Consider rapid new controller developments via

Originally appeared in:
October 2014, pgs 53-56.
Used with permission.
Special Report
Process Control and Information Systems
M. CALOSI and S. LODOLO, Aspen Technology, Italy;
and A. ESPOSITO and A. AUTUORI, Eni, Italy
Consider rapid new controller developments
via adaptive process control
A new set of innovative technologies are having a significant
impact on advanced process control (APC) methodology,
applications development and sustainability.a A traditional
application pushes for full linear programming (LP) optimization
that exploits all available degrees of freedom. A controller with such
objectives may seek a new solution that provides only incremental
benefit over the existing solution and the movement between
nearly identical solutions may be exacerbated by model mismatch.
Therefore, care is taken to identify quality models, and
model predictive controllers do not tend to return benefits
until late in the project when the application nears final
commissioning. This often results in many months going by
before the application is fully online and generating benefits.
The latest APC technologies, known as adaptive process
control, solve a number of problems.a They maintain the LP
steady-state target in an “optimum area” instead of an optimum
point. Optimization and stable behavior is achieved even in the
presence of a very large model mismatch.
APC can also utilize what are called “seed” models that are
built from pre-test and/or purely historical process data or even
from a rigorous plant model. This, coupled with the robust optimization behavior in the presence of model mismatch, means that
the controller can be deployed with these less-refined seed models and, in a much shorter period of time, begin accruing benefits
while the model is improved online by the adaptive technology.
Once deployed, the acquisition of process data to refine the
seed model is done using small perturbation background step testing while the controller is online. The underlying enhancements to
the model identification algorithm enable the effective use of significant amounts of closed-loop data that can be collected while a
user-specified level of emphasis on optimization is carried out by
the system. This approach puts in the hands of the users the ability
to make trade-offs in the speed of developing updated models vs.
the degree of loss of benefits while the model is being updated.
The collective set of innovations within adaptive process control is driving a radical change in the economics of APC applications. While it has been shown that the technology solves many
of the maintenance issues with traditional APC solutions, it is
also having a dramatic impact on new APC applications building.
In this case study, the implementation team discusses how
Eni’s Livorno refinery in Italy was able to use adaptive process
control to start accruing benefits from two hydrotreater units
just six weeks from the start of the project.
Testing. Several testing tools are available in the marketplace.
All of them deliver good plant data for model identification
purposes, generating ad-hoc signals for model inputs (i.e. manipulated variables) such as pseudo-random binary sequence
(PRBS) or similar. Some of them also ensure controlled variable (CV) limit violation protection, thanks to a raw unit model,
thus reducing the need for continuous monitoring by engineers.
However, none of these tools avoid the need to heavily test
the plant, which makes the process of generating accurate data
quite expensive.
The current generation of automated testing gives a binary
choice: operate the plant under the supervision of an advanced
process controller that optimizes it, or collect data in open-loop
giving up the benefits during the test. This conflict of interest promoted finding a solution that is able to generate data suitable for
modeling while continuing to optimize the operation of the plant.
But, if a controller (model) is needed to achieve some level of
process optimization during the test, this implies the need to safely control the process unit with an inaccurate model. When the
model mismatch is large, the steady-state targets calculated by the
APC controller may start jumping around, and weak process handles may suddenly be used, leading to large movements in these
handles. This situation occurs because the controller exploits all
available degrees of freedom, which may generate large manipulated variable (MV) movements, even if these movements bring
insignificant incremental changes in the cost function. This indicates that, with a low-quality model, pushing for full LP optimization every cycle may not always produce the best results.
APC controllers, in the presence of a very large model mismatch, behave in a much more stable way when all MVs are set
to minimum moves instead of minimum costs. In fact, most
instabilities are due to the optimization algorithm of an APC
application and not to the controller engine, which typically exhibits robust behavior even in the presence of a strong model
mismatch. However, with such configuration, the controller
stops optimizing and only cares about constraint protection.
