); these variables
are also used to model transition or switchovers.
Semicontinuous (SC) constraints represent a flow that can be
zero or between a lower or upper bound. Without SC constraints
the logistics problem would become a linear program and not a
mixed-integer linear program.
Standing gage (SG) constraints enforce the practice for tanks
where there can be flow in or out but not both at the same time
(mutually exclusive). SG constraints are useful to decouple the
production chain from the supply chain upon receipt of a crude oil
delivery for example, and to enable tank level differences to be
used as a cross check for custody transfer meters.
Mixing delay (MD) constraints restrict flow out of a tank
until a certain amount of time has past after the last flow in. A
tank must have SG constraints set for mixing delay to be used.
The MD constraints are useful to allow separating ballast-free
water after a marine vessel unload.
One flow in (OFI) constraints prevent more than one flow
into a piece of equipment at a time. OFI constraints are useful to
model cases where a pipestill can only be fed from one tank at a
time for example.
One flow out (OFO) constraints prevent more than one flow
out of a piece of equipment at a time. OFO constraints are useful
to model cases where a pipeline can only discharge to one tank
at a time.
Contiguous order fulfillment (COF) constraints define a
flow to be fixed quantity and fixed rate over specified start and
end times. They are typical of pipeline receipt and delivery orders.
In these cases the flowrate is equal to the quantity divided by the
difference between the end and start times. These order fulfill-
ment types are such that there is a contiguous or consecutive flow
between the order start and end time (i.e., an uninterrupted or
non-preemptive flow).
Noncontiguous order fulfillment (NOF) constraints define
order fulfillment to be the opposite of the COFs. Arrival and
departure dates are specified for supply orders and release and due
dates are specified for lifting orders. The NOFs are defined with
a specified order quantity such that between the arrival and depar-
ture date or release and due date, the cumulative quantity of mate-
rial that has flowed from the pipeline or to the pipestill equals
that specified by the order. This implies that there can be non-
contiguous or nonconsecutive flows from or to a piece of equip-
ment (i.e., an interruptible or preemptive flow). Arrival and depar-
ture dates are useful for handling marine vessel unloading when
arrival due to inclement weather conditions causes higher than
normal uncertainty levels. Release and due dates are useful for
specifying pipestill production mode orders because the planning
solutions will typically say how much crude oil to process within
a particular time horizon with the detailed flow scheduling to be
determined by the scheduling optimization program.
Lower up-time (LUT) constraints are identical to minimum
production run length-type constraints. They are used to specify
a minimum time a particular movement needs to be up or active
before shutting down or becoming inactive.
Upper up-time (UUT) constraints are used to specify a max-
imum contiguous time a movement can be up before it is required
to be shut down.
Equal flow (EF) constraints force the same flow value for a col-
lection of time periods where the movement is contiguously or
consecutively active. Either a lower or upper up-time must be
specified before the equal-flow constraints will be added for that
particular source-destination pair.
Switch-over-when-empty (SWE) constraints indicate a move-
ment cannot switch over or shut down until tank inventory is less
than some specified threshold. This is useful when a charge or
feed tank must be near empty before it can be shut down or before
another tank can be used to charge the pipestill.
Switch-over-when-full (SWF) constraints are very similar to
the SWE constraint where a movement cannot be shut down until
the tank is full. This is useful for receiving or storage tanks when
being fed from a pipeline because it tries to fill a tank before mov-
ing on to another one if the volume or quantity of the delivery
order is greater than the available ullage.
Startup opening (SUO) bounds are applied to a particular
startup variable for a movement and are used to restrict the time
of day when that movement can be started up. For example, it
may be useful to only have a switchover to a different tank of
crude oils feeding a pipestill during the day shift (between 8:00 a.m.
and 4:00 p.m.).
Shutdown opening (SDO) bounds are similar to the SUO
except that they tell the logistics optimizer when a possible move-
ment shutdown can occur.
Quality details (property specifications). The intensive
properties of the crude oil mixtures charging the refinery must be
carefully regulated for the pipestill to meet the downstream quality
stipulations or specifications when operated in a particular pro-
duction mode. These qualities are associated with the temperature
cutpoints or cuts of the different hydrocarbon streams being sepa-
rated by the pipestill and must be modeled as continuous variables
in the quality optimizer formulation. Quality balances or equations
must be associated with each quality throughout the entire blend-
shop where we model tanks as perfectly mixed vessels. The quality
balances force the subproblem to be nonlinear due to the product of
quantity (flow and inventory) times quality. Quality splitter equa-
tions model the situation of multiple simultaneous flows out of
equipment to ensure that each outlet stream has the same quality as
all of the other outlet streams. Following is a somewhat complete list
of the many streams produced by the pipestill or atmospheric and
vacuum distillation unit with typical properties that could be typ-
ically assigned or measured for the pipestill output streams.
Wet or saturated gas cut/properties include both the vol-
ume and weight yields of the pure components methane, ethane,
propane, iso- and normal-butane, specific gravity, etc.
Light and heavy straight-run naphtha cut/properties include
both the volume and weight yields, paraffins, olefins, naphthenes
and aromatics (PONA), Rvp, octane, specific gravity, sulfur, etc.
Jet fuel and kerosene cut/properties include both the vol-
ume and weight yields, cloud point, freeze point, pour point, spe-
cific gravity, sulfur, etc.
SPECIALREPORT PROCESS/ PLANT OPTI M I ZATI ON
HYDROCARBON PROCESSING JUNE/JULY 2003
Diesels and middle distillates cut/properties include both the
volume and weight yields, cloud point, flash point, pour point, spe-
cific gravity, sulfur, viscosity, etc.
Heavy distillates cut/properties include both the volume and
weight yields, basic nitrogen, metals (nickel, vanadium, iron),
refractive index, specific gravity, sulfur (total and reactive), vis-
cosity, etc.
