Академический Документы
Профессиональный Документы
Культура Документы
Aseema Singh
A thesis submittea i n conformity w i t h the requirements for the degree of 4Iaster of Applied
Science Department of Chernical Engineering and Applied Chemistry
University of Toronto
I*m
Canada
Canada
The author has granted a nonexclusive licence allowing the National Libraxy of Canada to reproduce, loan, distribute or sell copies of this thesis in microform, paper or electronic formats.
L'auteur a accord une licence non exclusive permettant a la Bibliothque nationale du Canada de reproduire, prter, distribuer ou vendre des copies de cette thse sous la forme de microfiche/nlm, de reproduction sur papier ou sur format lectronique.
L'auteur conserve la proprit du droit d'auteur qui protge cette thse. Ni la thse ni des extraits substantiels de celle-ci ne doivent tre imprims ou autrement reproduits sans son autorisation.
The author retains ownership of the copyright in this thesis. Neither the thesis nor substantial extracts fiom it may be printed or otherwise reproduced without the author's permission.
Abstract
This thesis provides a new approach to gasoline blender control that can out-perform conventional blend controllers. The proposed blend controuer includes appropriate noniinear bleiiding modcls and adopts concepts froni mode1 predictive control theory in order to allow it to liandlc stochastic dis t urbances in feedstock quali t ies. The resulting Real-Time Optimization (RTO) system is similar to model predictive control in that it predicts disturbances over one time horizon and optimizes the blender control problem over another. It then irnplenieiits coiitrol action for tlic current tiinc-step aiid repeats the process in a recediiig horizon fasiiiori. r\iiotli(:r irriportaiit coritribiitiori is tlint of parnrrictcr observability for stcady-statc RTO systciiis. Tlic ciirrcritly nviiiablc mctliod providcs only n iiecessary biit riot sufficient coiiditioii.
A iicw ;rpproncli lias 11ce1i prescntcd wliicii uses fundainentai statistical priiiciples nxid is
~ippli~~:il~li! t.o ;uiy stcady-stntc systciii wlicrc secoiidary measiirements arc used to est iiiintc uiiiiic;\siii-(*cl (liiaiitities. Obscrval>ility is tlicii cxtciided to a . iiicasiirc of degrce of observability.
Acknowledgments
I would like to thank my supervisor, Dr. J. F. Forbes, for his valuable ,gidance and great
patience during this work. 1 also wish to express my deep appreciation to the Department of
Chernical and Materials Engineering, particularly the Cornputer Process Controi group, at the
University of Alberta. F i n d y , a special word of thanks goes to the members of my family for
their support and encouragement t hroughout t his project.
iii
Contents
Abstract
11
..
Acknowledgements
iii vii
vi ii
1
9
1.1
RTOSTRUCTT,'RE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
...............................
5
9
1.3.1 B L E n COXTROLLERS . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4
1 . 5 THESIS CO>TTSTIOSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
15
STE\tTt91yTMETHOD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
18
TReVSFOR\L\nOX SIETHOD . . . . . . . . . . . . . . . . . . . . . . 20
2.1.6
2.1.7
EXCESS JIETHOD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Z M E D METHOD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
24
2.1.8
. . . . . . . . . . . . . . . . . . . . . . 26
INTERACTION METHOD . . . . . . . . . . . . . . . . . . . . . . . . . . 30 OTHER .ME THODS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 SUMMARY OF R V P BLEXDNG MODELS . . . . . . . . . . . . . . . . 31 INTERACTION METHOD . . . . . . . . . . . . . . . . . . . . . . . . . . 32 EMPIRICAL MODELS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
35
. . . . . . . . . . . . . . . . . . . 50
58
61
BLENDING PROBLEN . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
. . . . . . . . . . . . . . . . . . . . . . . . . . . 67
. 68
CONVENTIONAL CONTROLLER . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4.1
4.5
74
4.6
TIMEHORIZON CONTROLLER . . . . . . . . . . . . . . . . . . . . . . . - . . 76
4 . 6 . 1 TIME-HORIZON CONTROLLER PERFORMAVCE STUDY . DISCUSSION . . .
4.7
5
. .
..... .. ...........
80
82
84
Bibliography
Appendix A
Appendix B Appendix C Appendix D Appendix E Appendix F Appendix G Appendix H
108
101
List of Figures
1-1 RTO Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1-2 Refinery Flow Diagram (blender feedstocks indicated in boId) . . . . . . . . . . . 1-3 Gasoline Blending Process
3
6
CT
4 1 Blending Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
LP
+ Bias Controller
. . . . . . . . . . . . . . . . . . . . . 72
83
vii
List of Tables
2.1 Cornparison of Octane Blending Models . . . . . . . . . . . . . . . . . . . . . . .
2.2
26
Production Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Feedstock Economic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
63
viii
Chapter 1
INTRODUCTION
The econornic benefits of process control are weli recognized in the chernicai and petrochemical industries. These benefits have led to the widespread use of advanceci automation technologies in such plants. The control system in these plants are designed to operate the process using some pre-determineci policy, which may or may not be optimal. Additional benefits, which can be crucial in providing a cornpetitive edge, can be reaiized by operating the process in
an economicdy optimal manner. T h i s is usually achieved by the optimization of steady-state
operations policy [Forbes, 19941. The major categories of steady-state optimization methods are direct search and model-based optimization methods. Direct search methods 1e.g. McFarlane and Bacon, 19891 work on the actual process and involve plant experimentation to obtain new steady-state operations. The direction in which process economics are improved can be deduce iom the experimental results and the process optimum is found by iteratively rnoving the process in this direction after each plant experiment. However. the amount of experimentation required, especiaiiy for large dimensional problems, can be prohibitive. Moreover, for complex and integrated plants a i t h slow dynamics, it can take a long time For the process to reach steady-state after each perturbation. Using direct search methods on such plants would require the process to operate outside the normal region of operation for long periods of time, which may not be acceptable. Model-based optimization methods, on the ot her hand, work on mat hematical models t hat represent the plant, rather than the process itself. Optimization algorithms are used to find the optimum of the model, the results of which, after verification, are implemented on the plant.
RTO technology
is gasoline blending [Leung, 19851. This thesis considers RTO systems for gasoline blender control and offers solutions for improving blender efficiency. In doing so, obsenability issues in conventional steady-state RTO systems are also addressed.
1 . 1
RTO STRUCTURE
The structure of a typical RTO system is shown schematicdy in Figure 1-1 [Seferlis, 19951.
modek in the optimizer. 2. Data validation: at this stage, process measurements are examined to ensure that the
process has indeed reached steady-state. Also, Cross Error Detection and Data Reconcil-
iation are carried out. G r o s Error Detection involves detecting errors in measurements resulting from sensor failliles, leaks, etc., so as to eiirninate errornous data from being used
in mode1 updating. Data Reconciliation tries to adjust the data such that the material
3. Model updating: validated data is used to update the model in some way. Model updating involves obtaining new estimates of sorne of the model parameters using estimation techniques such as regression analysis.
4. Optimization: model-based, steady-state optirnization is used to calculate controiler set-
points. The optimization algorithm can be a linear program (LP) or a nonlinear program
-_i, 7
Controllen
Validation
Sensors
Process
(NLP) and, may require mixed integer programming (MIP) depending on the nature of
the mode1 equat ions.
5. Cornmand conditioning (also called post-optirnality or results analysis): at this stage, the
outputs from the optimizer are examined in order to ensure that only acceptable setpoints which wili result in a significantly improved plant operation, are passed t o the controllers [Seferlis, 19951.
6. Control: the acceptable setpoints are used by the controllers to implement the operations
policy on the plant. The controller structure can range from simple PID loops to advanced control schemes such as mode1 predictive control (MPC). More detailed treatments of the RTO structure can be found in Seferlis [1995] and Forbes [1994].
1 . 2
RTO DESIGN
The performance of any RTO system will depend on the characteristics of each of the constituent subsystems and their interactions with each other [Forbes, 19941. Al1 the components in the
RTO system depicted in Figure 1-1have been studied t o some extent by various researchers.
Krishnan [1990]has studied the problem of selecting appropriate measurements for model u p
dating. A measurement selection strategy based on minirnizing error in estimated parameters
has b e n provided. Issues in Gross Error Detection and Data Reconciliation have been addressed by Crowe [19881, Rosenberg e t al. [1987], and Tjoa and Biegler [1990]. Crucial to the success of the RTO systern i s the quality of models used on the optimizer. Biegler et al. [1985] have found that a model is adequate for use in RTO if model parameter \dues
can be found such that the gradients of the model with respect to the decision variables can
be forced to match those of the plant. Mode1 selection has also been studied by Forbes et al.
(19941 who have defined model adequacy in ternis of the model's ability to possess an optimum
at the true plant optimum. That is, a model is defined to be point-wise adequate if there exist
a set of values for the adjustable model pararneters such that the model-based optimization
possesses an optimum at the true plant optimum. The model adequacy approach taken by Forbes e t al. (19941d o w s for the selection of adjustable model parameters. That is, model adequacy can be used as a criterion for selecting adjustable model parameters. Sirnilarly, Roberts [1979], [1995] has stated t hat models and parameter estimation strategies should not be selected in isolation; the characteristics of the process mode1 should be taken into account in order to allow the system to effectively handle plant-mode1 mismatch. Krishnan [1990] has also looked at the problem of selecting the adjustable parameters in the model to be updated on-line and has suggested updating those parameters that have the most effect on the optimal value of the objective function or alter the active constraint set. Observability of the adjustable parameters from the available measurements has been studied by Krishnan [1990]. Optimization algorithms have b e n developed for several classes of optimization problems such
as LP, XLP, and MIP, and are being continually improved for increased computational el-
ficiency. Optimization theory and algorithms c m be found in Fletcher (19871, Avriel [1976], Balakrishnan and Boyd [1992], Duran and Grossman [1986], and Edgar and Himmelblau [1988]. Post optimality analysis has been studied by Seferlis [1995] and Miletic [19961. Tests have been provided to determine whether or not setpoints generate by the optimizer are to be passed to the cont rollers.
While most of the work has focused on individual components of the W O system, the interactions among the subsystems and their effect on the whole structure have received Little attention. Forbes and Marlin [1996] have provided the first approach to making F L T 0 decisions
based on the whole system. The approach is based on designhg RTO components such that
the loss of RTO performance i s minimized. The measure of RTO performance used by Forbes and Marlin [1996], the Design Cost Criteria' looks a t the loss of econornic opportunity resdting frorn a given RTO design. This mesure considers the deviation of the manipulated variables from the true plant optimum. An alternate approach due to Loeblin and Perkins [19961 uses the average deviation of the objective function rather than the setpoints. The past sections have focused on RTO system without application to any particular process. Since this thesis is concerned nrith RTO of gasoline blending, gasoline blending will be discussed
next .
1 . 3
GASOLINE BLENDING
Gasoline is one of the most important refinery products since it can yield 60 to ?O percent of a typical refinery's total revenue [DeWitt et ai., 19891. Fractional distillation of crude oil produces
several "cuts". Only a small portion of the distillation products, the light virgin naphtha or
light straight run (LSR) naphtha, is suitable for blending directly into gasoline. Other fractions have to be further processed. A simpiified flow diagram of the refinery processes is given in Figure 1-2. The flow diagram shows the main units and the refinery processes required for producing gasoline feedstocks in a typical refinery. Some of the refining processes are briefly described below [Givens, 19851:
1. Catalytic Reforming: this is an upgrading process that increases the octane nurnber of
the heavy virgin napht ha by catalytically convert ing low octane hydrocarbon into high
octane ones. The most important reactions taking place include the conversion of rnethyl cyclohexane (research octane number of 75) and n-heptane (zero octane by definition) to toluene which has a research octane of 120.
2. Catalytic Cracking: this is a co~versionp r o c w that breaks large, low volatility molecules into smaller, more volatile
Mixcd Butanes*
Batma
Alkylik
ButylLight Virgin N a p h h
.
b
Niphth.
UR
Virgin DiniIlau
n
I
Vacuum Gas Oil
c
oil
Cracking Caralytic
bcaulytic Cu.
Huw Cycle 0il1
Coke
Distillation
TL
Rcducui Cnidc
Reductd Cnide
Asphah
Figure 1-2: Refinery Flow Diagram (blender feedstocks indicated in bold) distillation colurnn, that separates the stream according to boiling range. About 50% of the catalyticdy cracked stream can be blended into gasoline without hrther upgrading. Another 20% contains light gases from which n-butane and isebutane are separated and used later for blending in gasoline.
3. Alkylation: down stream from the catalytic cracker, the alkylation unit further processes
light olefins produced during catalytic cracking. Although the light olefin stream has high octane quality, its boiling range is too low for gasoline. Alkylation involves catalytically reacting these olefins Rrith iso-butane to form high quality hydrocarbons such as i s e octane. 4. Hydrotreating: feedstocks (including straight run naphtha frorn the crude distillation unit) are treated with hydrogen in the presence of catalysts to reduce the amount of sulphur, nitrogen, and oxygen. Some of these refinery products are further separated into light and heavy fractions in order to provide more ffexibility in blending. A large refinery can have more than 20 blender feedstocks
that are blended into several grades of gasoline [Civens, 19851. Gasoiine blending i s the process of combining these blender feedstocks, together with s m d amounts of additives (such as antioxidants, corrosion inhibit ors, met a l deactivators, detergents, and dyes) , to make a mixture meeting certain quality specifications. S pecifications on gasohe qualities such as octane number, volatiiity, sulphur content, aromatics content, and viscosity are placed in order to ensure acceptable engine performance as well a s for environmental reasons. Feedstocks can be blended out of storage tanks, intermediate tanks into which upstream processes are being fed. or directly kom upstream processes (Agrawal, 19951. The gasoline blending process of blending feedstocks frorn storage tanks, is shown in Figure 1-3.
LSR Naphtha
Reformate Regular
Additives
Figure 1-3: Gasoline Blending Process Octane numbers are measures of a fuel's antiknock properties. Knocking occurs when a fuel/air mixture ignites prematurely in the cylinder, producing a *'knocking" sound in the engine. Knocking not only reduces the engine's power by working against the piston, but it also exerts mechanical stress on the engine parts. The octane number of a fuel is defined as the percent of iso-octane (2,2.4-trimethyl pentane) in a blend with n-heptane that exhibits the sarne resistance to knocking as the test fuel under standard conditions in a standard engine [Palmer and Smith,
1985). On the octane scale. iseoctane is assigned an octane number of 100 and n-heptane a
number of O. Two standard test procedures are used to characterize the antiknock properties of
fuels for spark engines - the ASTM D-908 test gives the research octane number ( R O N ) and the
ASTM D357 gives the motor octane number (MON). The RON represents antilcnock properties under conditions of low speed and frequent accelerations while the h1ON represents engine performance under more severe high speed conditions. The octane number posted at gasoline stations is the arithmetic mean of the RON and MON
19941. The main difference berneen automotive gasoline grades is their antiknock properties. For example. regular and prernium gasolines are specified to have posted octane numbers of 87 and 92, respectiveiy. Other important properties that afTect engine performance are volatility and boiling range.
