1. Implementation of Coordinator MPC on a Large-Scale Gas Plant
Elvira Marie B. Aske*&, Stig Strand& and Sigurd Skogestad*
*Department of Chemical Engineering
Norwegian University of Science and Technology (NTNU)
Trondheim, Norway
&StatoilHydro R&D, Process Control, Trondheim, Norway
Session title:
Distributed Optimization and Control of Large Scale Systems
20 minutes presentation
AIChE Annual Meeting, Philadelphia, USA
November, 2008

2. Outline Introduction and motivation
The Kårstø gas plant
Maximum throughput as optimal operation
Approach: Coordinator MPC*:
Maximize flow through linear network
Estimate feasible remaining capacity (R) in units using local MPCs
Application to Kårstø Gas Plant
Previous work*: Works well on simulations
Here: Actual implementation
Design
Tuning (plant runs)
Experiences
Conclusion
*Aske, E.M.B., S. Strand and S. Skogestad (2008). Coordinator MPC for maximizing plant throughput. Comput. Chem. Eng. 32(1-2), 195–204.

4. Kårstø plant Gas processing area Control room
A picture of the Kårstø plant. The main gas processing area where the distillation columns, and most of the utility is located, like boilers and compressors, and also export compressors.
The tank area and harbor where storage tank for ethane, butane, naphtha and condenaste. Liquid propane are stored at Kårstø in two large rock caverns.
These caverns provide a combined capacity of 250,000 cubic metres. Each cavern is 195 metres long, 20 metres wide and 33 metres high. The carverns can be filled with 570 cubic metres of liquid propane per hour.
Some producing numbers:
Sales gas produced: 27 bn scm per year= 21 million tons of liquid
Liquids production in tonnes/year:
Ethane
Propane
I-butane
N-butane
Naphtha
Condensate
Total
The gas via Kårstø corresponds to 300 TWh/year.

5. Nyhamna
Europipe II
Europipe I
Norpipe
Emden
ÅTS
Norne
Åsgard
Haltenpipe
Heidrun
Franpipe
Zeebrugge
Zeepipe I
St Fergus
Vesterled
Frigg
Statfjord
Kårstø
Kollsnes
Melkøya
Snøhvit
Ormen Lange
Easington
Langeled
Ekofisk
Sleipner
Troll
Dunkerque
Kristin
Tjeldbergodden
North Sea gas network
Kårstø plant: Receives gas from more than 30 offshore fields
Limited capacity at Kårstø may limit offshore production (both oil and gas)
Norwegian
continental
shelf
TRONDHEIM
Oslo
Kårstø is a gas processing plant that plays an important role in gas transport and treatment.
It receives rich gas and condensate from more than 30 offshore producing fields.
This set high demands, not only to the plant efficiency and regularity, but
also to the plant throughput. Limited gas plant processing capacity means that one
or more fields must reduce production or even shut down. Therefore it is important
for the Kårstø plant to process as much as possible and not become a “bottleneck” in the Norwegian gas transport system. UK
GERMANY

6. How manipulate feeds and crossovers?
Kårstø plant – 20 years of development
Europipe II sales gas
Halten/ Nordland rich gas
Tampen rich gas
Statpipe sales gas
Sleipner condensate
Propane N-butane I-butane Naphtha
How manipulate feeds and crossovers?
The diagram provides a schematic presentation of the development of the process plant over two decades.
This plant has been extended several times.
The plant has 3 different feeds and more than 10 crossovers.
How to manipulate on the feeds and crossover to obtain maximum throughput for this plant?
This is more than an operator can do manually.
Especially the extensions done in the latest years have been built with the possibility to distribute gas and liquid to different processing trains, wherever there are processing capacity.
1985 – Statpipe
1993 – Sleipner
2000 – Åsgard (KUP)
2003 – Mikkel (NET1)
2005 – Kristin (KEP2005)
Condensate
1985
1993
2000
2003
2005
Ethane