Light and heavy vacuum gas oils cut/properties include
both the volume and weight yields, base oils, basic nitrogen, met-
als (nickel, vanadium and iron), refractive index, specific gravity,
sulfur (total and reactive), viscosity, etc.
Vacuum residue or pitch cut/properties include both the
volume and weight yields, asphaltenes, base oils, carbon number,
metals (nickel, vanadium and iron), penetration, specific gravity,
sulfur (total and reactive), viscosity, etc.
Logistics and quality objective function details. Now
that we have enumerated the variables and constraints of the two
subproblems it is important to talk about the driving force for
optimization. This underlying forcing function is the objective
function that is continuously being maximized during the course
of the logistics and quality searches over the entire scheduling
horizon (start-to-end of schedule). Both the logistics and quality
objective functions are separated into three terms.
The first term is profit defined as revenue of crude oil mixtures
minus the feedstock costs of the delivered crude oils and any inven-
tory holding or carrying costs for both types of tanks. The profit
function is identical to both the logistics and quality subprob-
lems although the quality profit term can be extended to include
individual revenue generated from the cut yield flows. The sec-
ond term is required to maximize performance. Performance for the
logistics subproblem is defined so as to minimize the number of
active movements and the number of movement startups and
shutdowns (i.e., transitions or switchovers). Another term in the
performance category is to minimize deviation of any tank inven-
tory from a closing inventory target specified by the user. This is
also used in the quality subproblem but is extended to include
deviations from user-specified quality targets on the crude oil
compositions and cut properties. Ad hoc performance weights
are usually used for each performance type and can be tuned based
on the priority level dictated by the scheduling user.
The third term is very important when solving real-world prob-
lems. Not all input data required to solve for optimized sched-
ules is good or free of gross errors (see Kelly
10
for a list of possible
sources of error). Therefore, we must always anticipate that some
infeasibilities may occur before the data have been optimized and
carefully cross-checked for validity. In light of this, all quantity,
logic and quality constraints have artificial or penalty variables
associated with them. Each penalty variable is weighted and min-
imized in the objective function so that the most important busi-
ness practices at a site are respected when they cant all be met. If
the problem data are free of gross errors or flaws (as some people
refer to them) then the penalty variables will be driven to zero by
the optimizer meaning all business requirements are satisfied. The
penalties are also known in the planning domain as infeasibility
breakers or safety valves.
Ultimately, the scheduling optimization objective function is
used to balance the three costs of manufacturing: cost of renew-
able and nonrenewable resources (i.e., materials, equipment, labor,
utilities, chemicals, etc.), inventory (i.e., it costs money to store
materials and equipment) and transitions (i.e., startups, shutdowns,
changeovers, switchovers, sequencing, etc.).
Typical planning optimization systems only include the
resource and inventory costs and do not model transition costs.
The major reason is due to the mathematical intractability of
solving simultaneously for quantity, logic and quality given todays
state of optimization technology. Consequently, transition costs
are excluded from the planning models and only quantity and
quality details are formulated, except for minor logic details con-
cerning cargo or batch size increments for feedstock availability.
Because transition costs are relegated to the scheduling layer, all
planning solutions are overoptimized. This implies that all plan
versus schedule or plan versus actual analysis will have inherent
biases or offsets even if measurement, model, solution and exe-
cution errors are negligible
10
and strongly suggests that these
biases be interpreted carefully.
Time modeling. Both planning and scheduling involve time
considerations. There are principally two types of time model-
ing. The first and most used and studied is time discretization
into predefined fixed duration time periods but not necessarily
of equal duration over the scheduling horizon. All activities are
defined to start and end at the time period boundaries and are
piece-wise continuous over the time period duration.
The second time model is the most elegant and is that of con-
tinuous-time modeling whereby activity start and end times are
included explicitly as optimization variables. An example of con-
tinuous-time formulation of the crude oil blend scheduling opti-
mization can be found in Jia et al.
7
Continuous-time models also
have the notion of time periods except that these have variable
durations determined by the optimizer.
The recognized disadvantage of discrete-time formulations are
that they require a large number of time periods to model the
smallest duration activities, however, continuous-time modeling
enables each piece of equipment to have its own timetable. This
removes the need to artificially synchronize all equipment to be on
the same timetable and thus reduces the number of logic or binary
variables. There are nonetheless advantages of discrete time in
that it scales well when long time horizons are required for what-
if studies because larger time period durations can be used and it
can handle easily time-varying quantity bounds and out-of-ser-
vice orders. With continuous-time formulations, time-varying
tank inventory bounds, for example, require extra binary vari-
ables to be generated for the optimizer to assign which time period
the tank inventory capacity change is to take place even though we
know explicitly the event time of the change. Therefore, for the
immediate future, discrete-time formulations seem to have value
over continuous-time formulations given the previous discussion,
yet in the end both discrete- and continuous-time formulations
should be available to the scheduling user.
One final note on time models, the popular distinction now
between production planning models and production schedul-
ing models, with underlying structures of the lot-sizing problem,
is through the notion of big buckets and small buckets to discretize
time. This can be found in Belvaux and Wolsey
11
who also have
LOTSIZELIB, a library of diverse lot-sizing problems. The fun-
damental difference between big and small buckets, where big
buckets are used typically to model planning problems, is that big
buckets are those in which several materials can be produced on a
convergent-flow-path