A vapour pressure which is too high for the given ambient temperature wili resuit in vapour
stalling ana motor locking, while a vapour pressure which is too iow will lead to difficulties in engine startup [Palmer and Smith, 19851. The boiling range also affects the engine during
startup and driving, and is particularly important for good performance during quick accel-
eration and high speed operatiom. Like vapour pressure, the boiling range has to be tailored
for a given geographicd region and season. There should be a fairly reasonable distribution
of light (more volatile), intermediate, and heavy (relatively non-volatile) components: however, gasolines blended for use in cold climates need more iight components (front end volatility) than t hose intended for use in w m e r climates [Palmer and Smith, 19851. In addition, governrnent regulations place maximum restrictions on the vapour pressure to limit the emission of volatile organic compounds into the atmosphere [Lo. 1994j. Other environmental restrictions include
1 . 3 . 1
BLEND CONTROLLERS
Blend controliers are used for making gasolines that meet quaiity specifications whiie satisfjring
demand and availability limits. In addition, the controller shodd work to rnaxirnize profit by
using the least expensive feedstocks to make the most valuable products, minimizing "quality giveaway", and minimizing reblends. "Quaiity giveaway' refers to making gasoline with quaiities that either exceed the minimum required or fa11 below the maximum allowed. thus reducing the profit margin. It is estimated that consistent octane giveaway of 0.1 can cost a refinery several millions of dollars a year [Givens, 19851. Reblending can also lead to significant cwts to a refinery by taking up valuable tank space and blending time, reducing overd capacity Thus. the controller shauld be able to blend the products optimally and get the blends on specification the first time. There is a wide range of sophistication ic blending control technology presently employed by refiners. The typical state-of-the-art blending technology is built on t h e e levels: off-line optirnizer or scheduier, on-line optimizer, and regulatory control [Sullivan, 19901, [ Agrawal, 19951. The leveis of control strategy are shown schematicaily in Figure 1-4. At the top of the hierarchy lies the planner or scfieduler which uses off-Iine optimization to plan retinery operations for the long range (months), intermediate range (weeks). and short
range (days). Long-range planning can be performed at either the refinery or corporate level.
Long range forecasts of product dernands, crude oil prices, and process unit performances are used to plan refinery operations several months to a few years into the future [ S u l l i ~ n19901, .
(such as additional constraints) are used at t his stage to revise refinery plans as well [Sullivan,
19901. Short-range planning works at the unit level, rather than at the refinery level, and plans
blender operations over a shorter time frame (1 to 2 days) [Agrawal, 19951. At this stage, assumed feedstock qualities (or measured qualities, if available) are used to produce blend
Econornics inventory
On-Iine op=t
setpoints
iI
I
,
Controller
.
k-
Figure 1-4: Blending Cont rol Hierarchy recipes which are then downloaded to the on-line optimizer. Usuall. all levels of planning using an off-line optimizer are conducted using linear programming [Sulli\xn. 199O]. [Rarnsey Jr. and Truesdale. 1990]. (DeWitt e t al.. 19891. The on-line optimizer uses on-line blended quaiity information to rnodify the initial recipe during the blend and provides h a 1 blend recipes to be executed by the controller. On-line optirnizers are designed to minimize deviation from the initial blend recipe or optimize some economic performance (such as profit mauimization) while satisfjnng al1 constraints [Michaiek
e t aL. 19941.
However, the move has been towards optimizing steady-state process economics.
Thus the on-line optimizer is usually a conventional RTO system. Vermeer et al. [1996] have expanded on the conventionai steady-state RTO to perform dynarnic blend optimization. This is achieved by including blending dynamics into the controller t o calculate expected blended qualities and using these expected qualities as feedback to the LP based on-line optimizer. Incorporating blending dynamics in such a way allows the RTO layer to operate at a higher frequency.
The blend recipes generated by the on-line optimizer are then implemented at the control level. Final blend control is achieved through conventional distributed control systems (DCS) .
The DCS-resident controiler (usually PID loops) maintains flow ratio to match the recipe. maintains a target flowrate for the sum of all feedstodc flows, and stops when the required amount of product has been blended [Ajgawal, 19951. The control layer is aiso used t o ensure that quality specifications are met. It uses analyzer feedback and can deviate from the blend recipe determined by the on-line optimizer in order to keep the products on specification. The different subsystems in Fiewe 1-4 have to be integrated to provide information Aow (such
as tank and feedstock inventory) necessary for the success of the whole system. Databases are
used to keep track of feedstock inventory, tank availabiiity, feedstock qualities, and blended quaiities. The subsystems m i t e to and/or retrieve information rom these databases, thus linking the subsysterns into an integrated network [Agrawal?1995j, [Ramsey Jr. and Truesdale. 19901, [Palmer et al., 19951. [Ifichalek et al., 19941. At the heart of the entire blending control system shown in Figure 1-4 is the RTO Iayer which determines the final blend recipe and so the performance of the RTO system is crucial to the performance of the whole gasoline blender control effort. The design and performance of the gasoline blender RTO system is the focils of this t hesis.
The on-line optimizer is usually based on linear programming (e.g. Benefield and Broadway,
[1985], Diaz and Barsamian, [1996]?White and Hall, (19921. McDonald et al-,[l992], Serpemen
et al.' [1992]: Slichalek et al.. [1994]) even though most important gasoline properties have
been shown to blend in a nonlinear fashion [Rusin, 19751. -4s a result, blender control systems must deal with structural mismatch. Some refiners have tri& addressing this problem by using nonlinear models and employing nodinear programming (hTP) in the on-line optimizer (e.g.
Ramsey Jr. and Truesdaie? [1990]): however, linear programming remains as the predorninant
technolog-. oaing its popularity to reliability and ease of use [Magoulas et al., 19881. Another approach that has been ernployed is to linearize the nonlinear models and use sequential linear programming (SLP)(e.g. Diaz and Barsamian, j19961). In an attempt to maintain accuracy of the rnodels in the optimizer, the feedback ernployed is usually "bias updating" [e-g. Diaz and Barsamian, [1996], Bain et a l , [1993]). Bias updating
involves comparing measured blended qualities with those predicted by the models in the onIine optimizer; the difference between the two i s then added as an error term to the appropriate
blending models. Thus the RT0 system is most often formulateci as iinear programming ~ 4 t h
bias updating.
This approach of iinear programming with bias updating has proven quite successfui in practice
and is the basis of most commercially available biend control systems. Its success was largely due to the practice of blending f r o m well-mixed storage tanks. The current trends of in-line biending and blending from "running" tanks presents a problem that such a control structure cannot adequately address. This i s because upstrearn proces variations result ing fkom cat al-yst deactivation, heat exchanger fouling, etc.. will cause variations in the feedstock qualities. The
LP (or XLP) plus bias update formdation assumes constant feedstock qualities and so can
not adequately handle such time mrying feedstock qualities. Forbes and Mariin j1994j have shown that this traditional blend controlier approach can lead to a substantial loss in blender profitability given eLrena srnall cariation in feedstock qualities. Xote that in the case of tirne-varying feedstocks, the bias updating approach can still be used
to make products because the lou-er level feedback controllers in the RTO system
d ldek-iate
from the blend recipe determinecl by the on-line optimizer in order to meet specifications a s discussed above. Although products can be made in this manner to meet specifications, the resulting blend u-ill not necessarily be optimal in that the best cornbination of feedstocks may not have been used. Some refiners have tried addressing this problem by feedback of laboratory anal-mes of feedstock qualities and then using multiperiod optimization (e.g. McDonald et al., [1992] and Rigby et
al.. [1995]).The i n t e n d between optirnizations is usually a day and feedstock qualities are
assumed to remain unchanged during these penods. Thus this approach cannot effectively deal with higher fiequency disturbances arising Erom upstream process operation changes. Vermeer
et a l . [1996] have used blended quality measurements to predict disturbances in t hese qualities.
The disturbances, however, are assurneci to he step disturbances that remain constant during the optimization intenal. Therefore. this approach s d e n from the same drawback as the multiperiod opt imization strategy.
This t hesis wiU address the probiems associated with structural mismatch and inability of
the bias update approach to adequately handle fluctuations in feedstock qualitis. A new control approach that can incorporate nonlinear blending models and can deal with t i m e - w i n g qualities will be developed.
1.4
This thesis is concerned with the design of the blend controiler shown in Figure 1-4. and the
F E 0 layer in particuiar. A s has been discussed in the previou section. the W O iayer is
most often formulated as Iinear programrning with bias updating. The LP plus bias updating formulation suffers &orn plant /mode1 mismat ch since most important gasoline quaiit ies blend
s not capable of adequately handling disturbances in nonlinearly. Generaliy, the controiler i
feedstock qualities [Forbes and Mariin. N94!. The objective of this thesis is to proiide a blend controller t hat incorporates models which capture the inherent nonlinearities of the blending process. and one that can effectively deal with predictable kariations in feedstock qualities. Gasoline blending models are examineci in Chapter 2 with respect to predictive accuracy and their ability to be implemented for control purposes. Once scitable models have been selected. it is crucial that their predictive accuracy be maintainecl by updating some parameters on-line. Therefore it i s important to ensure t hat the adjustable parameters in these rnodels be obsenable rom the process measurements and the updating scheme. Chapter 3 look at parameter obsenability for steady-state systems. A measure of observability for steady-sr ate systems is dewloped and the concept is evtended to degree of obserrability. The observability measure is based on fundamental statistical principles and can be applied to the blending rnodels. The observability measure presented in Chapter 3 is not limited to the blending process or to RTO systems, but is applicable to any steady-state system where unrneasured quantities are to be
est imated using available measurements.
Chapter 4 presents a new approach to blender control that can incorporate nonlinear models
and can effectivel~deal with feedstock quality variations. The control approach is analogous
to 3lPC in that a time-horizon approach is taken. A gasoline blending case-study i s used to dernonstrate the problems associated with the conventional bias updating approach and the benefits of adopting the new blending strategy.
Although the tirne-horizon strategy has been applied to the RT0 layer in this thesis, it codd also be adopted into the off-line optimization iayer and used to improve the effectivenes of the
planning and scheduiing layer. Moreover, it codd be applied to on-line or off-line optirnization
1 . 5
THESIS CONVENTIONS
In accordance with the petroleum industry terminolog-, light and heavy components refer to
relatively volatile and nonvolatile cornponents, respectively. Ali terrns are defined on first use
and used thereafter. The Nomenclature section lists all syrnbols employed in this thesis while
the acronynis used are summarized in the Glossary of Terms section.
Chapter 2
2.1
OCTANE NUMBER
Octane numbers indicate the antiknock characteristics of gasoline or the ability of the gasoline to resist detonation during combustion in the combustion chamber. The two most comrnon
types of octane nurnbers for spark engines are the research octane number (RON) and the
motor octane nurnber ( M O N ) . The RON, defined by the Arnerican Society for Testing and Materials under the designation ASTM D-908, represents the engine performance under city
driving conditions (fiequent acceleration) while the MON (ASTM D-357) represents engine
performance on the highway [Gary and Handwerk, 19941. Since RON and MON of a gasoline blend in nonnear fashion, complex blending models are needed for accurate prediction of blended octanes [Rusin, 1975].
2 . 1 . 1
In this method, fictitious blending octane numbers (BON's) are used for RON and M0.V.
The BON'S blend linearly on a volumetric basis to give the octane number of the blend [Gary
where: y is the volume fiaction of component i, (ON)Md is the octane number (RON or
MON) of the blend, (BON)iis the blending octane number (RON or II/fON) of the component.
n is the number of cornponents in the blend.
Blending octane numbers are u s u d y obtained from regression analysis on small data sets and
may be user-tuned. Thus their application is often based o n user experience and judgment
[Rusin et a l . , 19811. BON's for some typical components can be found in Gary and Handwerk (19941. The limitations of using BON's has long being recognized which h a . lead to the developrnent of more reliable methods that do not depend so heavily on user judgrnent [Rusin et a l ,
19811.
2 . 1 . 2
ETHYL RT-70METHOD
The Ethyl method is one of the oldest models available in literature. This method has been used
as a benchmark against which newer models have been compared. Here, blending nonlinearity
is modeied explicitly as a function of the component sensitivity (RON - MON), olefin content,
and the aromatic content of the cornponents as given below [Heaiy et al., 19591:
where: r is
RON,m i s MON, s
coefficients. Quantities accented with an overbar represent volumetric average. Equations (2.2) and (2.3) contain a total of six parameters
( a l ,an, a3, a4, as, a6)for
fiON and
i W N blending. In order to estimate these parameters, the following data would be needed:
RON, MON, olefin content, and aromatics content of each pure feedstock,
historical blend data. Thus! the mode1 can be made adaptive so that its accuracy does not decrease over time. However, the models require knowledge of the olefin and aromatics content which may not be readily available.
Some indication of the predictive accuracy of the models has been provided by Healy et al.
[l959]. The standard deviations of prediction errors for unleaded blends sirnilar to the ones the
coefficients were determined from are reporteci to be 0.66 and 0.85 octane numbers for RON and MON, respectively. For independent blends, the standard deviations of prediction errors were found to be 0.92 for RON and 0.61 for MON.
2 . 1 . 3
STEWART M33THOD
This method was proposed by Stewart [1959a] about the sarne time a s ihe Ethyl method was published and is siMlar to the Ethyl method in that blending nonlinearity is attributed to the olefin content of the components. The blending models are:
where:
D,= , 1*
cent o l e h content of the blend and i s the volumetric average of the cornponent olefin contents,
are constants.
Stewart [1959a]has detennined the constants and C using least-squares analysis on 102 blends:
i was determined t o be 0.01414 for RON and 0.01994 for M O N , the value of C the value of i
was determined to be 0.130 for RON and 0.097 for MON. However, oidy 10 of the blends used were multicornponent; the rest were binary blends.
The standard deviation of prediction error reported by Stewart [1959a] is 0.77 octane numbers
for RON and 0.64 numbers for MON. However, this prediction error reported is for prediction within the data set used for evaluating the correlation constants. No indication of the predictive accuracy outside the development data set is provided. Also, any indication of the level of industrial acceptance of this method could not be found.
2 . 1 . 4 INTERACTION METHOD
This method is based on the two-factor rnodels where the overall effect is attributed to the main effects of each of the two factors and the nonlinearity is captured by an interaction terrn. The interaction term accounts for the effect of one factor on the other and can be determined using the design of experiments (Montgomery, 1991]. Similady, in the Interaction method, octane blending nonlinearity is attributed t O the twefactor interactions arnong the components in the blend and is accounted for by adding an interaction term to the volumetric average octane nurnber. Only the twsway interactions between pairs of components are considered; the t hreeway and higher interaction terrns are generally ignored [Morris, 19'751, [Morris, 19861, [Morris
et al.? 1994]. For an n component system, the octane number (RON or M N ) of the blend
The interaction coefficient, l&, between a pair of components is obtained by using the octane
numbers of the pure components and that of a 50:50 blend of the two components as given
l . , 19941: below [Morris et a
where
for
RON and the same for MON, giving a total of n(n - 1) pararneters to be determined. The
experimental data required for estimating these parameters are:
a a
RON and MON for each pure component, RON and MON for 5050 blend of a i l component combinations.
Thus, for either MON or RON in an n component system, the total number of fuels to be tested are n pure cornponents plus
n(n 1
v.
T h e total
number of tests required for RON and MON is n(n + 1). Since octane blending characteristics may change with changes in octane level and other feedstock qualitieso the interaction coefficients may become outdated if significant changes occur in the octane pool [blorris, 19861. Thus the predictive accuracy of the Interaction model m a y decrease with time as the feedstock qualities change with changes in upstream operations or crude srnitches. However, unlike the Ethyl RT-70 method, the Interaction model c a m o t easily are not correlation coefficients but be made adaptive since the interaction coefficients (IiVk) are determined experimentaily. Instead of determining the interaction coefficients frorn binary blends, the interaction coefficients could be obtained using regression analysis on blend data (data on al1 feedstock qualities as well as blended qualities). The interaction coefficients could then be updated using historical blend data and made adaptive. However, Interaction model requires a large number of parameters to be determined. For example, a 7 component blend would require determination of 42 interaction coefficients while the Ethyl RT-70 method will require estimation of only 6 pararneters. Therefore, a large amount of blend data would be
&O,
be re-evaluated as new feedstocks are entered into the blend. Some indication of the predictive accuracy of the blending models have been provided in literature. Morris [1975] has studied the accuracy of the Interaction method and has compared it to the Ethyl RT-70 approach. It has been reported that whiie the Interaction models fit data more closely than the Ethyl modek (this is to be expected since the interaction model has many more parameters that the Ethyl one), the predictive accuracy for a generai untested data set using predetermined parameters is about the sa-ne for the two methods. The standard dellation of
l 9 7 5 1 to be between 0.6 to 1.0 prediction error for both the methods were reported by Morris I
octane nurnbers. While there has been some industrial acceptance of the Interaction models, they have not been widely used. This codd be due to the large number of parameters that need to be determined.