The solution is a middle ground between running the controller with minimum cost setting and running it with no optimization, allowing for a trade-off between optimization and
robustness. The new adaptive process control technology has
the ability to control the process by making step-like changes
to the MVs and stabilizing the process even if significant model
mismatch is present, while preventing the LP targets from flipping around. It is intended to work as a robust controller where
minimal MV movement is made while active process constraints
(active CVs) are kept at, or close to, limits. In this way, the controller maintains process stability while at the same time performing
very small perturbation tests generating closed-loop data. The
HYDROCARBON PROCESSING OCTOBER 2014
Process Control and Information Systems
process unit can be operated close to the point desired by the operations staff, even if the LP costs are not entirely correct.1
Such an application is the perfect tool for existing advanced
controller model maintenance, where an initial controller (model)
is already in place, even if not performing well. However, the
ability to robustly control the plant, even with a highly inaccurate
model, gives the possibility to build from scratch and deploy a
new application in a very short time, thereby radically reducing
the time-to-benefit for new advanced control applications.
The new adaptive process control technology has been used
for the first time to build a new APC application in Eni’s Livorno
refinery hydrotreater unit, obtaining good results and proving the
solution to be suitable for this purpose.
The unit. The hydrotreaters’ naphtha feed, coming from the crude
distillation unit (CDU), has a sulfur (S) and nitrogen (N) content
Feed
(gasoline)
FC
F1
R1
TC
FC
Fuel
gas
Fresh H2
PC
PC
FG net
LC
H2 net
PC
V2
FC
V1
LC
C1
V101
LC
TC
LC
FC
F2
FC
Fuel
gas
To
reprocessing
tank
FC
To
reforming unit
FIG. 1. Schematic of Eni’s hydrotreater in Livorno, Italy.
322
320
A
318
WABT ref, °C
EOR)
nths (
6 mo
GE = 9
hs
2 mont
AGE = 7
Achievable benefits. The main economic benefits achievable
with an advanced process controller in the hydrotreaters
include fuel-gas consumption reduction in reaction furnaces
and feed re-processing reduction. Other minor benefits include:
minimized stripper reboiler duty up to product specifications,
longer catalyst life due to reaction pressure maximization and
hydrogen-to-hydrocarbon ratio maximization.
It was normal practice to maintain the same reactor inlet
temperature regardless of the feedrate. This is, by experience, a
temperature that guarantees the removal of S and N at the highest
feedrate. This means that the severity is allowed to change as the
feedrate changes. Minimizing the reaction preheater coil outlet
temperature (COT) up to a desired reaction severity, instead
of maintaining a constant reactor temperature, saves fuel gas
burned in the reaction furnaces (FIG. 2).
Another normal practice was to maintain the hydrotreater
feed higher than the continuous catalytic reforming feed so that
the V101 drum level was always controlled by the PID controller
that recycles the overflow back to feed for reprocessing (FIG. 3).
Regulating the hydrotreaters’ feedrate to keep the continuous
catalytic reforming feed drum level constant can reduce the
recycle to zero, saving the fuel gas needed for reprocessing.
hs
8 mont
AGE = 4
316
that is too high for processing in the continuous catalytic reforming
unit. It must be treated to remove pollutants. This is achieved by
sending the feed to two hydrotreater units configured in parallel. A
simplified process flow diagram of the described setup is shown in
FIG. 1. The units are primarily composed of a reactor (R1), a highpressure separator (V1) and a stripper column (C1).
The fresh naphtha feed is combined with hydrogen in a single
stream, preheated by reactor effluent, and then sent to a furnace
(F1) to achieve the desired reactor inlet temperature. The reactor
effluent is sent to a separator (V1) from which unreacted hydrogen
is recovered and sent to the hydrogen network. The liquid phase
from V1 is sent to a total reflux stripper (C1) to remove the
hydrogen sulfide. Heat is provided to the stripper by a furnace (F2)
that can be adjusted to achieve the desired Reid vapor pressure
(Rvp) for the naphtha exiting the column from the bottom. The gas
phase drawn from the overhead drum is sent to a sulfur-recovery
unit under pressure control. The bottoms of the stripper in the two
hydrotreater units go to the V101 drum, which is the feed drum for
the continuous catalytic reforming unit. Since continuous catalytic
reforming feed flowrate cannot be used to control the level (as it
must be as constant as possible), the level of this drum is controlled
by a proportional-integral-derivative (PID) level controller that
sends feed surplus to a reprocessing tank.
hs
4 mont
AGE = 2
AGE = 0
314
s (SOR)
month
312
310
45
65
Feed, m3/h
85
FIG. 2. Reference for reaction severity calculation.