piece of equipment, such as a blend header,

PROCESS/ PLANT OPTI M I ZATI ON SPECIALREPORT
HYDROCARBON PROCESSING JUNE/JULY 2003
during a single time period. Small buckets are typically used to
model scheduling problems where only one material can be pro-
duced on a single piece of equipment at a time (single-use or unary
resource logic constraints). Small time buckets are used to model
startups, switchovers and shutdowns as is the case in our formu-
lation of the crude oil blend scheduling optimization problem.
Segregating crude oils into tanks. A salient aspect of
crude oil handling and blending is that of segregating crude oils into
specific tanks. Segregation is used is to separate disparate crude
oil types into different tanks to maintain the flexibility or con-
trollability to blend to specific cut property values (i.e., specifi-
cation blending as opposed to recipe blending). The first require-
ment of crude oil segregation is to understand the key cut property
constraints.
From a degrees-of-freedom analysis, the number of key con-
straints must be less than or equal to the number of tanks used
to blend the crude oil mixtures (i.e., typically the number of receiv-
ing tanks). For example, in the example where there are two receiv-
ing tanks, at most two cut properties can be controlled at any
given time. Since there is a supply order of 5 Kbbl/hr this reduces
the number of degrees-of-freedom by one and hence, only one
cut property at any time can be controlled. Once the key cut prop-
erties have been identified then the crude oils should be separated
according to the level of each in the crude oil. All in all, effective
segregation can be difficult to figure out but can be automated
following the control and optimization techniques found in Kelly.
2
Usually very simple isolation rules are applied based on crude oil
bulk sulfur or density levels. Another relevant reason segregation
is used is to reduce complexity of the logistics subproblem. When
we preassign specific crude oils into tanks the number of choices
where an individual crude oil receipt can be stored is circum-
scribed by the segregation. In our scheduling formulation we han-
dle segregations by pruning the available connections between
pipelines and receiving tanks. For instance, in the example with two
receiving tanks and two segregations, light and heavy crude oils,
only light crude oils 3 and 4 are allowed in TK1 and only heavy
crude oils 1 and 2 are allowed in TK2. Hence, of the possible
eight crude oil-based connections (two tanks times four crude
oils) only four are allowed.
Continuous and batch blending. When most people think
of blending in the process industries they envision simultaneous
mixing of the blend constituents or components in some mixer or
blend header. This is known as continuous blending. When we
are solving the logistics subproblem, a fixed recipe or bill-of-mate-
rials is required that relates the blend volume size to the fractions
of each component material feeding the blend header. This is
known as recipe blending. In the quality subproblem, specifica-
tion blending is performed whereby the recipe is determined based
on the property specifications of the blended product. Continu-
ous blending is relatively straightforward to model because at
every time period we impose either the recipe constraints in the
logistics optimizer or the quality constraints in the quality opti-
mizer. However, in the quality optimizer, specification blending
makes the problem nonlinear.
Batch blending can be considered as the opposite to continuous
blending similar to batch distillation or separation. Batch blending
mixes the required components sequentially in a destination tank
with the components typically being fed one after the other. Both
recipe and specification blending are achievable using batch blend-
ing similar to the continuous blending. Yet instead of the blending
constraints being set-up for each time period, batch blending requires
the constraints to be specified over a time window made up of two
or more time periods so that the component additions are the equa-
tions cumulatively. In our example we employ batch blending at the
transferline with the restriction that components can flow into the
transferline one at a time. The time window we use for our example
is arbitrarily chosen at 20 hr.
It is also important to mention that components included in the
blending equations are not the individual crude oils

but the
crude oil segregations or mixtures. For instance in our example, the
two blending components are light and heavy crude oils.
Last of all, if we could solve the overall problem simultane-
ously for quantity, logic and quality then we would not have to con-
cern ourselves with the side issues of segregating crude oils into the
receiving tanks and specifying a nominal recipe for the blend head-
ers. These aspects would be dealt with effectively by the single
optimizer and it would determine where to put the crude oils
upon delivery and how much of each crude oil mixture from each
receiving tank should be set to through the blend header. The
only other effect that would preclude us from achieving almost
perfect crude oil blend scheduling optimization would be the
type, sequence and amount of each crude oil supply order and
potentially the production run schedule on the pipestills. Unfor-
tunately simultaneous quantity, logic and quality solutions are
not attainable given the present state of optimization technology
and, hence, puts the onus on the scheduling user to properly con-
figure the system to help overcome the solver limitations and to go
on to generate better-than-spreadsheet or simulator-type schedules.
Solving the problem f or logistics and quality. Since
both the logistics and the quality subproblems have been care-
fully formulated as mathematical programs, solving them using
commercially available optimization codes is our next step to
achieve better crude oil blend scheduling optimization. From the
perspective of finding optimized solutions, we can class all solutions
coming out of both the logistics and the quality optimizers as
infeasible, feasible, approximate (locally optimal) and globally opti-
mal, given that both subproblems are known to be nonconvex.