The transformation method involves transfonning the octane number, which blends nonlinearly,
to a quality that does blend linearly and ideally. This new quantity is then blended ideally and the result is transformed back to octane number [Husin et al.. 19811. This method is similar to
the Blending Octane Nurnber (BON) method in t hat transformed qualities are blended ideally; however. systemat ic methods of transforrning the octane nurnbers to the transformed qualities
and the blended transformed qualities badc to blended octane numbers are provided. While the
Transformation method is applicable to leaded as well as unleaded gasolines, only the unleaded
case will be presented here. Discussions on the application of the transformation method to
leaded gasolines can be found in Rusin et al. [1981]. For unleaded gasolines, t his method takes into account the interaction of major classes of hydrocarbons and the change in conditions (severity levels) inside the engine when measuring different hiels. The effect of severity levels are accounted for by relating component octanes and sensitivities to reference levels. The steps required for unleaded gasoline are outlined below.
Si
t=1
ROiv,,,_,
= RONtam,bl,nd - H R O N ~ ~ . , ~
where: ci,
cz,
CT,
cg,
and MOiV,,j are the reference RON and MON arbitrarily set at 90
[Rusin et al. 19811. H is effect of hydrocarbon type interaction, Ai, Pz, 0, are volume fractions
of aromat ics, p a r a h , and olefins for component i , respect ively. These blending models for RON and MON contain at least ten parameters (RON,.
ci, c2, c3. q,CS,
~ g ,c7,
f,
illON.,j ,
cg,
Predictive accuracy of the Transformation rnodels was tested by Rusin et al. [1981] who used data rom 564 gasoiines of varying compositions rom severai independent blending studies. They split the data from each source roughly into halves and used the first half for determining the model parameters. The second haif of the data set, together with data lrom two additional complete studies, was used to test the predictive accuracy of the rnodels. The standard de+ ations of prediction errors were found to be around 0.4 to 0.5 of an octane number for both
approach is much superior to the Transformation approach in terms of both mode1 simplicity
and ease of use.
2 . 1 . 6
EXCESS METHOD
This method de& with deviations hom ideality by applying an "excess" term. These excess values are added to the volurnetric average of the component octane numbers to predict the blended octane number (RON or MON) [Muller, 19921 as:
where: ( O N ) j is the octane nurnber of blend j, (ON): is the excess octane number associated with component i in blend j, and
uij
In this linear model, the excess octane number, (ON):, gives an indication of the contribution to nonideality in blend j by component i. These parameters for each blend are determined by preparing a series of "est-blends". The "rest-blends" are obtained by blending as many
sarnples of the fuel as there are components but omitting one component after another. Kote
that in preparing the "rest-blends" the volume of each component (except the omitted one) remains the same so the total volume and volurnetric composition vary. Detaiis of determining the excess numbers can be found in Muller [19921. The number of (ON)$ values to be determined for each fiel grade (blend) i s n and the experimental data needed to estimate these parameters are:
a
0
octane number (RON and MON) of each component, octane number (RON and MON) of the blend, octane number (Rollrand fi1O N ) of the n 'test-blends" .
The number of fuels that need to be tested are (n + d(n t 1 ) ) . Thus, a total of 2(n + d(n t 1)) tests would be needed for MON and RON combined, where d is the number of gasoline grades. The Excess met hod Ieads to linear blending models which has the adkantage of alloaing the use of linear programming (a-hichis widely used in industry [Magodas et aL. 19881) in determining the optimal blend recipe. However, the "excess" values are calculated for a given blend at its nominal composition. Therefore. the accuracy of this method will decrease as the composition deviates Erom this nominal point and nerv blending studies would have to be comrnissioned periodically to recalculate "excess numbers". Also, the parameters determined are specific to each blend. That is. "excess numbers" for different gades of gasoline (such as regular. prernium. etc.) have to be determined separately whereas for other models. the same set of parameters is applicable to ail gasoline grades. S o indication of the predictive accuracy or the level of industrial acceptance of this method codd be found.
This method correlates feedstock octane numbers to predict the blended octane number. The parameters of the equation are estimated using regression analysis on experimental data [Zahed
et a l . . 19931. The blending mode1 for ROM is:
blend.
Thus the parameters that need to be determincd are -Ido, dZl, ...' J i , , and k3 for R O S . and
Qa,
QI,
..-,Qn: and k4 for ?JlO:V.The data needed for estimating t h e pararneters are RO-\:
estimaring the parameters since blendhg characteristics and therefore mode1 pararneters may change as feedstock qualities changes over t h e . The rnodei parameters may be updated on-line using historical blend data. Zahed et <IL j1993! have eompared the predictive accuracy of the that of the Stewart method and have s h m n it to rnodels in Equations (2.9) and (2.10) ~ l t h
te better at predicting blended ROX for a set of expenmental data. Standard del-iations of
prediction errcr were not prorided.
2 . 1 . 8
Table 2.1 sumarizes blending models dixussed above and Table 2.2 compares the preictive
a c c u r q of the difhent models. Accufor interpolation (prediction aithin the development
data set j as well as extrapolation (prediction outside the development data set j are provided.
Surprisingly for the Transformation method. Rusin et al 1981: have reported a higher predic-
tion accu-
The Excess method. being lineu. is valid only u-ithin the vicinity of the nominal blend. .Uso. there is little information a~ailable with regards to the predictive a c c u r q of the Excess. Stewart or Z&ed methods. The Interaction method has been shown to be no more accurate than the Ethyl method i n predicting blended octar.es on general. untesteci data sets even though it has
many more parameters than the Ethyl rnodels. Xlthough the Transformation models seem to
provide slightly better predictive accu-
Mode1
Ethyl
Stewart Interaction
Parameters
al:a2,a3 Q4,aa.M a. c
# of Parameters
6 2 for ROX
\ f ~ ~
.
I
I
L,k
Cl. c2. c3. c4
"'"-'i
"("Lbr L 2
IO
kl, kz
Table 2.1: Cornparison of Octane Blending Modeis quality data and contain more parameters (les parsimonious). .\loreorer. the Ethyl models are
much more simple and easy to use than the Transformation models.
2.2
REID VAPOURPRESSURE
The Reid Vapour Pressure (RVP). defineci by the -4merican Society for Testing and Ilarerial.
under the designation ASTN D-323-56. giws an indication of the volatility of a gasoiine blend and is approximately the vapour pressure of the gasoline at 100F (3BC). The R V P of a gasoline blend &ects the gasoline performance in r e m of ease of starting* engine n-arm-up.
and rate of acceleration [ G a n and Handwerk. 19941. R V P is &O important because mavimurn
specifications on R V P Limit the amount of n-butane. a relatively cheap source of octane. that can be added to a blend.
2 . 2 . 1
THEORETICAL APPROACHES
One of the first theoretical approaches alailable is that by Stewart [1959b] who presented a met hod for predict ing blended RV P's by mat hemat ically simulating the processes occurring during the Reid test (which measures the equilibrium pressure for a constant-volume partial vapourizarion of the hiel in air). The Stewart approach uses component data (such as feedstock
1 Mode1
Ethyl
Stewart Int eraction Transformation
1
I
RON
Interpolation Extrapolation O-8e 0.99 O.77@ Y /A 0.28' 0.8' 0.51A 0.43A
1 Interpolation Extrapolation
0.79~
1
MON
0.61~
W.4 0 . 6 4 ' O.SA S/A Y/ -4
0.64~
0.33'
I
0.60'
Excess Zahed
S / , 4 S/ -4
&frorn Heaiy
et
S/A
X/A
a l [1359]
S/A S/A
S / A not available
Table 2.2: Predictive Accuracy of Onane Blending Modeis composition and component volatility ) and t hermodynamic relationships in simulat ing the test.
Several simplikng asjurnptions are made including: the presence of air and water vapour in
the test chamber are ignored. the ahsolute pressure i s taken as the RVP. volatile components are assumeci to have the density of butanes. and the nonvolatile components are assurneci to
The equations for estirnating the blended R V P provided by Stewart '1959bi are:
a-here: d, is the mole fraction of component i in the liquid phase. 7, is the acth-ity coefficient of component i. u, is the volume fraction of component i in the blend (feed) at 1 atm and 60
OF.
is the rnolar density of component i (pound-moles per barrel as liquid at 1 atm and
60 OF). pblmd i s the density of the blend as saturated liquid at 60 OF (pound-moles per barrel).
and k is the number of components in the b l e d (components over al1 feedstocks). The modeis given in Equation set (2.1 1) do not proxide (RVP)b(d explicitly as a function of component data: therefore. an iterative approach would have to be taken in olving for the blended RVP. The data required in using the Stewart method are RVP's of each feedsrock. volume fractions of a i l components in each feedstock. molar densities of all cornponents in each
feedstock, and estimates of activity coefficients of all components. Vazques et ai. [1992] have presented an algonthm for calcdating the blended RVP's based on the work of Stewart [1959b].The mode1 assumes additivity of Iiquid and gas volumes and.
as in the Stewart case, the algorithm is based on an air and water-Eree model.
T h i s approach
differs rom the Stewart method in the equation of state used and it requires that the molar composition of the feedstocks be knoun. An iterative procedure for calcdating the vapourliquid equiiibri~m for the flashing of the gasoline feed is provided. The steps required by this method are outlined below:
OFo
Step 3: Periorm "flash calculation" at 100 O F using the Gas Procesors Association
$ using :
Step 5: Use
converged
where: F is the m a s of the feed used in the Reid test, obtained by flashing F, and
are the molecular weights of the liquid phase, the vapour phase, component
2,
and
PLF.
and liquid feed (at 100 OF);p35, pcO,and plm are the densities of the Liquid blend at 35 OF,60
Details of how the blend densities at the different temperatures and the densities of the liquid
and vapour phases can be obtained have not been provideci. Also, it is not clear how d u e s
needed in Step 4 can be obtained Erom the equation of state in Step 3. The equation of state used is the Soav+Rediich-Kwong equation [Vazques e t al. 19921: &?'6T pz-(V
v-b v
+6)
b are constants.
where:
b are determined
where: rn is the number of components over al1 feedstocks, 4 and aj are the Soave-RedlichKwong constants for the pure component i and j,
pure component i, and Ko and Ki are the binary interaction parameters used to represent the vapour-liquid equilibria of binary mixtures..
This method requires that the molar composition of al1 the feedstocks be known which may not
be readily amilable. Also, correlations for estimating densities would be required. In addition,
experimental data may be needed to determine the parameters in the Soave-Redlich-Kwong [Vazques et al. 19921 equation of state.
The major drawback of the theoretical approaches i s that they are not easy to use on-line in
that they require cornplex calculations and iterative solutions. h o , they require data su& as feedstock molar compositions and component activity coefficients which may not 'De available.
2.2.2
INTERACTION METHOD
The interaction approach has &O been apped to R V P blending by Morris [1975]. The blending
mode1 is of the same f o m as in Equarions (2.6) and (2.7) evcept R V P is blended instead of
ON (octane numbers). Here also, only interactions between pairs of components is considered:
the three-way and higher interaction teare ignored. The Interaction mode1 is:
where: n is the number of feedstocks, iYkis the interaction coefficient between feedstock i and
racy of the Interaction method c m be found in Morris [1975] where the standard deviation of prediction error on an 8 cornponent. 42 blend study is reported to be 0.18 psi.
2 . 2 . 3
-4.neasy to use empiricai method developed by the Chevron Research Company is the Blending
Index method [Gary and Handwerk. 19941 . In this approach, blended RVP's are predicted
using the Reid Vapour Pressure Blending Indices (RVPBI) which blend linearly. The RVPBI's
are given in Gary and Handwerk [1994] and are of the form shown below [Gary and Handwerk,
19941:
RVPBI = (RVP)'.~'
(RVPBI)M,nd=
(2.12)
C (RVPBI),
1 Method
Stewart Ideal blending* Haskel and Beavon 1 1 9 4 2 1 r ~ o l a average r
*ldeel
1 Standard Deviation 1
(psi)
0.76 1.30 1 . 0 1
1.17
O T m R METHODS
Haskell and Beavon [1942] presented a method which involves the volumetnc averaging of the component RVPs except those of butanes which were treated differently; the butanes are assigned variable "blending pressure values" which were calculated based on the R V P of the butane and that of the debutanized blend (blend with al1 components except the butanes). Another simple approach that bas been used involves molar averaging (not volumetric) of the component
2.2.5
Cornparisons of predictive accuracy of some of the methods can be found in Stewart [1959b] looked at the standard deviation of prediction error for 67 blends wing different models.
The reported standard deviations are presented in Table 2.3. It is not surprising that the
method based on modeling the actual process provides more accurate predict ions. However, the thwretical rnethods by Stewart [1959bI and Vazques e t al. [1992] are not suitable for use
in on-line control systems due to their computational requirements. The interaction method is
also not suitable for on-line implementation since it requires that a lot of parameters be updated on-line as discussed in the octane blending case. Although not as accurate as the theoretical methos. simpler methods such as the Blending Index method can easily be incorporated in on-line controllers.
ASTM DISTILLATION
Gasolines, since they are mixtures of hydrocarbons with different boiling points, vapourize over temperature ranges. For example, the light virgin naphtha stream vapourizes throughout a temperature range of 30C
- 90 O C .
the ASTM D-86 distillation which is a standardized laboratory batch distillation that is carried out at atmospheric pressure wit hout fractionat ion. It gives amount (percent volume) dist illed to a given temperature. Several temperatures d o n g the distillation curve are used a s reference points in comparing the volatility properties of gasolines. Equivalently, distillation properties can be specified as temperatures at a given percent evaporated. Some of the methods for predicting blended distillation points available in Literature are briefiy discussed next.
2 . 3 . 1
INTERACTION METHOD
The interaction method, as described for octane blending, c m bo used t o predict the ASTN distillation points as percent evaporated at a given temperature [Morris, 19831, [Morris. 19751. Again, only the two-way interactions are considered and the interaction coefficients between al1 the component pairs would have to be determined for each temperature on the distillation curve. As in the octane and RVP blending cases, determination of the interaction coefficients require the quality (percent evaporated a t that temperature) of the individual components as well as those of the 50:50 blends and are calculated as described for the octane blending case. This method is only applicable to the intermediate points. not to end points of gasoline blends. The biggest drawback to using the Interaction mode1 is the large number of parameters that wouid have to be determined. For each temperature on the distillation curve,
n n 1) &f-parameters
are needed a-here n is the number of components. Thus for a 7 cornponent blend with 5 points on the distillation curve, 105 interaction coefficients would be required.
2 . 3 . 2
EMPIRICAL MODELS
Empirical models are quite often used in predicting the ASTM distillation points of blends and owe their popularity to their ease of use. Stanley and Pingrey [1954] have developed correlations for predicting the end points of distillation curves when only the blend composition and component ASTM dist ilIat ions are known. Correlation analysis on experiment al d a t a was
used (details of whidi were not provided) to produce tables of values that can be used for distillation end point estimation. This method does not provide estimates for intermediate points on the distillation cuve. Dhulesia [1984] has reported using models of the form:
l and a* are constants, is the percent volume evaporated, TI and T, are the final where: a
and initial boiling points, respectively, of the blend. Another empirical mode1 reported is of the form [DeWitt et al., 19891 :
where: a 3 and a4 are constants, k is the percent evaporated at temperature k for the blend, and
is the percent evaporated at temperature k for component i in the blend.
Indications of predictive accuracy of the models could not be found; thus sufficient information
is not amilable to ailow cornparisons to be drarvn arnong the diEerent models.
2.4
BLENDING SIMULATION
In order to conduct blending control studies by simulation, models that can closely represent the
blending process are required. Blending models are also needed for off-iine blend optimization,
as well as on-line blend optimization and control. For the purposes of the present research, only
octane number (RON and MON) and RVP blending will be simulated. Ot her qualities will not be considered. The blending simulation has been restricted to the three most important gasoline properties so that the important results can be highlighted wit hout adding unnecessary complications.