105
FIG. 3. V101 level control via feed recycling.
HYDROCARBON PROCESSING OCTOBER 2014
Process Control and Information Systems
Project timeline. A standard project has
Pretest
Test
Modeling
Commissioning
Monitoring
Troubleshooting
several phases needed to build, configure
• PID loop tuning • Supervised step • Data slicing
• Open and closed • Plant/model • Correct mismatch
and deploy an advanced process control- • Instrumentation testing
• Add constrains
• Identification loop testing
mismatch
• Fine tuning
review
• Data collecting
• Tuning
• Additional
ler. These include:
• Design
constraints
• Pretest, where PID loops are re3-4 months
tuned, instrumentation is reviewed
Start getting benefits
and a preliminary design of the
application is drawn
Pretest
Test (adaptive mode) and monitoring
Modeling
Comm.
• Plant test, where the plant test of
• PID loop tuning
• Automated step testing
• Fine
• Open- and
the unit is performed to generate
• Instrumentation
• Automated data collecting
identification closed-loop
review
• Automated data slicing
testing
quality data
• Design
• Automated modeling
• Fine-tuning
• Modeling, where the model of the
• Historical data gathering
• Process optimization with a robust controller
• Initial strategy and
unit is identified
“seed” model
• Commissioning, where the
3-4 weeks
Start getting benefits
controller is deployed online and
fine-tuned.
FIG. 4. APC project time line with standard and adaptive approaches.
The traditional APC application is
able to bring benefits only at the last
MV2
phase of the project. Moreover, the plant test phase is usually
Joptimal
expensive, since the plant is driven far from its optimum point,
and products off-spec can occur.
J
In contrast, the new adaptive process control technologies,
Jcurrent
Optimum point
which are able to control and optimize a process unit in a robust
manner even with a very inaccurate model, allow deployment
of an APC application in the early stages of a project.a Plus, the
Optimum area
CV1 limit
adaptive application will optimize the process unit while step
J CR
testing it in a nondisruptive way.
Actual point
FIG. 4 compares the schedules of the two project approaches,
and highlights the distinct difference in time to economic benefit. The Eni Livorno hydrotreaters project has been executed using the new technology that allowed for a different project path:
MV1
• Week 1: Pretest
FIG. 5. Adaptive control mode optimization.
o Instrumentation assessment
o Application design
o Historical data gathering
MV1 and MV2 represent two particular MVs. The technology
• Weeks 2–5: Preparation for adaptive process controller
ensures that the two CVs will be kept inside of the shaded triano Identification of the “seed” model from historical data
gle. In effect, an additional cost constraint is imposed so that the
o Instrumentation maintenance
cost function will not degrade more than the user-specified toler• Weeks 6–7: Adaptive process controller deployment
ance. When inside the shaded triangle, the engine will generate
o Application supervising
step moves for the requested MVs. This operation mode, where
o Operators training
the unit is tested and optimized is called adaptive control mode.
• Months 3–6: Background step testing
During the four months in which the application was run in
(adaptive control mode)
adaptive control mode, the unit was optimized and tested in a
o Four months of plant optimization
nondisruptive way implementing small steps on each MV. Given
• Fine modeling and final commissioning (one week).
the small step size (even 10 times smaller than usual step test
After only six weeks from the beginning of the project, the
moves), operators hardly noticed that a step test was ongoing;
APC application was online and running, stepping the unit and
they just saw the benefit of an MPC controller in action.
bringing benefits to the refinery.
An example of the application behavior is reported in FIG. 6:
1. The controller brings the unit near its optimum point,
represented by the minimum limit of the reaction severity
Background step testing. When in adaptive control mode,
2. Once the unit is close to the optimum point,
the optimization algorithm recalculates the steady state (LP)
the MVs are stepped
objective function at every time step and stores this cost func3. If the unit moves away, it is driven back to the optimum
tion, Joptimal. It then calculates the cost function at the current
area by the application before continuing the plant test.