Infeasible solutions do not satisfy all of the problem constraints, fea-

sible solutions do satisfy all of the constraints, approximate solu-
tion are feasible and are deemed to be of reasonable worth (best
within some neighborhood) and globally optimal solutions are
the best overall. For our purposes we concern ourselves with
approximate solutions given that running the optimization searches
to find the global optimum may take a very long time (i.e., more
time than we are willing to wait for an answer).
Before we begin our discussion on solvers, an essential proce-
dure known as roll-forward is required to determine tank opening
inventories and compositions at start-of-schedule (SoS).
SPECIALREPORT PROCESS/ PLANT OPTI M I ZATI ON
HYDROCARBON PROCESSING JUNE/JULY 2003

Convergent -f l ow -pat h t ypes of equi pment can consume many i nl et mat eri al s and produce onl y
one out l et mat eri al . Conversel y, di vergent -f l ow -pat h uni t s can consume many mat eri al s and
produce t w o or more out l et mat eri al s.

Not unl ess a segregat i on cont ai ns onl y one crude oi l .

Nonconvex i mpl i es t hat t he search space i s di chot omous, di scont i nuous or di sj unct i ve w here a
l ocal l y opt i mal sol ut i on i n one subregi on i s not necessari l y t he gl obal opt i mum over al l of t he
ot her subregi ons. Convex probl ems i mpl y t hat a l ocal opt i mum i s coi nci dent w i t h t he gl obal
opt i mum because t here i s onl y one regi on over w hi ch t o search.
Rolling the inf ormation f orward to start-of -sched-
ule using simulation. As mentioned previously, advanced
planning and scheduling solutions are used within a rolling hori-
zon construct to mitigate inherent effects of uncertainty in the
exogenous information of the problem. This is a best-practice
policy introduced by Bitran and Hax
3
as part of their hierarchical
planning and scheduling approach (see also Clark and Clark
12
for
a recent application to the lot-sizing problem). In the context of
crude oil blend scheduling optimization the business problem of
roll-forward is two fold and is typically carried out every business
work day on a daily cycle except of course for weekends. The first
operation of roll-forward, using prior opening information and
actual movement data, is to predict current or baseline (e.g., 7:00
am) tank inventory data. This is checked to ensure measured data
for tank inventories are consistent to what is simulated by the
scheduling tool. If there is a discrepancy between the measure-
ments and simulated values then it is up to the scheduling user
to resolve the differences by cross-checking with other information.
The second aspect of roll-forward is to predict changes from
baseline to SoS to arrive at the initial conditions for the next
scheduling optimization. This is also a simulation-type function,
which uses in-progress and any future movements that are or will
be occurring between the baseline and SoS. Any movement activ-
ities that cross the baseline or SoS are truncated and the move-
ment quantity is prorated so that only the amount within the
time frame in question is used in the simulation.
Two types of simulation technology can be used to solve the
roll-forward function: the sequential modular approach (SMA) or
the simultaneous equation approach (SEA). The SMA is some-
times referred to as the closed-formapproach found in process sim-
ulators. It requires external knowledge of the material flow path
and simulates each piece of equipment individually, which can be
somewhat complex when anywhere to anywhere-type of blendshop
networks exist. The main disadvantage of SMA is that it does not
handle reverse-, recycle- or recirculating-type flows well and requires
an iteration loop to converge when they are present. It has the advan-
tage of being able to handle discontinuous and complex nonlinear
functions to model difficult reaction kinetics and fluid mechanics.
The SEA is sometimes referred to as the open formapproach found
in process optimizers. It has the disadvantage that all nonlinear
equations must be continuous and once differentiable but has the
advantage of being able to handle the reversal-type flows easily. The
SEA requires the topology to be an implicit part of the model to
allow for easy handling of anywhere to anywhere-type of blend-
shop networks. For our scheduling application we use the SEA.
Specifically, the SEA is well suited to crude oil blendshop simulations
because we blend or mix linearly by either volume or weight.
Logistics solving methods. Although there is a paucity of
literature documenting the quantity and logic formulation of
continuous/semicontinuous (CSC)-type processes, there is, how-
ever, a remarkable amount of literature on the techniques being
used to formulate batch/semi-batch (BSB) type processes both
in the operations research (OR) literature and in the chemical
engineering journals on process synthesis engineering. That said,
the underlying mathematical programming theory used to aid
in formulating the crude oil blendshop problem was mostly found
in the OR literature
13, 14, 15
and relates to the classic problem for-
mulations of the fixed-charge network flow, lot scheduling and facil-
ity location problems.
At the core of the logistics optimization is use of the branch-
and-bound (B&B) search heuristic using linear programming (LP)
as the underlying sub-optimization method; this is also commonly
referred to as mixed-integer linear programming (MILP). It is well
known and can be found in many textbooks. B&B is an exact search
method in that if given enough time it would arrive at the global opti-
mum. The B&B begins by solving an LP with all of the binary vari-
ables relaxed to lie between zero and one. Then the search begins to
successively fix binary variables to either zero or one based on elab-
orate variable selection criteria and solving an LP for each newly
bounded binary variable. After each LP, which are called the B&B
nodes, another selection criterion is required to chose which node
will be branched on next. The B&B will terminate, kill or fathom
a branch of the search tree for two reasons. The first happens when
a node along the branch is recognized to be infeasible. The second
is called value dominance and happens when the nodes objective
function value is less than the value of the incumbent integer-feasible
solution for maximization problems. The incumbent integer-feasible
solution is the last solution found that has all binary variables at
the extremes of either zero or one. Consequently there is no sense
continuing a search on a branch that is infeasible and it does not
seem beneficial to follow a branch that is not as good as the cur-
rent integer-feasible solution found so far. This technique can have
other flavors to the search such as breadth first and depth first with
backtracking and more details can be found in standard textbooks
on integer programming. Moreover, other enhancements to the
B&B search include cutting planes, special ordered sets and vari-
able prioritization which in general can speed the search to find
good integer-feasible solutions.
Unfortunately even with the most efficient formulation, clever-
est B&B search and fastest LP code, finding good integer-feasible
or approximate solutions can be very time consuming. Hence,
we must be somewhat more pragmatic from the perspective of