7. Maximum throughput
Often: Economic optimal operation = maximum throughput
Operate with max feasible flow through bottlenecks
No remaining unconstrained DOFs (RTO not needed)
“Coordinator MPC”:
Manipulate TPMs (feed valves and crossovers) presently used by operators
Throughput determined at plant-wide level (not by one single unit)  coordination required
Frequent changes  dynamic model for optimization
The throughput at the Kårstø plant is presently set by the operators who manipulate
the feed valves to satisfy orders from the gas transport system (operated
by another company). The orders are stated as pipeline pressures, feed rates and
export gas rates, which may change on an hourly basis. The objective of this work
is to coordinate the throughput manipulators (uc) to achieve economic optimal operation.
The overall feed rate (or more generally the throughput) affects all units in the
plant. For this reason, the throughput is usually not used as a degree of freedom for
control of any individual unit, but is instead left as an “unused” degree of freedom
to be set at the plant-wide level.
In this study, a different approach is used. We assume that the main economic
objective is to maximize the plant throughput, subject to achieving feasible operation
(satisfying operational constraints in all units) with the available feeds. This
corresponds to maximum flow through the bottleneck(s) within the operational
constraints. This insight may be used to implement an optimal operation strategy
without the need for dynamic optimization based on a detailed model of the entire
plant (Aske et al., 2008).
One option for solving the maximum throughput control problem for the entire
plant is to combine all the MPCs in the plant into a single application. However,
we choose to decompose the problem by using several local MPC applications and
a coordinator MPC (Aske et al., 2008). The coordinator uses the remaining degrees
of freedom (uc) to maximize the flow through the network subject to given
constraints.
If feed is limited: optimum= max NGL recovery, can be included in the coordinator as well
TPM = Throughput Manipulator

8. Approach ? Objective: Max throughput, subject to feasible operation:
Remaining capacity (R) = Rs = 0 in bottleneck units
Throughput manipulators (TPMs): Feeds and crossovers
Approach: Use Coordinator MPC to optimally adjust TPMs:
Coordinates the network flows to the local MPC applications
Decompose the problem (decentralized).
Assume Local MPCs closed when running Coordinator MPC
Need flow network model (No need for a detailed model of the entire plant)
Decoupling: Treat TPMs as DVs in Local MPCs
Use local MPCs to estimate feasible remaining capacity (R) in each unit

9. ”Coordinator MPC”: Coordinates network flows, not MPCs

10. Remaining capacity (using local MPCs)
Feasible remaining feed capacity for unit k:
Obtained by solving “extra” steady-state LP problem in each local MPC:
subject to already given present state, model equations and constraints
Very little extra effort!
current feed to unit k
max feed to unit k within feasible operation
The remaining capacity for unit k is the difference between the current feed Flk
and the maximum feed Flk,max calculated at the current time.
The feed to the local unit Flk is a DV in the local MPC application and the maximum
feed to the unit k is then easily obtained by solving an additional steady-state LP-problem.
Here ulk is the vector of manipulated variables in the local MPCs,
and at the optimal solution, all these degrees of freedom (uk) are used to satisfy
constraints (feasibility limit)
To calculate the units current maximum feed, the
end predictions (steady-state model) for the variables are used. This assumes that
the closed-loop response time is faster for the local MPC than for the coordinator.
Note that Flk,max can change due to updated measurements, disturbances (i.e.
feed compositions changes), changes in the constraints and model changes (that
is, the steady-state gain in the models) in the local MPCs. The current feed to the
unit (Flk ) is measured, either by a flow transmitter or by a level controller output
(valve opening) if a flow transmitter is not available.
Backoff is needed due to unmeasured disturbances and possible long effective deadtime (at least for some units) in the plant. 10