MV targets, Jcurrent , with the minimum move setting (constraint
control), compares the two cost function values, and calculates
ΔJ = Joptimal – Jcurrent . If ΔJ is larger than the user-specified tolerModel improvement. Model accuracy improvement obtained
ance, or calibration ratio, CR, the engine will generate a move
with the adaptive process controller is shown in FIG. 8. The
plan and push the process toward the new optimal solution.
adaptive process controller was able to control and optimize the
In FIG. 5, CV1 and CV2 are the two active CV constraints, and
unit in a robust way, using a seed model that had up to 50% of
HYDROCARBON PROCESSING OCTOBER 2014
Process Control and Information Systems
Before adaptive mode
Adaptive mode
4
FIG. 6. Adaptive control mode behavior.
6
8
10
12
14
Feed recycle line valve opening, %
16
18
20
FIG. 8. Probability distribution functions of feed recycle line valve
opening before and during adaptive control mode.
Before adaptive mode
gain error in several curves and 30% of missing curves.
hydrotreater unit has a fuel gas consumption of 5,000 Nm3/h,
economic benefits would result in about $600,000 saved in one
year (considering a typical fuel gas price in Western Europe of
0.5$/Nm3).
At the end of the project, with the final advanced process
controller online, economic benefits achieved by the final
application were measured to be 16% of fuel gas consumption
savings. This means that in only six weeks of work, and with
an inaccurate seed model, the adaptive process controller was
able to achieve 20% of the overall achievable economic benefits,
while producing good data for the finalization of the model.
Most of the six weeks’ time was used to fix instrumentation
issues. Without such problems, benefits could start accruing as
soon as two weeks from the project kick off and increase week
after week.
Obtained benefits. During the four months in which the
a
Adaptive mode
315
316
317
318
319
Reaction furnace COT, °C
320
321
322
FIG. 7. Probability distribution functions for the reaction furnace COT
before and during adaptive control mode.
adaptive process controller was stepping the unit to generate
more accurate models, the plant was optimized and major
benefits sources were exploited. Reactor furnace COT had
been minimized, and, according to feedrate, a proper reaction
severity was guaranteed. FIG. 7 presents probability distribution
functions for the furnace COT during the adaptive control
mode period and immediately before.
A similar plot is shown in FIG. 8, where the feed recycle line
valve opening distribution is reported. The adaptive process
controller maintained the catalytic reformer drum level stable
and below the overflow discharging threshold, so as to avoid reprocessing of the feed almost completely.
Fuel gas savings were seen in the early project stages,
measured as 3.2% of the total fuel gas consumption for the
two hydrotreater units. Since the hydrotreaters at Livorno are
considered to be relatively small, if the 3.2% of fuel gas savings
is transferred to a hypothetical hydrotreater unit in a mediumsized refinery, some great results could be seen. If an average
NOTE
Aspen Technology
LITERATURE CITED
Harmse, M., Q. Zheng and R. Golightly, “An Enhanced Iterative Process for
Maintaining APC Applications—Adaptive Process Control,” Aspen Technology.
2
Lodolo, S., M. Harmse and A. Esposito, “Use adaptive modeling to revamp and
maintain controllers,” Hydrocarbon Processing, October 2012.
1
MIRCO CALOSI is a principal engineer at Aspen Technology in Italy. He has worked
on several refining and petrochemical process units developing advanced process
control applications, automation solutions and quality inferentials.
STEFANO LODOLO is a senior advisor and advisory business consultant with
Aspen Technology in Italy. He works with customers all around Europe in
advanced process control and energy management areas.
AUGUSTO AUTUORI is responsible for APC project coordination at Eni refineries.
He works on oil movement systems implementation, innovative systems
implementation, plant monitoring and operator training.
ANDREA ESPOSITO is a senior advanced process control engineer at the Eni
Livorno refinery in Italy. He is in charge of project development and application
maintenance in the advanced process control area.
Eprinted and posted with permission to Aspen Technology from Hydrocarbon Processing
October 2014 © 2015 Gulf Publishing Company
www.aspentech.com