the quality of the logistics solutions that can be found in reason-
able time. To help speed the search, a myriad of heuristics have
been the focus of much research in both OR and artificial intelli-
gence (AI). These are referred to as primal and meta-heuristics.
Primal heuristics use results of the LP solutions and successively
round and fix binary variables to either zero or one. Examples are
the pivot-and-complement,
13
relax-and-fix,
15
dive-and-fix,
15
smooth-and-dive
16
and chronological decomposition.
17
Meta-
heuristics use a metaphor usually found in nature to devise a search
strategy that exploits a particular nuance of the natural mecha-
nism. Examples include the genetic algorithm, tabu search, scat-
ter search, simulated annealing, ant colony optimization and
squeaky wheel optimization.
Many other heuristics or approximation algorithms can be
found in the OR and AI literature and are basically separated into
two categories: greedy search and local search. Greedy searches are
typically used to find quickly integer-feasible solutions in some
greedy fashion with a myopic view of the search space. Greedy
searches tend to exploit some detail of the problem to enable some
fixing of the binary variables. Local searches are basically a refine-
ment on top of greedy searches to try and find better solutions,
essentially using a trial-and-error approach, in the neighborhood
of the greedy search solutions. An interesting example of local
search applied to the lot-sequencing can be found in Clark.
18
All
in all, most relatively successful heuristics for practical size prob-
lems require some form of a B&B search with backtracking and will
typically embed a commercial B&B code in the algorithm.
PROCESS/ PLANT OPTI M I ZATI ON SPECIALREPORT
HYDROCARBON PROCESSING JUNE/JULY 2003
Quality solving method. Solving for the quality variables of the
problem is carried out using well-established successive linear pro-
gramming (SLP). SLP technology is the cornerstone of all solving
methods found in oil refinery and petrochemical large-scale plan-
ning systems. An example SLP algorithm can be found in Palacios-
Gomez et al.
19
which in spirit is used by many of the SLP solvers
today. Success of SLP as the method of choice for solving industrial-
size planning and scheduling arises from it use of the LP. As LP tech-
nology improves SLP technology improves because the major itera-
tions of the SLP are simply the LP solutions. Although SLPs are well
documented to be more suitable for mildly nonlinear problems with
either none or only a few degrees-of-freedom at the optimum (i.e.,
otherwise known as superbasic variables), the maturity of LP tech-
nology plays a major role in the SLP success over other nonlinear
solvers such as successive quadratic programming or conjugate-gra-
dient methods for example.
One of the biggest advantages is the use of presolve.
14
Presolve
is applied before any LP is solved and can dramatically reduce LP
matrix size (i.e., fewer rows and columns) through clever tight-
ening, consistency and probing techniques, and can remove eas-
ily vacuous and redundant constraints and variables; presolve is
also used in the MILP solutions. While the other nonlinear solvers
could also take advantage of presolve, these nonlinear solvers often
do not employ third-party commercial LP codes that have many
man-years of development implementing incredibly efficient and
fast presolving techniques. A second advantage is the use of inte-
rior-point and simplex (both dual and primal) LP solving meth-
ods where needed in the SLP algorithm. Since commercial LP
codes offer both interior-point and simplex methods, the SLP
program can be tailored to use the appropriate LP method at each
step. Nonlinear solving codes usually use only one solving tech-
nique. For example, for large problems it is appropriate to solve the
initial LP using interior-point then any subsequent LP resolves
use the dual-simplex; this is also true for MILP problems.
Need for the SLP formulation is of course borne out by the
product of quantity times quality or a flow times a cut/yield for
instance. When blending is performed linearly either by volume
or weight, in the absence of any antagonistic or synergistic effects
requiring nonlinear blend laws, this makes the problem both bilin-
ear, trilinear and quadlinear. Its trilinear because of the flow times
cut/yield times cut/property and quadlinear because of the density
property required when performing the weight balances. Unfor-
tunately this makes the problem nonconvex as mentioned, and
to solve it to global optimality necessitates use of global opti-
mization techniques found in Adya et al.
20
To solve to global opti-
mality requires a spatial B&B search similar to the MILP B&B
search except that the branching variables are continuous and not
binary. In our case they would be the flow and quality variables.
Due to the fact that global optimization is very slow and no com-
mercial software is available, we claim only to search for locally
optimal or approximate quality subproblem solutions.
A side benefit to solving for the logistics subproblem first in
series, then solving for the quality subproblem, is the actuality that
the SLP solves faster than if we were to solve for the qualities first
(as in the planning systems that solve for quantity and quality).
The reason is that the logistics solution provides us with an excel-
lent starting position or local neighborhood for the flows and inven-
tories. This aids the SLP where it is well known that all nonlinear
programs do better when better initial guesses are provided.
Example results. Fig. 3 shows one penalty-free logistics solu-
tion with a 10-day time horizon. The blue horizontal bars are the
supply and demand orders. The yellow bars are flow out of the equip-
ment and the green bars represent flow into the equipment. The
trend lines superimposed on the tank equipment show the inven-
tory profiles that are within the limits of their respective upper and
lower bounds. The major ticks on the x-axis are strategically spaced
at a distance of 20 hr and the minor ticks are positioned at every 5 hr.
The y-axis shows the renewable equipment resources starting from
the pipeline at the top down to the pipestill displayed at the bot-
tom. If either quantity or logic penalties were encountered they
would be shown as red bars between the flow to and from bars for each
equipment. It is clear that because there are no penalties there were
no logic constraints (standing gage, mixing delay, minimum up-
time, etc.) violated and all inventory and flow bounds were simul-
taneously respected without incident, i.e., this schedule satisfies
100% of the business practices and needs over the entire horizon.
As can be seen in the figure we have satisfied all of the six supply
orders for the pipeline (PL1) and segregations are properly main-
tained in that only those crude oils that belong to a segregation can
SPECIALREPORT PROCESS/ PLANT OPTI M I ZATI ON
HYDROCARBON PROCESSING JUNE/JULY 2003
0
PS1
TK4
TK3
TL1
TK2
TK1
PL1
20 40 60 80 100 120
Time, hr
E
q
u
i
p
m
e
n
t
140 160 180 200 220 240
This Gant t chart shows one penalt y-f ree logist ics solut ion
wit h a 10-day t ime horizon.
FIG. 3.
0
0.850
0.855
0.860
0.865
0.870
0.875
0.880
20 40 60 80 100 120
Time, hr
W
h
o
l
e