A good blending mode1 should provide acceptable predictive accuracy, it should be parsimonious
and easy to use, and it should not require a large data set to be trained on. Also, for on-line
optimization, a mode1 that can be readily updated on-line would be desirable. This is because models are not perfect and their predictive accuracy u s u d y degrades over tirne (as feestock properties and t hu blending characteristics change wit h tirne) unless they are appropriately updated periodicaily. Octane blending rnodels have been discussed and cornpared in Section 2.1. .4mong the models found in published literature, the Ethyl RT-70 [Healy et al., 19591 modek were found to exhibit the best combination of predictive accuracy and parsimony. Also, the Ethyl models c m readily be made adaptive for on-iine implementation. Healy et al. [1959]have w d data from 135 blends to determine the correlation coefficients. Thus the Et hyl correlations wit h the published coefficient values were chosen over the other octane blending models.
R V P blending models have b e n discussed in Section 2.2. Very iittle information is currently
amilable in Literature that can be used to choose the best R V P blending model. .4iso, blending data is not available that can allow the different models to be compared for predictive accuracy.
Thus' based on ease of use and a high degree of industrial acceptance, the Biending Index
method was chosen for modeling R V P blending. The selected models are presented again for completeness in the following equation set and &il1henceforth be used to model the gasoline blending process:
The parameters used for octane nurnber blending are [Healy et al., 19591: ai = 0.03224, a2 =
Chapter 3
PARAMETER OBSERVABILITY
IN RTO SYSTEMS
In Chapter 2, gasoline blending models were examinai for use in the gasoline blending RTO system s h o w in Fibwe 1-4. Since the selected blending models are to be ued on-line, the adjustable parameters in the models would have to be updated on-iine in order to maintain mode1 accuracy. Therefore. it is important that al1 adjustable models pararneters be observable for the available measurements. That is, the measurements should d o w unique estimates of the adjustable parameters to be obtainod t hrough some estimator. The success of any RTO system is dependent on al1 the adjustable parameters being observable from the process measurements
extended to degree of observability. The chapter concludes with examples to illustrate bow the new observability measure can be applied to diEerent estimation schemes.
3 . 1
BACKGROUND
Observability was originally defined and studied for linear dynarnic systems by Kalman [1960]
as a dual to controllability. Kalrnan observability is used to determine whether a given set
of measurements contain sdficient information from which the states of the system can be
determined (Brogan [1991], Friedland, [l986]).It will only be briefly discussed here since the
focus of the present work is RTO of steady-state operations policy. Consider the standard
state-space representation of a dynarnic linear process:
where: y is a vector of state \&ables: u is a vector of control moves, z is a vector of rneasurements or outputs, and c ( t ) and v ( t )are noise terms. The system given in Equation set (3.1) is said to be completely obsenable if every initial state
y(to)can be detennined through knowledge of the system control, u ( t ) ,and output. y ( t ) , over
some b i t e time i n t e r d to 5 t 5 ti [Ray,19813. For systerns where the matrices A and C do not vary with time, observability can be checked by constructing the observability matrix M:
where: q is the number of state variables. The linear dynamic system with constant matrices A and C is completeiy observable if and only if the rank of the observability matrix M is q, where q is the number of state variables [Ray, 1981]. Here complete observability means that every state variable is observable. If only
a subset of the state variables are obsenable, the system is said to be partially observable.
Conditions for other classes of dynamic systems can be found in Bryson and Ho [1976] and
Astrom [1970].
So te t hat checking for observability by test ing the rank of the observability mat rix can yield only
has looked a t the singular values of the observability matrix provided by the test. Gudi [1995]
to characterize degree of observability. Gudi (19951 has used the condition number, the ratio
A smdl condition number indicat es t hat the state variables are "strongly observable" while
a large condition number indicates "weak observability" . Since condition numbers are scale
dependent [Golub and Van Loan, 19891, such a measure of degree of observability will also be scale dependent.
The concept of obsenability was extendeci rom linear dynarnic systems to nonlinear steadystate systems by Stanley and Mah (19811 while retaining the fundamentai idea behind observability. That is, it is used to check if changes in all the state variables c m be determined through knowledge of the process outputs (measurements) and the steady-state models that describe the process. Stanley and Mah [1981] considered a steady-state system of dimension q with m measurements shown in Figure 3-1. They defined the system as the triplet (S,
C,
V)
where S is the feasible set of States that always satisfy equality, inequaiity and strict inequality constraints,
C is the vector
and V is the set from which measurement noise realizations are obtained. The measurements are given by the measurement functions as:
where: y is a vector of state variables, v is the noise vector, z is the measurement vector, and S is the closure of set S consisting of S and dl its limit points. In the case of perfect measurements, V = {O}, v = 0,and z = C(y).
C,
{y : w ( y ) = O, y E N} where vector w contains plant mode1 equations such as material and energy balances, thermodynamic relations, etc., N is the union of a finite number of convex sets, N,
c W, that
C and/or
have provided some definitions regarding observability that are useful Stanley and hlah [1981] for the present work. These are given below.
Cl V ) let
y , E S and I b e an index
Measurement Noise
1 .
Feasible Set
,
5
1
569
&l
Steady-State System
Estimator
4 Y^
Figure 3-1: Steady-State System and Estimator Considered by Stanley and Mah (19811. set. y ' is locally unobsentable at y , if there exists a sequence { y k } h ,such that:
That is?the states included in the index set I are defined to be locally unobsemble if there are changes within this set of states that do not result in a corresponding change in measurements.
For the situation described in the defution of local unobservability. even though y, differs frorn
y , , they both result in the same measurement value. As a result. one cannot be distinguished
, . Observability is then frorn the other and the system is said to be locally unobservable at y
defineci by Stanley and hl& [1981] in ternis of unobservability:
Definition 2 Locd Observability: in a system {S, C, V ) let y, E S'. Y ' zs locally observable
ut
y, if it is nat unobsentable.
Stanley and ~Mah[1981] have also provideci a method of testing local observability which is
Theorem 3 For a system gzven in the standard fonn S = { y : #(y) = O, y E N}?y , E S. let
w and
V,w(yo) vyC(ya)
T h e system
= q d e n q i-s the n u m h i
of states in the steady-state system. Vyw(y.) and VyC(yo)are the Jacobian m a h c e s ofw and
('?
Thus. instead of checking the rank of the observability matrix M. as in the case of linear dynamic systems. the rank of the augmenteci Jacobian rnatrix
the linear dynamic case. if the rank is found to be l e s than q: the state variables are not completely observable. The rationale behind the rank requirement for local obsenability can
be illustratecl using first-order Taylor series approximation [Cornin and Szczarba. 1995i to the
1 VYC(YO)
1 Vyw(YO) 1
J
should be checked. As in
In order to ensure that changes in states (Ay) are uniquely obsenable from the changes in
r
1
The concepts of state observability for steady-stat e s-wtems was extended to parameter observability by Krishnan [1990] a-ho considered a generai nonlinear system where the stead-state model of the process and the measurement equations are defined as:
where:
B E W'
As an extension of the rank condition in Theorem 3 for local state obwrvability, for al1 the . , the observability criterion presented pararneters and states of the systems to be obsenable at y
by Krishnan [1990] requires that:
rank
V Y ~ Y O )
where: q is the number of state variables and p is the number of parameters to be determined.
The need for the above rank requirement can be iilustrated usinp the Implicit Function Theorem
[Corwin and Szczarba, 19951:
d u =
dC =
Thus for this Linear system of equations to have a unique solution with respect to dy and dfl
at y , , the rank of the matriu:
quantities (pararneters and states) can be determineci from the measurements. Even when these quantities are observable through the mode1 and measurement equations, a poorly designed estimation procedure may not allow the measurements to be uniquely mapped ont0 est imates, rendering the unmeasurable quantit ies uno bservable. Therefore, the characteristics
of the estimator should be considered-
3.2
PARAMETER OBSERVABILITY
The obsemability criteria presented by Stanley and Mah [1981j and K r i s h n a [1990] do not consider the effect of the estimation procedure. Their methods for determining obsenability have
b e n based only on the models that describe the steady-state mode1 equations and process measurements. If the system is unobservable according to the rank cnterion in Equation (3.2). then the variables will cert ainly be unobservable through the whole updating procedure. However. cornpliance with Equation (3.2) does not guarantee observability through the entire updating process. That is, Equation (3.2) provides a necessary but not sufficient condition for local obsenability. For a sufficient condition, properties of the estimator would have to be included
in formulat ing a measure of observability. In t his section pararneter obsembili t y for nonlinear
stead-state systems wili be considered and a new measure of obsenability will be developed based on statistical principles. The concepts udi then be extended to provide measures of degree of observability. First, a definition of pararneter observability is provided:
Definition 4 Parameters are dejned to be obsentable fmm a given set of pmcess rneasurernents
if a unique set of pammeter estimates c m be determined wing a speci,5ed updating scheme.
In order to gain insight into how pararneters may become unobservable, consider a steady-state
system for which the adjustable parameters
(fi).
measurement error
may be normally dist ributed. the same is not necessarily true for the resulting est imates. Figure
3-2 shows the propagation of measurement noise through what may be a nonlinear updating
Updator
Fi-we 3-2: Parameter Updating
scheme, a-hich will distort the distribution characteristics The developments of this chapter
assume that the parameter estimates may be locally approximated in some small neighbourhood
of the nominal estimates using multikw-iate normal distribution (ie.. probability density hinction of
8 --~p(&
Qd)j. The
For a bivariate case. this funcrion represents a beU-shaped surface whose probzbiiity contours
on the pararneter plane are concentric ellipses around
parameter space are pdimensionai efipsoids. Thus. the function in Equation (3.4) exhibits a
single extremum which is a maxima located at
P.
The surface would not possess a unique maxima if the surface were to become Bat in which case
the updator would not be able to use the rneasurements to obtain a unique set of parameter
estirnates. The conditions under which such a situation couid arise can be examined by looking
at the distribution of
fi more closelc
the exponential t e m in Equation !3.4) and the determinant of Qo appears in the denorninator
s [Chatfield and Collins. of the pre-exponential terrn. One definition of the determinant of Qg i
The eigenvalues of 4s'. q, are reciprocals of the eigenvalucs of Qg (2. e., qi = i = 1, ....p ) B [Chatfield and Collins. 1980j. Soa-. as an eigendue q, decreases. lQbl increases and the p r o b
ability density function decreases in height. At the same time, the hvper-ellipsoid elongates in
k,
limit as rlj
Tj,
0, the surface becomes Bat and a h i b i t s no extremum; a t this point a unique set
of estimates does not exkt and the parameters are unobservable. Then observability can be
dehed in term of the properties of QB. Conditions for observability can
&O
m a t h Qg, the probability density hinction can be written as [Chatfield and Collins. 19801:
Taking logarithms of both sides and multiplying by -1 gives the following quadratic equation:
The random variable c2 follows a chi-squared distribution with p degrees of fieedom [Chatfield
and Collins. 19801. For a prescribed d u e of 2, this equation describes the surface on which the
probability density h c t i o n has a constant value. Since Qg is required to be positive definite for validity of Equation (3.4):the surface in Equation (3.7) is a pdirnensional ellipsoid that describes the joint confidence region of the parameter estimates [Sorenson. 1980]. The lengths of the semi-axes of t hese hyper-eUipsoids indicate the uncertainty in the comesponding eigenvector directions in the space of the parameters [Sorenson. 19801. [Krishnan. 19901. The semi-axes 2 and have directions defined by the eigenvectors T, of the hyper-eilipsoid have magnitudes ?; [Sorenson, 19801. For good estimates then. large eigenvalues of Q-' are desireci. D Alternatively? conditioning of the joint confidence region can be studied by exarnining Q-'
very ill-conditioned. As 11, approaches zero. the confidence contours in the parameter space stretch out in the direction of r, and in the Iimit become unbounded as depicted for a bivariate
case in Figures 3-3(a) to 3-3(d).
In Figure 3-3(a). the contours are concentric circles which inicate a perfectly conditioned Q;'
(ail eigenvalues are q u a i ) . As one of the eigenvalues decreases. the contours elongate as shown
Figure 3-3: Confidence Contours in Figures 3-3(b) and (c). Figure 3-3(d) shows the limit case when one of the eigenvahes of Q: l reaches zero and the contours become unbounded. At th% point, a unique
the parameters become unobservable. Hence an alternate definition of parameter observability can be provided in terms of the eigenvalues of the inverse of the conriance matrix:
normally distributed) are to be determined using measurements (z), the parameters are obseniable zf al1 the eigenvalues of Q:' are strictly positive (positive and nonrem) where Q p is the P covariance mat& of 6. In order to demonstrate how parameter obsenability relates to the covariance rnatrix of the estimates, a simple linear example is used below. The exarnple will illustrate how parameters
(2)
v is a vector of errors
x are variables
The estimates
Case A:
22.944
The eigenvalues of 4 8 'are 0.48,53.02, and 2.49. Since di the eigenvalues of Q:' nonzero, the parameters are observable.
.&O,
are
(B) are quite close to the true value of the parameters (P).
Case B: If rank[x]
< 3, the m a t h
rneters would not be observable. Consider the case where the third row is dependent on the first two:
Here Q-' =
P
and its eigenvalues are 2.489, 0, and 73-511 . The zero eigenvalue
of Q F ~ indicates that the parameters are unobsernble and the least squares estimate cannot be obtained.
Case C: In this case
329.875
-445-000
165.313
-222.500
-222.500 300.000
quite f a r rom the true values of the parameters) because the parameters are prac-
tically unobservable. This can be seen by examining the eigenvalues of QY' which
are 8.9*10-~. 2.43, and 72.3. Although al1 the eigenvalues are nonzero, one of the
eigenvalues is very small, causing the parameters to become practically u n o b ~ n ; able. Also, the joint confidence region is very ill-conditioned - the condit ion number,
Q
direction corresponding to
vmin
the direction corresponding to qmi, is almost 105 times larger than in the direction corresponding to q,, . In this rather simple linear example it is easy to recognize the estimation problem as being pooriy posed by exarnining the condition number of X . Nonlinear estimation problems can be
checked in a similar manner by linearizing the nodinea. problem about some nominal operating point and then checking the condition number of the resulting matrix. Such an approach ni11 only locally test the estimation probIern and wili not be able to detect celineuity of the nonlinear mode1 equations at other operating points. That is, c~linearityof equations at
certain points, which can cause the parameters to become practicdy unobservable, may not
be easily detectable. In such situations, test ing for local observability by exarnining the inverse
of the covariance matrix of pararneter estimates as proposed here. will identiS the obtained estimates as being poor. Note that the least squares estimates obtained in Case C are very poor
(B is not
close to ,6) even though ail the eigenvalues of QT' are nonzero and so the
B
updating problem passes the observability test. This highlights the need to consider degree of observability rather than a binary observability test.
3.2.1
DEGREE OF OBSERVABILZTY
The observability criterion for pararneter observability avaiiable in literature by Krishnan [1990] given in Equation (3.2) determines whet her or not quantit ies (states/parameters) are observable. Thus there are only two possible outcornes from the observability test: observable or unobservable. This test provides no information about the quality of estimates. For dynarnic systems. Gudi (19951 has used the condition nurnber of the obsenmbility m a t r k M to differentiate between "strong" versus "weak observability. The approach taken by Gudi [1995] could
be applied to pararneter observability of steady-state systems by examining the condition num-
ber of the augmented Jacobian matrix in Equation (3.2) to get an indication of the degree contains no of obsembility. However. as has been pointed out earlier, the augmented m a t r ~ u information about the estimator. In this section, it will be shoan how the proposed parameter observability measure naturally leads to a measure of degree of observability that is based on the statistical properties of the pararneter estirnates.