11. Local MPC applications
Kårstø: Most local MPC applications are on two-product distillation columns:
CVs: Distillate- and bottom products quality (estimated)
+ differential pressure and other constraints
MVs: Temperature setpoint (boilup) and reflux flow
DV (disturbance): Feed flow
New: Local MPCs estimate their feasible remaining capacity (R)
These MVs correspond to the local degrees of freedom (ul) and the CVs correspond
to the local outputs (yl).
Some of the columns have additional CV constraints, like valve opening, temperatures and levels. One column has an additional
MV and some columns have additional DVs, but in principle, all the columns have the same control configuration.
The product qualities are given by mole fraction of the key component and a logarithmic transformation is used to linearize over the operating region (Skogestad,
1997).
The high limits on the product qualities follow from the maximum levels of impurity in the sales specifications and the differential pressure high limit is placed just below the flooding point.
The MPC problem is solved at each sample time using a standard two-step approach, where first a steady-state problem with constraint relaxation until the predicted
final steady state is feasible, and then the “standard” (dynamic) MPC problem is solved with the possible recalculated set points and constraints. The priority order
for solving the steady-state feasibility problem in the local MPC (Strand and Sagli, 2003) is:
High limit differential pressure
2. Impurity limits
3. Impurity set points
This priority hierarchy may lead to a relaxation of the impurity set points (and in worst case the limits) to avoid exceeding the differential pressure high limit. By
using relaxation, the column can handle the given feed rate without flooding the column, but note that the exceeding the limits may result in an unsellable product.
In the dynamic optimization part, constraints are handled by adding penalty terms to the objective function.
The local MPC applications are based on experimental step response models.

12. Coordinator MPC
Objective: Maximize plant throughput, subject to achieving feasible operation
MVs: TPMs (feeds and crossovers that affect several units)
CVs: total plant feed + constraints:
Constraints (R > backoff > 0, etc.) at highest priority level
Objective function: Total plant feed as CV with high, unreachable set point with lower priority
DVs: feed composition changes, disturbance flows
Model: step-response models obtained from
Calculated steady-state gains (from feed composition)
Plant tests (dynamic)
The coordinator uses individual (SISO) step response models, or more precisely a
single-input multiple-output representation of a multi-input multi-output system.
The advantage with SISO models is that it is easy to adjust the models independently
for input-output pairs. However, SISO models imply that the structure of
the process is lost and, for instance, disturbances do not propagate as they would
in a state-space model. The loss of structure leads to some additional work.
The sampling time for the coordinator MPC is 3 minutes. The prediction and
control horizon are set to 6 hours, whereas the longest response models reach
steady state at approximately 4.5 hours

13. 6 MVs 22 CVs 7 DVs KÅRSTØ MPC COORDINATOR IMPLEMENTATION (2008)
Export gas
Rich gas MV CV CV
Export gas MV CV CV
Rich gas CV CV CV CV CV CV MV
Half of the plant included:
6 MVs
22 CVs
7 DVs MV
Condensate CV MV CV CV MV CV CV CV CV CV

14. Step response models in coodinator MPC
Remaining capacity (R) goes down
when feed increases…
NB: Ikkje alle modellar får plass på bildet!
+ more…

15. Coordinator MPC in closed loop
Test runs January to April 2008

16. TEST 07 FEB 2008 Export gas Rich gas MV Export gas MV1 Rich gas MV2 MVCV CV
Export gas
MV1
CV3 CV
Rich gas CV CV
CV2 DV CV
CV1 CV
MV2 MV
Condensate CV MV CV CV MV CV CV CV CV CV 16

17. TEST 07 FEB 2008 MV1 MV2 CV1 CV3 DV CV2 t = 0 min: Turn on
t = min: Change model gains (tuning)
t = 500 min: Adjust back-off for R in demethanizer
t = 580 – 600 min: Feed composition change (DV)

18. Experiences
Using local MPCs to estimate feasible remaining capacity leads to a plant-wide application with “reasonable” size
The estimate remaining capacity relies on
accuracy of the steady-state models
correct and reasonable CV and MV constraints
use of gain scheduling to cope with larger nonlinearities
Crucial to inspect the models and tuning of the local applications in a systematic manner
Requires follow-up work and extensive training of operators and operator managers
“New way of thinking”
New operator handle instead of feedrate: Rs (back-off)
Important prerequisite it that the local MPC applications are installed. Then we are able to operate coloser to constraints. In this case for distillation columns it means that lower internal trafic in column gives ”extra capacity”.