c
r
u
d
e

o
i
l

s
p
e
c
i
f
i
c

g
r
a
v
i
t
y
140 160 180 200 220 240
Trend of whole crude oil specif ic gravit y cut propert y. FIG. 4.
fill a storage tank (TK1 and TK2). Flows from TK1 and TK2 to the
transferline (TL1) comply with the 3-hr minimum run length as well
as the 9-hr mixing delay specification. To observe mixing delay on
tanks count the number of hours from the end of a green in-flow bar
to the start of the first out-flow yellow bar. The long run lengths
for flows from each of the two feed tanks (TK3 and TK4) charging
the pipestill (PS1) also comply with the 19-hr up-time minimum
constraint and 3-hr mixing delay. All standing gage restrictions were
also obeyed since no green and yellow bars overlap for TK1, TK2,
TK3 and TK4. The demand order of continuously charging 5
Kbbl/hr or 120 Kbpd to PS1 was additionally met.
This logistics solution took approximately 60 seconds to gen-
erate on a 1-gigaHz PC which involved solving an MILP. No spe-
cial heuristics except for the default settings in the B&B search
were used. Table 5 illustrates the power of presolve. The number
of inequalities or rows is reduced by 58% and the number of con-
tinuous variables or columns is even more dramatically reduced by
77%. The number of nonzeros in the constraint matrix is corre-
spondingly reduced by 61%. Thus, matrix density has increased
or conversely, sparsity has in fact decreasedit has become less
sparse after presolve.
The logistics solution was then used as input to the quality opti-
mizer where the same 1-hr time period and 240-hr time horizon were
used to generate the quality time profiles. The quality solver took
approximately two seconds to solve. In this case study, only an LP is
required to solve the quality optimization given that no flows were
adjustable and hence, no nonlinearities present; the lower and upper
flowrate bounds are equal. Figs. 4 to 6 trend the profiles of whole
crude oil/specific gravity, kerosene/pour point and heavy gas oil/sul-
fur cut/properties respectively for only those flows leaving the charg-
ing tanks and entering the pipestill; we do not show any internal flow
or tank qualities. The black line is the actual trace of the cut/property
and the blue line is the planning proxy target found in Table 3.
For the whole crude oil/specific gravity we observe a step-type
function that is due to the business practice of preparing mixes of
crude oils in the feed tanks and then charging that mix from one
tank at a time, emptying each tank before swinging to the other.
The result is an approximate 19-hr run length given the feed tank
capacity and pipestill charge rate. The approximately 0.011 maximum
excursion from the proxy would be improved by using a 50:50 recipe
on the transferline. The 50:50 recipe is driven simply by the fact
that the light and heavy crude oil segregations for whole crude oil spe-
cific gravities would mix to 0.862 if there were also 50:50 mixes in the
receiving tanks of each appropriate crude oil. Unfortunately, this
recipe would cause undue variation in the other two qualities. Because
there are only two receiving tanks and three qualities, and total flow
to the pipestill to potentially respect, at most we could only reason-
ably control two of the variables. Since throughput is rarely sacri-
ficed for overall refinery stability and profitability, only one quality
could potentially be controlled. The other two qualities would dis-
play an offset from target (i.e., only one quality can possess reset or
integral action in the context of control theory).
The kerosene/pour point trend does not show any obvious cycle
or periodicity as seen in Fig. 4 and there is a large excursion from
target between hr 60 and 80 when crude oil #2 is delivered at hr 7,
with a pour point of 42, starts to percolate through the blendshop
to the pipestill. If it were the most important quality bottleneck,
there would be four avenues to reduce this variability. The most
powerful effect is to change the delivery schedule of crude oils to
better manage the pour point quality. This is not always possible
nor is it an option unless sufficient lead time is available to the crude
oil traders and procurers. The second avenue would be to focus on
a segregation recipe that would better control the pour point to the
planning or operational target, although as mentioned this would
be at the expense of the other two qualities. The third approach
would be to alter the segregation policy. The current policy is based
on the whole crude oil/specific gravity whereby crude oils #1 and
#2 are deemed to be heavy and crude oils #3 and #4 are deemed to
be light. If for example kerosene/pour point is the quality of most
importance then it would seem prudent to segregate crude oils #2 and
#3 together as a low pour segregation and crude oils #1 and #4 as a high
pour segregation. This type of analysis can be found in more detail in
Kelly and Forbes.
2
PROCESS/ PLANT OPTI M I ZATI ON SPECIALREPORT
HYDROCARBON PROCESSING JUNE/JULY 2003
0
-42
-40
-38
-36
-34
-32
-30
-28
-26
20 40 60 80 100 120
Time, hr
K
e
r
o
s
e
n
e