Degree of observability, in this context, refers to the quality of estimates that can be obtained
from a given set of measurements via the updator. In designing experiments for precise parameter estimation, the joint confidence region of the estimates has been used to obtain measures of estimate quality [Sutton and MacGregor, 19771, [Box and Lucas, 19591' [Box and Hunter,
19651. Quality of parameter estimates is related to the characteristics, both size (volume) and
shape, of the joint confidence region of parameter estimates [Sutton and MacGregor, lW7]. In general, updators yielding confidence regions with s m d volumes provide better estimates than those resulting in large volumes (of similar shape). Confidence regions that take the shape of
a hyper-sphere are more desirable than very elongated regions since the more elongated the
and Box hyper-eliipsoid gets, the more correlated the estimates become. Box and Lucas [1959]
and Hunter 119651 have used the determinant of Qg as measure of the volume of the joint
confidence region. The rationale behind using the determiant is that
IQDJ
is the product of
the eigenvalues of Qp which represent magnitudes of the semi-axes of the hyper-ellipsoid (joint confidence region). Therefore. the larger the determinant , the greater the volume. However, the volume criterion gives no indication of the shape of the confidence region. An indication of the shape or conditioning of the region can be obtained by looking at the condition number of the Hessian of the function c2, QT' P
:
2 takes , on
is a hyper-sphere and increases as the region becomes elongated. Another measure of shape
Qil
and so incorporates information about uncertainty in al1 the p independent directions in the
parameter space whereas directions. However, adopted here. Thus both the size and shape of the joint confidence region have to be considered in determining the degree of parameter observability. For "strong" observability, small lQBl and nz are
r;2
~2
considers only the best (rnost certain) and worst (least certain)
required. Among regions wit h similar volumes, the one yielding the smallest value for
r ; wodd ~
be preferred, while among regions with similar conditioning the one encompassing the smallest volume (for a given confidence level) shodd be chosen.
4 simple example is used below to illustrate the use of the shape criterion in selecting among
estimators that yield confidence regions of s i d a volumes.
EXAMPLE 3.2
5 sing two different estimators A and B. The inverse of the covariance rnatrk of the
=
j
a=[
p,
] [
1 ( Q ~ )1 = 0.20, the joint confidence regions frorn both the estimators enclose the same volume for a given confidence level. However. the 4 ; ') and (Q; '), are vastly different condition numbers of (
and 0.43. Since
A
The eigenvalues
P1)B
are 11.57
The confidence contours resulting horn the two estimators are shown in Figures 3-4
(4 and (b).
The confidence regions resulting hom estimator B are much more elongated than t hose from estimator A. Therefore est imator A provides much stronger parameter observability and so shouid be chosen over B. In this section a criterion for determining whether or not parameters are observable from a given updating scheme and process measurernents has been provideci. The criterion is based on
evarnining the inverse of the covariance matrix of estimates (QT' ). The idea was then extended
P
to obtaining an indication of the degree of obsenmbility which again looks at the characteristics
of
3.3
COVARTANCE MAT-
APPROXIMATION
which produces parameter estimates rom process measurements. When O is at least once
continuously differentiable in z, changes in the adjustable parameters
(AB) in
response t o
srnall changes in the rneasurements (Az) can be approximated using a first-order Taylor series
as:
where:
Thus small changes in z, including measurement noise, are approxirnately mapped to perturbations in
B through
to produce a variance in the parameter estimates, Qg, can also be locally approximated using
Thus measurement noise propagates through the updator, the properties of which are captured
in
will be examined
It has already been estabiished that any measure of observabiiity should include information
about a l 1 components of the parameter updating systern. Qd depends on
Qz)
2 (in addition to
which in turn depends on 0 , the updator. Thus Qg contains information about the whole relates to the cornponents of the updator will also be discussed
3 . 3 . 1
ESTIMATION EXAMPLES
for which the pararneters 0 have to be estimated using the measurements z. There are many alternative estimation techniques which may be used in updating the adjust able pararneters. This section will focus on t hree common met hods (back cdculat ion, optimizat ion- based estimation, and prediction filter based estimation) and illustrate how
c m be obtained for each
case. The covariance rnatrix of estimates, Q8, can then be approxirnated using
g.
Note that since t here are no output equations or state variables involved for steady-state systems described by w ( z ,P ) = O, Krishnan's [1990] obsenability criterion in Equation (3.2) will reduce
rank [V& = p
BACK CALCULATION
The parameters can be back-calculateci using a v e n values for measurements (z) by solving the
following sirnultaneous set of (nonlinear) equations for 8:
as required by the rank criterion provided by IOishnan [1990] aven in Equation (3.10).
The sensitivity matm can be determined by taking the total differential of Equation (3.11):
h m which
parameters. QB can then be approxirnated using Equation (3.9). S o t e that calcdation of requires that rank criterion by Krishnan j19901 in Equation (3.10) be satisfied. Since in back-calculation al1 the characteristics of the estimator are included in the mode1 equations ( w ) thernselves. the approach provided by Krishnan [1990] includes information about
dl cornponents of the estimator. Also. from Equation (3.12) it can be seen that
(and therefore
Q B )contains a l 1 the characteristics of the estimator as well. Thus for back-calculation. both approaches d l provide necessary and sufficient conditions for parameter obsemability.
OPTIMIZATION-BASEDESTIMATION
Parameters are often updated using optimization based methods such as nonlinear regression.
P.
subject to :
where:
B. and
E.
problem transforrns it into its reduced space [Fletcher, 19871 where it becornes an unconstrained optirnizat ion problem:
mjn
P
o(D:z)
(3.14)
Sufficiency conditions for a minimum require t hat the reduced gradient (Oro) ~ a n i s h at the optimum and that the reduced Hessian be positive definite for a unique minimum:
~f O be positive definite
Forbes j19941 has provided an expression for the reduced gradient (nhere the appropriate derivat ives e i s t ) :
The sensithlty matrix can be obtained by taking the total differential of the reduced gradient and rearranging:
Thus
depends on the inverse of the reduced Hessian and the effect of perturbations in
parameter updating procedure including the characteristies of the objective function as n-el1 ai
at.-O the steady-stare modeis that descnbe the procev (contained in the vector f ) . Similarl- 7
n-iU &O
contain information about the objective function as weli as the mode1 equations.
local sensitivity analqsis [Seferlis. 1995: which is based on the Implicit Function Theorem :Coruin and Sznarba. 19951 and describes the fim-order changes to the optimum subject t o small perturbations about a nominal point. The approach to sensitivity analysis taken by Ganesh
cf3 and Biegler il9871 rields a linear system of equations that allous for easier computation of z.
The aut hors develop sensiti\-ity analgis under the assumption that the fol1oning condit ions are
satisfied at the local minimum.
e':
B and
(B'. 2').
b'
and strict complementary
slachess [Edgar and Himmelblau. 1988j holds for the Functions in Problem 3.13 at
a'
3. The second-order sufficiency condirions are met [Ganesh and Biqler. l98j.
From the Karush-Kuhn-Tucker (KKT) conditions !Edgar and Hirnmelblau. 1988::
a-here: L = 0
mult ipiiers.
of Lagrange
Applying the Implicit Function Theorem [Cornin and Szaarba. 1995i and rearran,$ng Dives:
v & '
!or
can be obtained for maIl optimization problems. numerical methods =-ould have to be utilized
for larger ones. The cm.ariance rnatnx of parameter esrimares !Q8 j c m then be apprmimated using Equation 13.9).
ponents of the updating procedure. The Lwangian L contains the objective function O as nell
as the model equations (contained in f ) . Thus the characreristics of
O
T,.,LS. V : L ' . Tg,L ' . and T , L'. Information on the model equations are contained in these
teas well as in T,f.Tf. and Vcfs.
s u b j e d to :
L = E'AE
R o m Equation (3.18) it can be seen that
t pT[o(,&z)- E]
estimator - the model equations in the constraints as u d as the objective function. Characteristics of the objective function are containeci in the ternis VB,LalVr,cL', v~L'.VB,L? and
V,,L*? while mode1 information i s included in these terms as well as in V,w'. 'VjwT*. V,w'.
and V.ok. The method presented by I(rishnan jl99O! examines o d y the r e m V D wand so looks only at the constraints in the estimator in Probiem (3.17) and neglects the objective function ent irely. providing a necessary but not sdicient condition for parameter obsemability. Test ing for obsemibility using Qg on the other hand. provides a sufficient condition since i t includes al1 aspects of the estimator.
applied very successfuily over the past three decades for state estimation. Prediction filters
based on Box and Jenkins :19T6i methods are also widely used. Other prediction filter design
rnethods include those baseci on Discounted Least Squares approach [llontgomeq-and Johnson.
1976; and the Bayesian approach :Ilontgomey and Johnson. 1976:.
a-hich involves the current measurernents ( z t , ) and their predictions from the previous time step
(itkIt4 [Ray. - , ) 19813. Thus the expression for updating the adjustable pararneters is given by:
a-here: PtkitC is the vector of parameter estimates at time t , using data up to time tt' ,Btki,-,
tk
at time t t , itkl,,-, are predicted measurements at time tk using data u p to time t t - l ' and K is the ilter matrix. Step ahead predictions of parameters
I)
using appropriate prediction models. When such information is not amilable, one approach is to assume that the measurements and the parameter estimates wiii remain unchangeci over the
next t ime-step:
m i t ten
as:
Comparing Equation (3.19) with Equation (3.8): the sensitivity of the curent estirnates to the measurements.
d4tk'tk. is
&tk
Thus the sensitivity matrix ail1 depend on filter design. The co~ariance matrix of parameter estimates Qq can be obtained using via Equation (3.9) as:
-1
The met hod for determining parameter obsembility presented by Krishnan [1990] ws steadystate models ( w ) that describe the process as given in Equation (3.10). These models are commonly used in the design of the Kalman farnily of filters [Chiu, 19871 (in addition to other information such as noise characteristics). However. certain other classes of filters such as
prediction filters b a s 4 on tirne-series andysis [Box and Jenkins, 19761 may not directly contain these steady-state models. Moreover, the filter matrix may be user-tuned in which case the filter matrix (K) will not contain any mode1 information. Thus for filter based-estimation. determination of parameter observability using the approach of Knshnan [1990] uses only partial 6iter information (as in the case of the Kaiman f d y of filters) or no information at dl. Therefore, Knshnan's approach has only lirnited application to some fiiter designs and no application to ot her designs. The st atist ical approach presented in t his chapter, however. can be applied to all filten and will provide a sufficient condition for parameter observability since it uses the filter matrix directly in calcdating Q p as shown in Equation (3.20).
3.4
DISCUSSION
Estimation of unmeasured quantities fiom secondary measurements for steady-state systerns is an important problem in the chernical industry. For example, concentrations in distillation
colwrins are
to ensure that unique estimates of these quantities can be obtained from the available mea-
surements and estimator. Moreover, the estimates must have an acceptable level of statistical certainty. The approach to determining parameter obsenability for steady-state systems provided by Krishnan [1990] considers only the steady-state models that represent the plant. It does not take into account the characteristics of the estimator. It has been shown in this chapter that
t his approach cannot provide a sufficient condition for pararneter obsenabili ty. A statistical
approach to determining observability of parameters from the measurements and estimator has been presented which takes into account d l components of parameter estimation to provide a sufficient condition. The ideas behind parameter observability have been extended to degree of observability which look at quaiity of estimates.
The proposed approach to observability and degree of observability involves examining the
eigenvalues of the inverse of the covariance matrix of pararneter estimates (QT'). Since eigenB
values of a matrix are dependent on the scaling applied to the elements of the matriw, degree of
observability will also be scale dependent. Gudi (19951detemines degree of state observability
in dynamic systems by looking at the singular values of the obsembility matriw (M) which are also s a l e dependent. Therefore, the matrices
(by scaling a i l elements to the same order of magnitude) before degree of observability can be determined. In the steady-state parameter obsemability case, Q-' should be based on al1 parameters
In RTO systems, parameter observability can be used in selecting adjustable pararneters. Krishnan [1990] has provided a met hod for selecting adjustable paramet ers which invoives identifjing "key" pararneters and then updating these parameters on-line. Key parameters are determined by perturbing each parameter at a tirne and noting the resulting change in the objective funct ion value and any active set changes. Parameters that cause the most change in objective function value per unit change in parameter. or thase that cause the active con.stra.int set to change. are selected as adjustable parameters. However, the objective function in RTO system is based on process econornic which is usually o d y poorly known. Therefore, the predicted profit could be quite different than the actual profit. Thus selecting adjustable pararneters based on objective function changes could lead to poor RTO design. Forbes and Marlin [1996] have provided an alternative approach to adjustable parameter selection which is based on minimizing loss of
RTO performance. Loss of RTO performance is determined as the Design Cost which considers
performance loss due to: offset from true plant optimum, and variance in predicted optimal setpoints. In the Design Cost approach. those pararneters which when updated on-line result
in the l e s t Design Cost. are selected as adjustable pararneters. However, this s t r a t e 3 (as
weIl as that presented by Krishnan [1990j) does not consider parameter observability. Thus. it could in some cases, lead to the selection of parameters that cannot be estimated from the measurements and estimator and so cannot be updated on-line. The first step in any parameter selection scheme is to ensure that the selected parameters d l be observable frorn the a~xilable measurements and chosen estimation scheme. Parameter observability can be tested in two steps. First, the method presented by Krishnan [19901c m be used which provides a necessary condition for observability. Parameters that fail to satisfy the rank criterion in Equation (3.2) can be discarded from the set of candidate adjustable pararneters at this point. Next. the observability approach presented in this chapter should be used or1 the remaining pararneters to test if the parameters would be observable from the
measurements and the estimation scheme. Parameters that are not observable or those that are
o d y weakly obsenxble can also be eliminated. Note that the proposed statistical approach to
determining parameter obsembility will recognize and eliminate paramet ers that fail to satisfy the rank criterion in Equation (3.2) since it provides a necessary and sufficient condition for pararneter observability. However, the second test is specific to a given estimation scheme. Thus
if obsenabiiity t hrough ot her estimation schemes have to be tested, the parameters eliminated
by the h s t step wodd not have to be tested again. The above procedure will narrow the set of
possible adjustable parameters to those t hat exhibit an acceptable degree of observability. The Design Cost approach to adjustable parameter selection presented by Forbes and Marlin [1996!
can then be applied to the resulting srnder set.
Thus the observability measure developed in this Chapter can be applied to RTO systems
as well a s to any steady-state system where parameters are to be estimated frorn secondary measurements.
Chapter 4
BLENDING CONTROL
TECHNOLOGY
This chapter builds on the modeling and parameter observability work presented in Chapters 2 and 3 to provide an efficient blend controller. The chapter begins by presenting the perfect or
ideal blend controller which is then used as a benchmark for comparing controller performance. In addition to providing a performance bendunark. the ideal/perfect controller provides insight into better controller design. Xext , the nidely used control strategy and the problems associated
with it are examined. Finally, a new control approach that overcomes the shortcornings of the
current strategy is presented. Performance of diEerent blend controllers are demonstrated using
4.1
INTRODUCTION
linear programming with bias updating. The success of this control strategy has been largely due to the practice of blending from well-rnixed storage tanks. However, increased cornpetition has forced refiners to reduce on-site tankage inventory by blending out of "running" tanks (or direct in-line blending). "Running' tanks are those intermediate tanks into which feedstocks from upstream processes are being pumped. Also, process strearns can be routed directly
to the blender without first being pumped into either an intermediate storage or "running" tank [Agrawal, 19953. Since blending feedstock qualities are dependent on upstream process operations and upstream process changes ( e.g. catalyst deactivation, heat exchanger fouling,
crude oil switches, etc.), they may not be constant and may vary significantly during blending.
This chapter focuses on the problem of controlling a gasoline blender which is subject to stochastic disturbances in the feedstock qualities. Such disturbances nat uraily arise when the blender is using "running" tanks as feeds and upstream process operations change. The objectives of this chapter are to demonstrate the problems that existing blender control systems may encounter when faced with feedstock quality disturbances, to provide a control strategy that can effectively deal with such disturbances, and to illustrate the performance improvements associated wit h the proposed cont rol approach.
4.2
An automotive gasoline brending problem is used to compare the performance of biend controllers in this chapter. Different blend controllers will be applied to the benchmark problem.