19. Conclusions
Frequent changes in feed composition, pipeline pressures and other disturbances require a dynamic model for optimization
Coordinator MPC is promising tool for implementing maximum throughput at the Kårstø gas plant.
More focus among operator personnel on
capacity of each unit
Plant-wide perspective to decide the plant- and crossover flows
At present, the estimate is based on experimental
modelling. However, rigorous models for local units can also be used
to predict the remaining capacity. This is attractive for units where experimental
modelling is difficult, for example, due to nonlinearities. This illustrates the
flexibility with this decomposition where the best available model can be used to
predict the remaining capacity.
Income? Only estimate available. Processing more gas  oil production not limited by gas handeling capacity.  $
Furhter plan: some challenges at some distillation columns must be solved, in addition to personell resources.
Implement this, further extensions like the most important issue now: CO2 content in the export gas. (not exceed 2.5 mole%)

20. Acknowledgements References StatoilHydro and Gassco
Kjetil Meyer, Roar Sørensen
Operating managers and personnel at the Statpipe and Sleipner trains.
References
Aske, E.M.B., S. Strand and S. Skogestad (2008). Coordinator MPC for maximizing plant throughput. Comput. Chem. Eng. 32(1-2), 195–204.
Full paper: E. Aske, E. Ph.D. thesis, NTNU, Trondheim, Norway, 2009 (Chapter 6). Available from the home page of S. Skogestad:

21. COORDINATOR
IN CLOSED LOOP
DATE=?

22. DATE=? Export gas Rich gas MV Export gas MV Rich gas MV MV CondensateCV CV
Export gas MV CV CV
Rich gas CV CV CV CV CV CV MV MV
Condensate CV MV CV CV MV CV CV CV CV CV 22

23. CV: Pipeline pressure MV: Feed MV: Crossover 6 hrs 9 hrs
DATE=?
CV: Pipeline pressure
MV: Feed
New constraint
from pipeline
network operators
CV: Remaining capacity
MV: Crossover
Increase
backoff
An example from MPC tuning: 7. feb 08 (t=538, internal use only)
Increase back off for remaining capaicty
Can use crossover flow only because of plantwide view, do not need to reduce the feed rate. Constraint tuning
There is a crossover available to route the flow through the plant without reducing the plant feed.
The crossover is increased to reduce the flow to the unit.
2. New guidelines from the gas network operators. The pressure must be increased the end-line of the pipe to operate the pipeline network differently (Tampen link).
The coordinator takes action and reduces the plant feed.
Here, there are no ideal value on the crossover, so it is remained constant.
Shorter prediction period for pipeline pressure: too conservative to reduce plant feed when it predicts for 6 hours. Do not measure or know how flows into the pipelines are.
6 hrs
9 hrs 23

24. COORDINATOR
IN CLOSED LOOP
07 FEB 2008
Plant, from test run nr. 2

25. CV: Pipeline pressure MV: Feed 6 hrs 9 hrs MV: Crossover
07 FEB 2008
CV: Pipeline pressure
MV: Feed
6 hrs
9 hrs
MV: Crossover
CV: Remaining
capacity
Upper right and left: balanced pipeline pressure and plant feed. CV contraint tuning here.
Remaining capacity balanced with crossover. Here the model has too little gain and this leads to an “aggressive” use of the crossover and actually generates oscillations in the downstream remaining capacities because of the delays in the flow network if the model gain was too low.
To avoid this, the model gain was almost doubled crossover is now able to control the remaining capacity towards its low constraint.
3. A large feed composition change enters the plant, first heavier feed (more feed flow to the column), then lighter feed (less feed to column).
The feed composition is estimated from flows and temperature. Its absolute value has no measing, but the relative change that matters.
Crossover starts reducing, then because of the lighter feed, the crossover starts to increase again. See that the remaining capacity reach its low constraint within its horizion.
DV: Feed composition
Composition
disturbance
Model adjustment 25