p
o
u
r
p
o
i
n
t
140 160 180 200 220 240
Trend of crude oil kerosene pour point specif ic gravit y. FIG. 5.
0
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
20 40 60 80 100 120
Time, hr
H
e
a
v
y

g
a
s
o
i
l

s
u
l
f
u
r
140 160 180 200 220 240
Trend of heavy gas oil sulf ur cut propert y. FIG. 6.
TABLE 5. Logist ics opt imizat ion problem st at ist ics.
# Rows # Columns # Non zeros # 0 1 Variables # SOS1
16391 33025 79611 1500 9540
6814 7651 30745 1500 9540
The fourth avenue along the lines of changing the segregations is
to add a third or even fourth receiving tank. This would be an expen-
sive alternative but may provide a level of flexibility and controllability
well worth the investment. For instance, if a third tank is added then
instead of only being able to control theoretically without offset one
quality, we would now be able to control two qualities. With four
tanks we would be able to control without offset all three qualities.
The somewhat intangible benefit of controlling more qualities implies
that the downstream processes will have less disturbances to battle. A
more tangible benefit quantifiable by the planning optimizer would
be the ability to ride closer and closer to the real refinery constraints
or quality bottlenecks as shown in Fig. 2.
Finally, we show the quality profile for the heavy gas oil/sulfur.
An interesting consequence of this trend is the periodicity or cycle
of the variation (i.e., up-down-up using the blue line as the datum).
It appears to be in the range of 100 hr for this set of cycle data. This
means that from a production standpoint, the heavy gas oil inter-
mediate tankage must have sufficient capacity to store 100 hr
worth of heavy gas oil production. The reasoning behind this is that
given the relatively uncontrollable quality variation, due to the
inherent delivery schedule disturbances and limitations in the
blendshop, we need up and down or positive and negative varia-
tion around the planning target over some time frame to acquire
reasonably on-specification quality in the intermediate tanks. The
best alternative of course is to have constant or steady quality (i.e.,
the blue line) since shipment or blending of the intermediate can
be performed at any time during production and there will be
less likelihood of off- or over-specification product.
The next best thing is to have as short a perturbation cycle as
possible. And for this example, it would seem that qualitatively
we should have larger capacity intermediate tanks for kerosene/pour
point than the heavy gas oil/sulfur due to the irregular nature of the
pour point trend. This ultimately implies that the more a given cut
varies in quality variation the more tankage is required to buffer it
so that it is more consistent for blending or as charge to a down-
stream process unit. Minimal tankage will cause sharp swings in
quality forcing controls to react to the upsets.
It should be emphasized that scheduling is an important deci-
sion-making tool not just to create pro-forma operational sched-
ules but to also help answer tactical business questions such as
can we run feasibly by trading cargoes and delaying a crude oil
delivery of Arabian light by two days or can we accept another
crude oil delivery three days from now of Arabian heavy to fill
out our fluidized catalytic cracking unit? In the same way plan-
ning system users have more than just one planning model such as
for facilities, budgetary, feedstock selection and operation, so can
many different types of scheduling models be employed to answer
these questions timely and accurately. HP
LITERATURE CITED
1
Norman, A., K. Mahmood and M. Chowdhury, The need for a paradigm for
innovation, http://www.eco.utexas.edu/homepages/faculty/Norman/long/
InnParadigm.html, Department of Economics, University of Texas at Austin,
August, 1999.
2
Kelly, J. D. and J. F. Forbes., Structured approach to storage allocation for
improved process controllability, AIChE Journal, 44, 8, 1998.
3
Bitran, G. R. and A. C. Hax, On the design of hierarchical production plan-
ning, Decision Science, 8, 28, 1977.
4
Pardalos, P. M., and M. G. C.Resende, (editors), Handbook of Applied