4 . 2 . 1
BLENDLNG PROBLEM
The blending problem studied here is based on a case-study by Forbes and Marlin [1994] and
involves simultaneously blending two grades of automotive gasoline using five feedstocks as shoan in Figure 4-1. For the purposes of this illustrative problem, the product specifications for each grade of gasoline will be lirnited to: minimum limits on RON and MON? and a
maximum limit on the RVP. Constraints are also placed on maximum and minimum product
demands and feedstock availabilities. Product specifications and economic data are provided
in Table 4.1 and feedstock data are in Tables 4.2 and 4.3.
In this problem, gasoline is to be blended for 24 hours with an RTO interval of 2 hours. As discussed in Chapter 2, the gasoline blending process is simulated using the Ethyl RT-70 [Healy
et al.! 19591 and the Blending Index [Gary and Handwerk, 19941 models. These models are
Regular Premium
value ($1b bl) ma.. demand limit (bbl/day min. demand fimit (bbl/dav)
33.00 8000 7000 37.00 10000 10000
l
reformate
( Available 1
LSR naphtha
n- but ane catalytic gas. alkylate (bbllday) 12000 6500 3000 4500 7000
Cost
Feedstock
RON MON Olef in (%) Armatics (%) R V P (psi)
reformate
(C14a)
LSR napht ha
(clh)
n-butane
(C25a)
catalytic gas.
(C12a)
alkylate
('34
94.1 80.5
1 .O
58.0 3.8
pnrantlicscs
i i ~ i r l e rfccdsrock
++
LSR naphtha
m!
*
n-butane
Regular
*
C
.c
catalytic gas.
Premium
STOCHASTIC DISTURBANCES
For the illustrative purposes of this case-stud- disturbances are to be added to RON. MON, and R V P of the feedstocks. In order to sirnplib the benchmark blending problem, it was assumed that disturbances were present oniy in the feedstock that affects the optimal operations the most. In order to determine the feedstock whose qualities the optimal blend recipe is most sensitive to? a local sensitivity analysis [Seferlis. 19951 was conducted around the nominal blend.
The nominal optimal blend was obtained using NLP with MINOS in GAMS [Brooke e t al., 19921
using data given in Tables 4.1, 4.2, and 4.3, and the blending models in Equation set (2.16). T h e results are swnmarized in Appendk A. After scaling a l 1 the decision variables and parameters to the same order of magnitude, the linear system of equations given in Ganesh and Biegler
[1987] was solved analytically using MAPLE [Char et ai., 19911 for the pararnetric sensitivity
of Vgx*, each corresponding to a feedstock, were obtained and their norms (largest singular
values) computed. The results are summarized in Table 4.4.
Subrnatrix il'orm
Table 4.4: Feedstock Sensitivity The submat rut corresponding to the reformate qualities yielded the Iargest singular value. Thus. stochastic disturbances were added to the reformate stream' for the purposes of this example. a s :
where:
%f.t
[ RON
if~.Y RVP
The scalar disturbance, d t , was modeled as steps of random height entering a well rnixed tank.
The tank was assurned to hold one third the availability of reformate (as given in Table 4.3) per eight-hour shift, resulting in the following transfer function:
where:
2 '
.S noise vector was generated for each 2 h o u time interval over the blend horizon and noise
was assurned to remain unchanged during each
the
reformate qualities over the whole blend horizon. The resulting RON, 6 1 0 N . and R V P of the reformate Stream over the 24 hour period is shoam in Figure (4-2).
4.2.3
CONTROLLER PERFORMANCE
Controller performance is measured by the profit earned from applying the controller on the benchmark problem posed in this section. When a controller produces blends that meet al1
Reformate Qualities
RTO Interval
Figure 4-2: Reformate qualites dunng 24 hour blend. quality specifications. the profit level can easily be computed using the economic d a t a provided
in Table 4.2. However, if the products at the end of the 24 hour blend horizon are found to be
off-specification. the- may not be suitable for sale as produced, and would likely be routed to .slop' tanks and reblended. In such situations! profit calcdation is not as simple.
One option is to include the cost of reblending in determining profit level. Alternatively. the
off-specification portions of blends (products made during RTO intervals that are causing the whole 24 h o u blend to be off-specification) can be "removed" from the blend. T h e profit can then be calcuiated based on the rnodified srnaller blend that does meet quality specifications. ..Uthough blend portions cannot be 'kemoved" or unblended from a blended product in practice,
it can easily be done for simulation studies for the purposes of accessing controller performance.
Since the economic data required for determining loss of profit due to reblending is not available for this study, the second alternative of "removing" off-specification products is adopted.
Small violations of quality specifications can often be tolerated since product quality measurements are limited by analyzer sensitivity. For the purposes of this study, it is assumed that
violations of up to -0.03 octane number for RON and MON, and +O.005 psi for RVP cm be tolerated. That is, blended quaiities that off-specification.
4.3
An ideal blend controller is one that gives the maximum possible profit rom the blending process. while meeting all blended quality specifications, and sat isfying demand and araila bility limits. In order to achieve perfect control, the controller must have perfect knowledge of
the blending process. That is, it should employ blending rnodels that capture the t m e blending
behaviour. In addit ion, the controller must have perfect knowledge of al1 feedstock qualities at d l times throughout the entire blending tirne horizon. For batch blending of gasoiine (blending to a storage tank for a given length of time as in the benchmark problem) , the ideal blending control problem c m be formulated as an optimization problem where the objective is to ma.irnize profit over the entire blending time horizon, subject to meeting quaiity constrainrs and demand/availability limits at the end of the blend. This problem has the form:
subject to:
where: c is a constant vector containing process economics, x are the feedstock or component
flows. R is a matrix containing known feedstock qualities. g is a vector of blending equations
that represent the true blending behaviour. w is a weighting vector (usually containing only ones and zeros to indicate the presence of a specific Stream in a blend). s are t h e blended quality specifications, h contains product demand and feedstock availability constraints. and t,, and tf
are the initial and final blend times, respectively.
The controller works to maximize profit o\*erthe whole blend by integrating blending economics from the start of the blend (t,) to the end of the blend (tf), rather than at any point in time.
ensure that quality specifications are met at the end of the blend. These constraints use perfect knowledge of the feedstock qualities (R) over the entire blend horizon (hom t , to tf) as well as perfect blending models in vector g. The second set of constraints:
integrate mass balance information over the blend horizon in order t o satisfy d l product demand
4.3.1
The perfect controller u formulated in Problem (4.3) was applied t o the benchmark blending problem. Vector g contained blending models for R O N , MON, and R V P for both gasoline grades. The blending models and contraints used here were the same as those used for simulating the blending process given in Equation (2.16). The dernand/aiailability constraints in vector h use m a s balances and assume no density changes upon blending. Feedstock flows xi
where:
X ~ J .
catalytic gasoline. and alblate. respectively: 16. .... xlo are B o a s at time t to premium gasoline
of reformate. LSR naphtha. n-butane. catai'ic
The perfect controller was applied to the blending problem by approximating t h e integais in Problem (1-31 n-ith m r n a t i o n s . and wlving using IIISOS/G;\.\IS YBrooke et a l . 1992:. The optimal blend trajecton and the active constraints during each K I 0 i n t e n d are provideci in
Xppendix B. The theoretical maximum profit &-asfound to be S66.276.55 for the 2 1 hour blend.
4.4
CONVENTIONAL CONTROLLER
as:
max
X
cTx
subject to:
where: c i s a constant t-ector containing process economics. R is a constant matrLv containing the feedstock qudity blending indices. s are the blended quality specifications. x are the f&to& or component f l m . w is a weighting vector (usually containing o d y ones and zeros
to indicate the presence of a specific Stream in a blend), b are the biases wd to update the
blending model to maintain its accuracy. and h i s a vector of linear equations for the maximum
and minimum product demands and feedstock availabilities.
The matrix R contains blending indices such as Blending Octane Sumbers (BO.\'-s) that
b l e d linearly as volumeric averages. However. as discussed in Chapter 2. most important gasoline qudities (such as octane number. Reid Vapour Pressure. and the -%S'Pd Distillation points) blend nonlinearly. Since linear blending models are unable to capture the true blending behariour. structural mismatch is introduced into the optimizer. Bias updating is used to compensate for such piant/model mismatch as weU as other modeling inaccuracies.
2. cdculating biases as the difference between measured blended qualities and those predicted by the linear model &cc-i.
where
xb-1
3. solrlng Problem (4.1) using the newly calculateci biases to obtain new blender feedstock
f l o( ~x ~ )
The process iterates over the whole blend horizon. That is. the above four steps are carried out
at each RTO i n t e n d until end of blend is reached.
In the bias update formulation of Problern (4.4). the only adjustable pararneters are the bias
terms whle ail the feedstock qualities are treated as fixed pararneters. Forbes and 4Iarlin (19941
have shown that a bias update approach can lead to a substantial l o s in blender profitability
given even a s m d variation in the feedstock qualities. Thus the current blender control structure
cannot adequat ely address the blending problem where the cornponent qualities are varying.
The conventional LP with bias updating controller sufTers from two limitations: inability to
incorporate the inherent blending nonlinearities. and ineffectiveness in dealing with fluctuating feedstock qualities. The effect of t hese limitations on cont roller performance is demonstrated
by applying the controller to the gasoline blending problem descnbed in Section 4.2.
4.4.1
The traditional LP with bias updating blending controller given in Problem (4.4) was used on
the blending problem posed in Section 4.2. Blending Octane Nurnbers (BON'S) [Gary and
Handwerk, 19941 were used for RON and M U N and blending indices (RVPBl's) were used
for RV P blending [Gary and Handwerk. 19941. Hence, matrix R contained the BON'S and
RVPBl's for al1 the feedstocks. The linear models used in the RTO layer result in structural
mismatch in octane blending.
The demand/availability constraints use mass balances and assume no density changes upon
blending. Feedstock flows at time t. xt, and the weighting vector. w, are given as:
where: x t , ~..., , xt,s are flows at time t to regular gasoline of reformate, LSR napht ha. n-butane, catalytic gasoline, and alkylate, respectively; xs,..., xlo are flows at time t to premium gasoline of reformate, LSR napht ha, n- butane, catalyt ic gasoline, and alkylate, respectively. The demand/availability limits at time t are formulated as:
The following steps were carried out at the kth RTO interval:
xk-1
Blended Octane
+ Bias Controller
The above steps were repeated at each RTO interval, and Problem (1.4) was solved using linear
prograrnming in GAMS [Brooke et ai., 19921. The optimization results (feedstock flowrates and
the end of the 24 hour blending period are s h o w in Figures 4 3 and 4 4 as deviations from specification.
Whiie there is significant quaiity giveaway in RON and MON of reguiar and RON of preMum gasoline, RVPs of both grades of gasoline are above the maximum specified and the
ib.iON of prernium grade is below the minimum dowed. In addition, RVPof both grades is
off-specification throughout the biend. As a result, both grades of gasoline produced are offspecification and may not be suitable for sale. Since both grades are off-specificationt hroughout the blend horizon, portions of blends cannot be removed to obtain blends that do meet specifications. Hence, a profit level for the conventional controller could not be cornputed.
In this case-study, the conventionai controuer fails to make products meeting quaiity specifications. Poor controller performance can be attributed to a combination of structural plant/model mismatch resulting from the use of linear blending models, and parametric mismatdi aiising frorn fluctuations in reforrnat e quali t ies.
Blended RVP
RVP Regular
RVP Premium
+ Bias Controller
4.5
As discussed in Chapter 2, linear gasoline blending models are not able to capture the tme
blending behaviour because most important gasoline qualities blend nonlineariy. Since blending nonlinearity is one of the major sources of blending inaccuracy, accurate modeling of properties such as octane number and boiling points is crucial to the success of the blend controiler [Ramsey Jr. and Truesdale, 19901. Recognizing the problems associateci with blending nonlinearity, some refiners have tned incorporating nonlinear blending models in their controllers.
An approach to incorporating nonlinear blending models has been to linearize the models and
then use sequential iinear programming (SLP) which retains the controiler formulation given in Problem (4.4) (e.9. Diaz and Barsamian, [1994]). Another approach is to incorporate the nonlinear rnodels directly in the blender control problem by formulating the on-line optimizer as a nonlinear programming (NLP) problem with bias updating (e-g. Ramsey Jr. and Truesdale.
19901:
subject to:
where: g(x) is a vector of nonlinear blending models. Sote that Problem (4.5) differs from the LP
the blending models used and can be solved using standard NLP techniques [Avriel, 19761.
-+ bias update controller given in Problem (4.5) was used on the benchmark autome
tive gasoline blending problem posed in Section 4.2. The controller was designed to eliminate al1 sources of plant/model mismatch by using the same blending models as used in simulating the blending process (see Section 2.4). Since the feedstock qualities used in the
their nominal values for al1 feedstocks except reformate, bias updating was used to compensate for fluctuations in reformate qualities. Product demand and feedstock availability limits were
( Quaty
Deviation
MON 1 RVP (reguiar) (regdar) (pr emiurn) (prernium) 0.015 psi -0.05 -0.1 0.01 psi
RON
RVP
Table 4.5: Deviation rom Specification at End of Blend the same as used in the LP
+ bias update case and the quaiity constraints were of the form:
where: gl and g2 are the RON blending mode1 applied to regular and premium grades, respectively, g3 and g4 are the MON blending mode1 for regular and premium grades, and g~ and g6 are the R V P blending mode1 for regular and premium gasoline. The blending models are given
in Equations ( 2 . 1 6 ) .
Simulation of blending control was conducted as descri bed for the LP+bias update case except that the blend recipe at each RTO interval was obtained using ,ULP with MINOS in GAMS [Brooke et al.? 19921. The cornplete blend recipe can be found in Appendix D. Upon completion of the blend, four of the blended qualities in the product tanks were found to be off-specification
as summarized in Table 4.5.
For the purposes of this study, deviations of up to -0.03octane nurnber for RON and MON and f0.005 psi for RVP were deemed acceptable. These tolerance levels were satisfied for both grades when gasoline produced during the l S t , g t h , loth, and 121h intervals were excluded form the final products. The excluded gasoline was routed for reblending. Eliminating products from 4 of the 12 RTO i n t e n d s resulted in much smaller batches and the profit earned from the blends is $44,105.96, much Iess than the theoretical maximum of
$66,276.55. However, performance of the NLPtbias update controuer is far better than t hat
of the conventional LPfbias controller used in Section 4.4.1. Thus eliminating structural m i s mat ch from the conventional LP +bias cont rouer improves cont roller performance drarnat i c d y . However, the bias updating strategy is not able to adequately handle pararnetric mismatch arising from stochastic feedstock disturbances, resulting in the production of off-specification products.
4.6
TIME-HORIZON CONTROLLER
Cornparing the performance of the LPtbias controller of Section 4.4.1 and the XLPtbias controller of section 4.5.1 clearly shows that the use of appropriate nonlinear blending models is crucial to the success of the blend controller. By eliminating structural plant/model mismatch. a very significant improvement in controller performance was obtained for the benchmark problem. However, as shown in the case-study, the resulting NLP+bias updating controller can still lead to the production of off-specification blends when disturbances are present in feedstock qualities. This is because in the bias updating structure, the feedstock qualities are treated as Lied parameters in the controller while they vary in the actuai plant. The resulting parametric mismatch cannot be compensated for by bias update [Forbes anci Marlin, 19941. Thus, such a blender control strategy is not well suited for control application where feedstock qualities are fluctuating. Several control practitioners have reco,dzed the problems associated wit h Auctuating component qualities and have offered some partial solutions by incorporating multi-period optirnization capabilities into t heir blender control systems (e-g. SlcDonald et al. [l992], Rigby et al. [l9951). These methods approach the problem of feedstock quality changes by infrequent feedback of feedstock properties hom laboratory analyses, and predicting future control actions using these laboratory based measurements. The update intervai for such systerns is typically a day and component qualities are assumed to remain unchangeci in the future. Thus, they cannot effect ively deal with the higher frequency dist urbances associated wi t h upstream process operat ion changes. Another approach involves using mode1 predictive control technology where on-line blended quality measurements are used to predict disturbances in t hese qualit ies [Vermeer et
a l . , 19961. However, in their work Vermeer e t al. assumed disturbances in the blended qualities
as step disturbances that persist over the b l e d horizon. Although a move in the right direction.
t his met hod does not utilize noise modeis that exploit feedstock quality disturbance dynamio to predict feedstock qualities in the future. As a result, existing methods cannot adequately
15 years since their development . MPC systems have been widely appiied to systerns subject to
stochastic disturbances [Garcia et al., 19891, [Ogunnaike and Ray, 19941. The success of MPC systems c m be attributed in large part to their ability to exploit knowledge of process dynarnics to predict future plant behaviour using appropriate modeis. However, to date, such strategies have not been appiied to RTO systems. In this work, the disturbance handling capabilities
of MPC systems are adopted in developing a new blend controller which is based on insights
provided by the ideai controller, The ideal controller in Problem (4.3) cannot be implernented because it requires future knowledge of feedstock qualit ies. Although perfect knowledge of feedstock qudit ies at al1 t imes during
the blend is not realizable' past feedstock qualities c m be used to predict these qualities into
the future. Thus at each RTO intend! predicting feedstock qualities during the remaining blending time horizon ni11 allow future control actions to be based on knowledge of feedstock behaviour over the whole blend horizon ( p s t and future). This requires the blending problem to be posed as an optimization probIern over the whole blend horizon as in the case of the ideal controller. This approach generates blend recipes for the rernaining blend horizon. The recipe for the current time-step can be implemented and the process of prediction/optimization repeated at each time-step in a receding horizon fashion until end of blend is reached. uch a blender control strategy is very similar to MPC in that they both involve predicting disturbances over a time horizon and basing control action by optimizing over a time horizon. In both cases, only the current control action is implernented and the whole process is repeated in receding horizon fashion. However, while the length of the prediction as well as the control (optimization) horizons are tuning parameters in MPC, only the prediction horizon is w r tuned in the blend controller. This is because the optimization horizon in the blend controller
is h e d at each RTO i n t e d as the remaining blend horizon. The blender control strategy discussed above and the ideal controller work towards meeting
quality specifications only at the end of the blend. If blending has to be tenninated prior t o
reaching end of blend, the products made thus far could be far Erom being on specification. In order to guard against such situations? an additionai constraint m a y be added to the opt irnization problem that would ensure that product specifications are met at ail times in the future. This control strat egy can be mat hematically formdated as:
subject to:
where: R is a rnatrix containing past and present measured feedstock qualities. R is a matrix containing the predicted feedstock qualities, g is a vector of blending equat ions which rnay be linear or nonlinear. and t,. t,, tf are the initial. present, and final blend times, respectiveiy, and h contains product demand and feedstock availability constraints. x~te that the Problem (4.6) is very similar to the perfect controller in Problem (4.3). The proposed blend controller differs from the ideal controller in that the blend horizon is divided into two parts: past (and present) and future. Also, Problem (4.6) contains an additional constraint
to ensure that product specifications are met during blending. The objective functions in P r o b
lems (4.6) and (4.3) are e q u i d e n t even though Problem (4.6) integrates blending economics
only from t, to tl rat her than over the whole blend horizon. This is because the economics due
to the past blended products is a constant and will not affect decisions on future control action.
In other words, future control action cannot change economics due to past blended products.
look a t predicted end of blend quaiities which are composai of past blended qualities (toto t p )
and future predicted blended qualities. The future predicted blended qualities are obtained by
integrating the blending models Erom t p to t f using predicted feedstock qualities. ~ ( t Feedstock ) . qualities c m be predicted using conventional prediction filter techniques (e-g. Box and Jenkins.
[1976], Chiu, [1987]). These constraints emure that products meet quaiity specification at end
look at future blended quaiities based on blending models and predicted feedstock qualities. These constraints attempt to blend products such that each partial batch made during individual RTO intends meet specifications. Finail- the third set of constraints:
integrate m a s baiances fkom the start of the blend to the end of the blend in order to satisfy ail demand and availability limits over the whole blend. Like the ideal controller. the time-horizon based controller in Problem (4.6) can be solved using dmarnic programrning [Bellman and Lee. 19841 or the problem can be discretized and then solved with SLP [Avriel. 19761. The blender control problem in Problem (4.6) is unique compared to other blend controllers in that . at each RTO inten-al. the controller look back in time to what has been bIended so far while anticipating future trends in feedstock qualities. It then adaptively alters the blend recipe in order to maxirnize profit while satisfymg al1 constraints. The advantages of the time horizon formulation are that during each RTO interval the controller has an opportunity to:
1. cornpensate for past rneasured off-specification products:
blending case-study of Section 4.2. The time-horizon based controller requires that some of the feedstock qualities be forecasted into the future. First. though, the feedstock qualities that are to be updated on-line and forecasted have to be identified. It is desirable to keep the number of adjustable parameters Iow
since the more parameters that need to be updated on-line. the more measurements t hat would
o capture be required. However. the selected adjustable feedstock qualities should be able t
the effects of major disturbances. Selection of adjustable parameters for RTO systenis is discussed in Section 3.4 where a selection strategy based on the Design Cost criterion (Forbes and Slarlin. 19961 is recommended. The suggested adjustable parameter selection strategy incorporates parameter observability information. Parameter obsenability is discussed in Chapter 3 where a new mesure of obsembility is dewloped based on fundamental statisticd principles.
Even when a rigorous selection strategy is not applied. those parameters that exhibit strong
obsenability (have well condition4 joint confidence regions) from the anilable measurements and aven estimation scheme should be selected orer those that only have a Ioa- degree of observability.
constraints were as describeci for the ideal controller in Section 4.2. In this simulation case-study. disturbances are known to be present only in reformate RON.
not have to be selected. h o . it was assurneci for this case-study that the reformate qualities
test on these
parameters was not conducted. Generdy, however, feedst ock qualities are not rneasured and have to be estimated from other rneasurements. In such situations, obserwbility (fiom the available measuremeats and estimator) of the qualities to be forecststed wiIl have to be ensured using the methods developed in Chapter 3.
The t hree reformate qualities were forecasteci in the time-horizon based controller using a predicn this simulation study. the actud noise model was used in designing the prediction tion filter. I
fter. In general. the noise mode1 would not be h o a n and can be developed using identification techniques [e.g. Ljung, 19871. The prediction filter made minimum mean squared error forecasts [Box and Jenkins, 19761 of the reformate qualities using the noise model given in Equation
(4.2). The prediction filter equations are provided in Appendix H.
Three time-horizon based controllers were designed which differed in their size of the prediction
horizon. 1-step. Zsteps. and Psteps ahead predictions were w d . The prediction filter used
the noise model given in Equation (4.2) and two samples of historical reformate quality data
to make 1-step. 2-steps. and 3-steps ahead minimum mean squared error forecasts [Box and Jenkins, 19761 of the reformate qualities for each F U 0 i n t e n d . The qualities were forecasted at each RTO intenal rom t , to tf-i and the predictions were held constant for al1 i n t e n d s b e o n d the predict ion horizon. The three controllers (with 1. 2. and 3 step ahead prediction horizons) were used on the blending problern with SLP usinp MISOS in GAMS [Brooke et aL. 19921. The results from the controllers at each RTO instant are provided in Appendices ElF. and G. While controllers with prediction horizons of 2 and 3 steps yielded products meeting quality tolerance levels?
ROX and R V P in regular g a ~ l i n e produced using 1-step ahead prediction filter failed to meet
quali ty specifications wit hin the specified tolerances. Evcluding regular gasoline produced during the last inten-al brought the ROX up and the R V P d o m to within tolerance a t the cost of reducing the size of the blend. thus decreasing the profit level. Profit earned kom the controllers for the 24 hour blend are given in Table 4.6.
S66.3r6.55 but performance of tirne-horizon controllers increase slightly as the size of the prediction horizon increases since longer prediction horizons are able to exploit more of the noise dynarnics.
4.7
DISCUSSION
-4 gasoline blending case-study has been used in this chapter to compare the performance of
- bias
update controller
U'~LSfound
to perfom quite
poorly on the benchmark problern and it has been thown that the conventional controller can lead to a large performance l o s when stochastic disturbances are present in feedsrock qualitiec.
- bia. con-
u ' e d ro compensate for parametric misrnarch resulting from fluctuations in reformate qudities.
The SLP - bias controller a-as found to proride si,dcant performance improvements O\-erthe conventional LP
turbances are present in the feedstock qualities. b~ using the proposeci the-horizon based controlier. T h e p r o p o d controller diffen from the 5 7 2
it
optirnizes the blending problem over the a-hole blend horizon rather than at a point in time.
In addition. it updates the feedstock qualities directl- thus addressing the issue of parametric
mismatch directly. The performance of three tirne-horizon based controllers and the hT,P
bias controller are compared in Figure 4.5- Controikr performance i s based on the amount
of theoretical maximum profit uncaptured by each controiier. The profit levels for the h;LP
Controller Performance
approach. T h i s is because the time-horizon controllers update feedstock qualities directly and
work to rnavimize profit over the whole blend horizon rather than at a point in time.
Chapter 5
T h i s thesis has focused on the design of the gasoline blend controller. The widely used blend
controiler (LP
not capable of effectively handng feedstock quaiity disturbances as shoan in Chapter 4. In t his t hesis, a blend controller that incorporates suitable gasoline blending rnodels and one t hat
can effectively handle predictable changes in feedstock qualities has b e n presented and shown
to out-perfonn the conventional controllers. Chapter 2 examines gasoline blending modeis for use in the blend controller. Slodek for octane number. Reid \=pour pressure' and A S D I distillation points are discussed with respect to predictive accuracy and ease with whch the model can be updated on-ne in order to maintain
its predictive accuracy. The Ethyl RT-70 [Heaiy et a l . . 19591 model for octane number blending, and the Blending Index method [Gary and Handwerk. 19941 for R V P blending were found to
be the most suitable for contrai purposes arnong published models. Since some of the mode1 parameters would have to be updated on-line in the blend controller, it i s important to ensure that estimates of these parameters be can obtaine using the available measurements and chosen estimator. Chapter 3 develops a method of determining parameter observability for such steady-state systems.
The design of the new blend controiler that incorporates appropriate nonlinear blending rnodels
is then presented in Chapter 4. The proposed controller adopts concepts bom MPC systems
5.2
CONCLUSIONS
The use of LP with bias updating for the control of gasoline blending can lead to a large performance loss when stochastic disturbances are present in feedstock properties as shown in the case-study in Chapter 4. A s a first step towards capturing more profit, plant/model mismatch should be rninimized. The use of 'JLP with bias updating increases controller performance, but
can still lead to the production of off-specification blends as illustrateci in the case-study. This
is because in. the bias update approach, the feedstock qualities are treated as fixed parameters even though they are changing in the real process. The time-horizon b a s 4 blend controller proposed in this thesis directly addresses the issue
of feedstock quality fluctuations by treating these qualities as adjustable parameters that are
updated on-line. Moreover, the control strategy predicts feedstock quali t ies over the blend horizon to plan blending operations over the entire blend horizon. At each RTO interval, the controller predicts feedstock qualities over a time horizon. and optimizes the blending problem over the remaining blend horizon. It then implements control action for the curent time-step and repeats the process at the next time step in a receding horizon fashion until end of blend. Thus the proposed time-horizon controller can be thought of as an economic optirnizing model-predictive control system t hat is analogous to the process control technology that petroleum refiners have widely deployed throughout t heir plants [e.g. Cut ler and Ramaker,
1979j. Although computationally more intensive, the time-horizon based controller offers a new
approach to handling feedstock quality disturbances, and c m provide efficient control of the gasoline blender as iUustrated in Chapter 4.
The tirne-horizon RTO approach developed for the blend controller can be applied to on-line
optimization of any process that experiences stochast ic disturbances t hat c a w the optimal operations to drift with time. The proposed RTO approach would predict trends in major disturbances entering the plant and optimize the steady-state process over a time horizon so as
to include predicted future plant behaviour in determining optimal operatiom. In addition. the time-horizon optimization approach can be applied to off-line optimization applications such
as refinery planning and scheduling where predict able trends exist . Here again, predict able
disturbances can be exploited in order to optimize operations over a period of time rather than at a point in time. The parameter observability test developed in Chapter 3 is based on fundamental statistical principles. and has been shown to provide a necessary and sufficient condition for obsenability of parameters hom available measurements and estimator for steady-state systems. The approach naturally leads to a measure of degree of observability that looks at the joint confidence region of the parameter estimates. The observability and degree of observability concepts developed in this thesis can be used in the design of al1 RTO systems. The tests can be used for selecting adjust able parameters such that t hey have an acceptable degree of observability. Moreover, the observability concepts and tests developed in Chapter 3 are not limited to RTO systems. Rat her, they are applicable to any steady-state system where secondary measurements are used
to obtain estimates of unrneasured quantities.
Future work stemming fiom this thesis include applications of the parameter observability work of this thesis for RTO design. as well a s further work on the design of the gasoline b l e d cont rouer. The approach to determining parameter observability developed in this thesis can be used for developing RTO design practices. That is, it can be used as a basis for selecting adjustable mode1 parameters as well as for selecting process measurements for updating these parameters. In developing a met hod for select ing adjustable parameters and/or measurements, the test developed in this thesis can be used to ensure that the adjustable parameters exhibit an acceptable degree of observability.
The time-horizon blend controller is based on forecasting some of the feedstock qualities into
the future using a prediction filter. The prediction filter only looks at the behaviour of the feedstock qualities in the past in order to make future predictions. It does not directly include
any information on upstream process operations changes and how these changes affect blender
feedstock qudities. The proposed blend controller can be improved by developing models
between some upstream process variables (measured disturbances) and the feedstock quaiities of interest. The models can then be used in a feed forward configuration in order to aUow the contrder to handle feedstock disturbances (arising fkom measured changes in upstream process operations) more effectively.
Also, the developments of this thesis have only been applied to the design of the RTO layer in
the blend controller. The ideas should be extended to the off-Iine optimization layer as well. That is, blending models selected in Chapter 2 and the time-horizon optimization approach developed in Chapter 4 shouid be adopted in the off-Iine scheduler and planner for increased bIender efficiency.
Nornenclature
system matrix in Problem (3.1) aromat ics content ( % volume) correlation coeEcients in Ethyl RT-70 model, i=1,..,6 constant in the Stewart model for octane blending constant in Soave-Redlich-Kwong equation of state constant in Soave-Redlich-Kwong equation of state matrix in Problem (3.1) that relates rate of change of state variables to controlled variables bias vector constant in Soave-Rediich-Kwong equation of st ate matrix in Problem (3.1) that relates observations to states cost vector random variable defined in Problem (3.7) random variable defined in Problem (3.6) constant in the Stewart method for octane blending parameters in Transformation method
variable for feedstock i defined in the Stewart method for octane blending
number of gasoline grades scalar disturbance at time t mass of feed to RVP test chamber probability density function of parameter estirnates vector of blending models adjustment factor for hydrocarbon type in Transformation method vector of demand and availability limits
interaction coefficient between feedstocks i and k in octane blending interaction coefficient between feedstocks i and k in R V P blending filter matrix constant in Soave-Redlich-Kwong equation of st ate number of components over all feedstocks parameters in Transformation met hod pararneters in Zahed rnethod Lagrangian of optimization problem
mass of liquid phase
observability matrix for dynarnic systems model parameter in Zahed method model pararneters in Zahed method, i = 1,...,n motor octane nurnber molecular weight of i union of a finite number of convex sets that defines plant operation region normal distribution convex regions that define the plant operating region, Nt number of feedstocks in blend olefin content (% volume) paraffin content (% volume)
pressure
cW
number of adjustable pararneters covariance mat rix of parame ter est imates covariance matrix of process measurements model parameter in Zahed method model parameters in Zahed method, i = 1, ..., n
vector of reformate qudities
research octane number feasible set of states in steaciy-state system closure of set S consisting of S and al1 its limit points vector of quality specifications
octane sensitivity (research octane number - motor octane number)
steps of random height at time t random impulses at tirne t ternperat u r e vector of contolled variables volume fraction of feedstock i in blend set from which rneasurement noise realizations are obtained m a s of vapour phase in R V P test chamber volume of feedstock i in blend measurement noise vector
gas volume
percent volume evaporated expanded liquid volume weighting vect or decision variables in gasoline blending problerns (feedstock flowrates)
P state variables E 3
rate of change of y initial states of process process measurement vector
backward shift operat or
constants in blending mode1 in Equation (2.14) constants in blending model in Equation (2.15) adjustable parameters E !RP adjustable parameter estimates E 9ip
mean of adjustable parameter estirnates, fi
activity coefficient of component i vector of steady-state equations that reiate measurement to States
L2 n o m
ith eigenvalue of Q8
Lagrange mult ipliers mole fkaction of cornponent i in the liquid phase vector of residuals in Problem (3.13) noise vector in Problem (3.1) initial noise vector in Problem (3.1) density of i at 100 OF (pound-moles per barrel) molar density of component i (pound-moles per barrel a s liquid at 1 atm and 60 O F )
Pb1end
O F
o2
objective h c t i o n of optimization-based estimation problem in Problem (3.13) map (possibly nonlinear) representing pararneter estimation probiem matrix of operating conditions in Example 3.1 vector of steady-state plant model equations
OPERATORS
V
gradient Hessian determinant of matrix A
transpose o f matrix A
V*
IAl
SUPERSCRIPTS
*
optimalvalue predictedvalue
Glossary of Terms
ASTM bias updating American Society for Testing and Materials mode1 updating by applying an error tenn
BON
LP
LSR 1MIP
MON MPC NLP
ON
PD
RON RTO
Real-Time Optimization Reid Vapour Pressure Reid Vapour Pressure BIending Index Sequential Linear Programming
RVP
RVPBI SLP
Bibliography
Agrawal, S.S. (1995). Integrate blending control, optimization, and planning. Hydrocarbon
Avriel, M. (1976). Nonlznear Programming: A nalysis and hiethods. Prent ice-Hall. Xew Jersey. Bain, M.L., K.W. Mansfield, J.G. Maphet, R.W. Szoke, W.H. Bosler and J.P. Kennedy (1993).