Optimization, Oxford University Press, London, U.K., 2002.
5
Lee, H., J. M. Pinto, I. E. Grossmann and S. Park, Mixed-integer linear pro-
gramming model for refinery short-term scheduling of the crude oil unload-
ing with inventory management, Industrial Engineering Chemistry Research,
35, 5, 16301641, 1996.
6
Shah, N., Mathematical programming techniques for crude oil scheduling,
Computers & Chemical Engineering, 20, Suppl. B, S1227S1232, 1996.
7
Jia, Z., M. Ierapetritou and J. D. Kelly, Refinery short-term scheduling using
continuous-time formulationcrude oil operations, Industrial Engineering
Chemistry Research, February 2002.
8
Jain, V., and I. E. Grossmann, Algorithms for hybrid MILP/CP models for a
class of optimization problems, INFORMS Journal of Computing, 13,
258276, 2001.
9
Graves, S. C., A review of production scheduling, Operations Research, 29, 4,
646675, 1981.
7
Jia, Z., M. Ierapetritou and J. D. Kelly, Refinery short-term scheduling
using continuous-time formulationcrude oil operations, Industrial
Engineering Chemistry Research, February 2002.
8
Jain, V., and I. E. Grossmann, Algorithms for hybrid MILP/CP models for
a class of optimization problems, INFORMS Journal of Computing, 13,
258276, 2001.
9
Graves, S. C., A review of production scheduling, Operations Research, 29,
4, 646675, 1981.
10
Kelly, J. D., The necessity of data reconciliation: some practical issues,
NPRA Computer Conference, Chicago, Illinois, November 2000.
11
Belvaux, G. and L. A. Wolsey, Lot-sizing problems: modeling issues and a
specialized branch-and-cut system bc-prod, CORE Discussion Paper
DP9848, Universite Catholique de Louvain, February, 1998.
12
Clark, A. R. and S. J. Clark, Rolling-horizon lot-sizing when setup times are
sequence-dependent, International Journal of Production Research, 38, 10,
22872308, 2000.
13
Nemhauser, G. and L. A. Wolsey, Integer and Combinatorial Optimization,
John Wiley, New York, 1988.
14
Williams, H. P., Model Building in Mathematical Programming, 3rd Edition,
John Wiley & Sons, 1993.
15
Wolsey, L. A., Integer Programming, John Wiley & Sons, New York, 1998.
16
Kelly, J. D., Smooth-and-dive accelerator: a pre-milp primal heuristic
applied to production scheduling problems, Computers & Chemical
Engineering, 27, 827832, 2003.
17
Kelly, J.D., Chronological decomposition heuristic for scheduling: a divide
& conquer method, AIChE Journal, 48, 29952999, 2002.
18
Clark, A. R., A local search approach to lot sequencing and sizing, IFIP
WG5.7 Special Interest Group on Advanced Technologies in Production Planning
and Control, Florence, Italy, February 2000.
19
Palacios-Gomez, F., L. Ladson and M. Enquist, Nonlinear optimization by
successive linear programming, Management Science, 28, 10, 1106, 1120,
1982.
20
Adya, N., M. Tawarmalani and N V. Sahinidis, A Lagrangian approach to
the pooling problem, Ind. Eng. Chem. Res., 38, 5, 19561972, 1999.
SPECIALREPORT PROCESS/ PLANT OPTI M I ZATI ON
HYDROCARBON PROCESSING JUNE/JULY 2003
J. D. Kelly is a chemical engineer and has a masters degree
in advanced process control from M cM aster University. He has
worked as an advanced control engineer at both Shell Canada
and Imperial Oil including implementing real-time optimization programsand tac-
tical planning and scheduling solutions in their refineries. M r. Kelly has installed
plant-wide data reconciliation packages in several oil refineries around the world
and he haswritten many academic publicationson the subject. He isnow a solu-
tions architect for advanced planning and scheduling at Honeywell Industry
Solutionsin Toronto, Canada.
J. L. Mann isa chemical engineer with a bachelor of applied
science degree from the Universityof Toronto. He hasworked as
a design engineer and as a simulation engineer at Imperial Oil,
including developing refinery-wide simulation tools to support planning and
scheduling activities. M r. M ann hasworked on a number of plant information sys-
tem projects with a focus on integrating plant data collection systems to plant-
wide yield accounting systems. He now is a business architect for Honeywell
Industry Solutionsin Toronto, Canada.