Gasoline blending and integrated on-ne optimizat ion, schedung and control systern.
Box. G.E.P. and G.M. Jenkins (1976). Time Series Anolysis: Fonoc~stzngand Control. HoldenDay. Oakland, CA.
Box, G.E.P. and H.L. Lucas (1959). Design of experirnents in nonlinear situations. Biometrika
46, 77-90.
Box, G.E.P. and W.G. Hunter (1965).Sequenriai design of experiments f o r nonlinear models.
iBM Scientifi Synposium in Statrstics pp. 113-137.
Brogan, W.L. (1991). Modern Control Theoq. Prentice Hall. Sew York. Brooke. A.. D. Kendrick and A. Meeraus (1992). GAMS.. A Userts Guide. Release 2.25. Boyd and Fraser Publishing Co. ~lassachusetts.
B q w n . A.E. and Y.C. Ho f 1976). Applied Optimal ContmL Halsted Press. Sew York.
Char. B.iV.. K.O. Geddes. G.H. Gomer. B.L. Leong. 31.B. Monogan and S.M. Watt 119911.
Maple Lunguage Reference ManuaL Waterloo Sfaple Publishing. iVaterloo. OS. Canada.
Chatfield. C. and A.J. Collins (1980). Intmduction to .hfdtivariate .-inaly&. Chapman and Hall.
Sew York.
Chiu. C.K. (1987). KaZman Filtering u?th Real- Time Applicattons. Springer-Verlag. Sem-York. Corain. L.J. and R.H. Szczarba (1995). Calculus in k t o r Spaces. Marcel Dekker. Sew York. Crom-e. CAf. ( 1988). Recursive identification of grossi errors in Iinear data reconciliation. -4IChE
J. 34. .%l.
Curler. C.R. and B.L. Rarnaker (1979). Dynamic rnatrix control: A cornputer control algorithm.
Dhulesa. H. f 1984). Equation fits ASTM distillations. Hymcurbon Processzng Sept .. 17S180.
Diaz. -4 and J.A. Barsamian (1996j. Meeting changing fuel requirements uirh online optimization. Hydroclzrbon Processing Feb.. 11-76. Draper. S.R. and H. Smith (1967). Applzed Rqessron Analysis. John IViley 61 Sons. Xew York. Duran. .\I.A. and LE. Grossman (1986).A mixed-integer nonlinear progamrning algonthm for process -cystems s_vnthesis. -4iChE J. 32. 593-606.
Edgar, T.F. and D.M. Hirnmelblau (1988). Optirniration of Chernid Pmcesses. McGraw-Hill.
Yew
York.
Farley. K.R. and J .P. Kennedy (1996). Constructing a refinery informat ion systern. Hydmcarbon Prvcessing Sept ., 43-44. Fletcher. R. (1987). Pmctical Methods of Optimization. John Wley 6; Sons. Te%York. Forbes, J.F. (1994). Mode1 Structure and Adjustable Parameter SeIection for Operations O p timizat ion. P hD thesis. Mcblaster University. Canada. Forbes. J.F. and T.E. klarlin (1994). Model accuracy for economic optimizing controllers: The bias update case. Ind. Eng. Chern. Res. 33, 1919-1929. Forbes. J.F. and T.E. Marlin (1996). Design cost: A systematic approach to technology selecrion for mode1 based real-time opt irnizat ion systems. Computers C'hem. Engng. 20. 717-734. Forbes. J.F.! T.E. Mariin and J.F. SlacGregor (1994). Model adequacy requirements for optirnizing pIant operations. Computers Chern. Engng. 18. 497-510.
Friedland. B. (1986). C o n t d System Design: A n Introduction to State-Space hfethods. SlcGraw-
AIChE
Garcia. C.E.. DAI. Prett and SI. Slorari (1989). >Iode1 predictive control: Theory and practice
Golub. G.H. and C.F. Van h a n (1989). Matriz Computations. John Hopkins University Press. Baltimore.
Haskell, 3.B. and D.K.Beavon (1942). Front end volatility of gasoline blends. Ind. Eng. Chem.
(2). 167-170.
Healy, W.C.Jr.. C.W. Maassen and R.T. Peterson (1959). A new approach to blending octanes. In: Pmc. 24th Midyear Meeting of the Arnerican Petmleurn Institute 5 Diuision of Refining. Vol. 39. Sew York.
Krishnan, S., G-LV. Barton and J.D. Perkins (1992). Robust parameter estimation in on-line optirnization- part 1 : Methodology and simulated case-study. Cornputers C'hem. Engng. l6(6), 545-562Kuttelwascher. H. and H. Rhemam (1995). Tank farm automation wirh a DCS. Hydmcarbon
Loo P.T. ( 1994). On-iine rvp analysis improves gas blending. INTECH Sept.. 3 1-33.
Loeblein. C. and J.D. Perkins (1992). Estimated parameter and measurernent selection for on-line optirnization based on economics. ICHEME Reseanh Euent, Eumpeun Conf. for
Young Engrneers 2. 757-759.
MacFarlane, R.C. and D. LV. Bacon ( 1989). Empirical strategies for open-loop on-line opt imization. Can. J. Chem. Eng. 67. 665-677.
Magodas. K.. D. .\larinusKouris and A. Lygeros (1988). Instructions are given for building
McDonald, J.M., L M . Grupa Jr., M.P.McKibben and D.E. Roe (1992). Gasoline blending using
McLellan, P.J. (1981). Thesis. A Nonlznear Refinery Model for Schedulzng and Econornic Evaluation, M.A. Sc. Thesis. University of Waterloo. Canada. Michalek, T.F., K. Xordeen and R. Rys (1994). Using relational database for blend optimization. Hydrocarbon Pmcessing Sept., 47-49. Miletic, 1. (1996). Thesis. Results Analysis in Real-Time Optirnitution. ter University. Canada. Montgomery, D.C. (1991). Design and AnaZysis of Eqerirnents. John Wiley & Sons. Xew York.
Morris, W.E. (1983). Prediction of mogas blend distillations can be improved. Oil
G u J.
Morris, W.E., W.E. Smith and R.D. Snee (1994). Interaction blending equations enhance reformulated gasoline profitability. Oil
Gus
J. Jan. 17,54-58.
Muller, -4. (1992). New method produces accurate octane blending values. 0 2 Zt 3 Gas J. Mar.
23,80-90.
Ogunnaike, B.A. and W.H.Ray (1994). Process Dynamics, Modeling, and ContmL Oxford
University Press. New York.
Palmer, F.H.and A M . Smith (1985). The performance and specification of gasolines. In: TechHancock, Ed.). Blache11 Scientific Publications. London. nology of Gasoline (E.G.
Palmer, R.E., J. Bar-Dov and S. Whitaker (1995). Use an integrated approach to product in-iine blending. Hydrocarbon Processing Feb., 62-66. Ramsey, J.R. Jr. and P.B. Truesdale (1990). Blend optimization integrated into refinery-wide strategy. Oil 4 Y Gas J. Mar. 19,40-44. Ray, W.H. (1981). Advanced Pmcess Contml. McGraw-Hill. New York.
Rigby, A.' and L.S. Lasdon and A.D. Waren (1995). The evolution of texaco7sblending systems:
From omega to starblend. interfaces 25, 64-83. Roberts, P.D. (1979). An algorithm for steady-state optimization and parameter estimation.
Rusin, M. (1975). The structure of nonlinear models. Chem. Engng. Sn'. 30,937-944.
Rusin, M.H.. H.S. Chung and J.F. Marshall (1981). A "transformation7' rnethod for calculating
the research and motor octane nurnbers of gasoline blends. Ind. Eng. Chem. Fundam.
20. 195-204.
Seferlis, P. (1995). Coilocation Nodels for Distillation Units and Sensitivity Analysis Studies in Process Optirnization. PhD thesis. MccMaster University. Canada. Serpemen, Y., F.W. Wenzel and A. Hubel (1991). Blending technology key to making new
gasolines. Oz1 6 GUSJ. Mar. 18, 62-74.
Snee, R.D. (1981). Developing blending rnodels for gasoline and other mixtures. T~cf~nometrics 23, 119-129. Sorenson, H.W. (1980). Pammeter Estimation: Princzples and Problems. Marcel Dekker, Inc.. Xew York.
Stanley, G.M. and R.S.H. Mah (1981). Obsenrability and redundancy in process data estimation.
leum products. Industrial and Engzneehng Chemistry 46, 2 182-2185. o r gasoline blends. Petmleum Refiner 38, 135-139. Stewart, W.E. (1959a). Predict octanes f
Stewart, W.E. (19596)- Predict rvp of blends accurately. Petmleum Refiner 38: 231-234.
Sullivan, T.L. (1990). Retiery-wide blend control and optimization. Hydmcarbon Pmcessing
May, 93-96.
Sutton, T.L. and J.F. MacGregor (1977). The analysis and design of binary vapour-liquid equilibrium expenments, part II: The design of experiments. Can. J. Chem. Eng. 55,602608.
Tjoa,
1-B. and L.T. Biegler (1990). Simult aneous stratedes for data reconciliation and gross
error detection of nonlinear systems. AIChE J. Annual Meeting. Turing, A.M. (1948). Quart. J. Mech. Appi. Math. 1, 285. Vazquez-Esparragoza, J.J., G.A. IglesiasSilva, M.W. Hlavinka and J. Bullin (1992). How to estimate RVP of blends. Hydrocarbon Pmcessing Aug., 135-138. Vermeer, P. J.. C.C. Pedersen, W . 1 1 . Canney and J.S. Ayala (1996). Desing and integration issues for dynamic blend optirnization. NPRA Comp. Conf. Nov., paper no. CC-96-130. White, J. and F. Hall (1992). Gasoline blending optirnization cuts use of expensive components.
Oil fi Gus J. Nov. 9, 81-84.
Zahed, A.H., S.A. Muiiah and M.D. Bashir (1993). Predict octane number for gasoline blends.
Hydrocarbon Pmcessing May, 85-87.
Feedst ock
reformate
Regular
5430
LSR naphtha
n-butane catalytic gas. alkylat e
2570
1.99
O
O
1-
RVP (psi)
10.8
Lagrange multipiier maximum RVP in regular gasoline maximum RVP in premium gasoline minimum RON in regular gasoline minimum MOX in premium gasoline $1737 $2023
9
$
$3110.5 @
$5679.3
$1.57
$4.58
"
bbl
$2.38
$0.15
$4.36
$0.43
LSR Naphtha
Cataiytic Gas.
Reformate
LSR Naphtha
Regular
Premium
1
Constraint
l
RON
MON RV P (psi)
10.8
maximum availability of catalytic gas. maximum demand of regdar gasoline maximum demand of premium gasoline
Activc coristraints ovcr wholc blciid
10.8
F i i i d Blendcd Qunlitics
C0NSTRA.INT
maximum RV P in regdar gasoline
1
1
~
I N T E R V A L
2
X
1
,YX
4
X
5
X
7
X
8
~
1 0 1 1 1 2
X
x x
X
x
X
x
~
x
X
x
X
x
X
x
X
x
X
x
X
x ~
x
x x x x x x x
x x x x
x
x x
x
x x x x x
x
x
x
x
x
x x x x
x x x x x
x
x
x x x x
x
x
x
.Y
x ~
x
x
x
x x x
x x
x x
x x
Reformate
LSR Naphtha
O. 157 O. 175
Cataiytic Gas.
Alkylate
0.204
0.306
0.191
0.186
O. 186
1 RTO Interval
1
Reformate
LSR Naphtha
Catalytic Gas.
Alkylate
O
O
Regular
Premium
92.8
ROrV
89.6
2
x
3
x
4
x
5
x
7
x
8
x
1
x
x
0
x
1
x
1
x
2
x
4
x x
x
,x
1
x
x
x
x
x l x I
x
~
x
x x
x
x
x
x
~
x
x
x
x
x
x
x
x
x
x
x x
x
x
x x
x x x
x
)x
.Y
x
x
x
x
x x
x
x
x
x
x
x
x
x x
x
x
x
x
x x
x
x x
x
x
I
x
x
x
x
x I x
x x
x
x
x
x x
The blend recipe implemented by the controller at each RTO intervai is given below:
- - -
1 RTO Interval
Reformate
o. 220
0.1 15
0.111 0.114
0.113
O. 109
O. 109
0.110
o. 102
O .O90
0.087
The folloning constraints were found to be active during al1 RTO i n t e n d s oves the blend:
1. maximum R V P in re,oular gasoline
gasoline
9. maximum *late
to premium gasoiine.
Reformate
LSR Xaphtha
Reformate
LSR Naphtha
Regular
Premium
Constraint
1
gas.
R V P (psi)
Final BIcndrvl Qiialitic~
The following constraints were found to be active during each RTO intenal for the "current"
t i meet ep:
111
i n premium gasoline
---
Reformate
Catalytic Gas.
Alkylate
The following constraints were found to be active for the whole bknding problem:
1. maximum availability of catalytic gas.,
The foilowing constraints were found to be active during each RTO interval for the "current"
time-step:
1- maximum R V P in reguiar gasoline,
114
--
. -
RTO Interval
Reformate
Catalyt ic Gas.
0.072
O O
O
O
O O O
O
5.2*10-~
O 0.007
2.4*10-~
O O O O
"
Reformate
LSR Naphtha
The foilowing constraints were found to be active over the blending problem:
1. maximum availabili ty of catalyt ic gas..
The following constraints were found to be active during each RTO interval for the "current" time-step:
1. maximum R V P in regular gasoline,
In order to predict the reformate qualities? q,.f,, , the scalar disturbances are first calculated
using the above equation (with known Be,)The scalar disturbances are then forecasted and
, ! used to calculate predictions of q
where: si are random impulses that are integrated to give steps of random height,
st.
One, two. and three step ahead forecasts of dt are made using minimum mean squared error forecasts [Box and Jenkins? 19761. The filter equations are given below:
where:
Forecasting was started using two samples of historicai data and by initidy setting si-, and
si to zero. The impulses were then calculateci using:
APPLIED & IMAGE. lnc 1653 East Main Street -. Rochester. NY 14609 USA
-- Fa: --I