NARNGULU 18/12/07 9:37 am Page 3

PROCESS DEVELOPMENT

Optimum process scheduling

Expert Optimizer saves more than A$1.2 million annually

at the wet processing section at the Narngulu synthetic rutile
plant, Doug Clark* and Dario Castagnol† explain

O
ptimum scheduling for the wet section of the Becher process at batches with ammonium chloride in 22 sparged agitated tanks, to remove
Iluka, Narngulu was achieved using ABB’s Expert Optimiser. A metallic iron from the mineral grains. The time required for the batch
mixed logical dynamical model of the process is solved in real reaction varies but is generally in the range of 16 to 20 hours. The
time in order to find the optimum operating strategy for a 43 hour time reaction is exothermic and completion is detected by temperature. After
horizon and the strategy applied. The resulting smoother operating gives aeration the resulting slurry is fed to a cycloning system where the iron is
savings in both material and energy leading to a dramatic reduction in removed with the overflow and the mineral is recovered in the underflow.
annual operating costs. Savings of A$1.2 million/y were achieved, giving a This is a semi-continuous process. What happens to the underflow after
project payback of less than six months. the cycloning depends upon whether the product is ‘standard’ or SREP.
Iluka Resources use the Becher process to produce synthetic rutile from SREP heavy metals are leached out using sulphuric acid. In the ‘standard’
ilmenite at its Narngulu site in Western Australia. The process has two case, there is a bypass system which simply pumps the slurry from one
distinct elements, the dry and wet sections. In the dry section ilmenite is tank to another providing some additional buffer capacity. Finally the
mixed with coal and is processed in rotary kilns to reduce the iron oxide in slurry is fed to a filtering and drying system. The kilns’ off gasses provide
the mineral to iron metal. The reaction that takes place is: heat for drying which can be supplemented by a gas burner if required.
Optimising the wet part of this process is highly problematic because
CO + FeO.TiO2 ◊ CO2 + Fe + TiO2 of the need to predict the state of the process in two days time and
produce the most cost effective strategy to give smooth operation. To
Two different products are produced ‘standard’ and SREP (Synthetic compound the difficulty, some of the equipment can be assigned to
Rutile Enhancement Process). In SREP hydroborasite flux is added to the either product line and the optimisation must be able to cope with this
kiln so that heavy metals can be removed from the reduced ilmenite (RI) without requiring re-engineering. ABB was selected to work with Iluka to
in the wet processing. After the kilns any unburnt coal (char), is separated optimise this process due to the pre-existing good relationship between
from the RI in a magnetic separation plant to be returned to the kilns. the two companies and due to the good pedigree of the Expert
In wet separation, iron is separated from the titanium oxide as the RI Optimizer system.
passes through a separation system, and in the case of SREP heavy metals
are also removed. Optimisation tasks
Aeration is the first phase of wet separation. Here the RI is reacted in There were two optimisation tasks. The first was to control the kilns to
improve output and quality, the second was to schedule the operation of
*CPM Solutions, ABB Switzerland Ltd the wet section in real time to meet various objectives. This discussion
† Dario Castagnoli, Corporate Research, ABB Switzerland Ltd covers the latter part of the project.

30 International Mining JANUARY 2008

NARNGULU 19/12/07 10:11 am Page 4

PROCESS DEVELOPMENT

In order to schedule the wet system the optimiser system decides: constraints may be violated but this adds a very high cost to the model
■ When to charge and discharge the aerators so that violation becomes unlikely.
■ When to start and stop the cyclone systems The equations above are matrix ones, covering the whole process time
■ When the leach system should run and, if running, what the setpoint horizon. The user does not see this because the graphical model is
should be automatically converted to a matrix representation by the Expert
■ When the dryer systems should run and, if running, what the setpoints Optimizer system. These matrices end up being extremely large, making
should be. the solving of the model in real time difficult. The level of difficulty is
Expert Optimizer provides recommendations to the operators to charge then massively increased by including Boolean variables (example,
or discharge aerators; automatically starts and stops the cyclone systems; run/stop) and for this reason we believe this is probably the most complex
changes the setpoints for the leaching and drier systems and can real-time MLD system ever deployed on an industrial scale.
recommend stops/starts of these systems as required. It is anticipated In no particular order, the constraints considered in the model include:
that in the future the charging and discharging of the aerators will be ■ Keeping the initial feed bin at a low level (improves quality, reduces
automated. aeration time & prevents kiln stops in case of wet section breakdowns)
■ Not charging more than one aerator on each line at one time (physical
The model constraint)
The chosen solution method was to model the process as a Mixed Logical ■ Not discharging more than one aerator on each line at one time
Dynamical (MLD) system using the ABB Expert Optimiser toolkit and then (physical constraint)
to apply Model Predictive Control (MPC). Expert Optimiser provides a ■ Not having mature aerators that cannot be discharged due to lack of
unique graphical tool for programming MPC models that brings the tank space (improves quality and reduces power usage)
techniques into a factory safe environment. ■ Keeping the level of the tanks within target ranges and ensuring no
The general method of working is: overflows
■ The process is decomposed into control volumes (in this case, RI bin, ■ Not stopping the SREP leach system unless absolutely necessary but if it
aerators, cycloning, leaching & drying) must be stopped then not starting it until a minimum time period has
■ Each control volume is modelled independently using a combination of passed (improves quality and reduces acid consumption)
blocks from a palette and, if required, imported code written in the ■ A minimum consecutive up time for the cyclone systems (operator
Hysdel language preference)
■ Control volume models are linked together graphically to produce a ■ Running the dryers at steady output (maximises use of waste heat)
model that closely matches the process flowsheet. The model for ■ Allowing equipment to scheduled as not available (for planned
‘standard’ is shown in Figure 1. maintenance).
■ The MLD blocks are parameterised using drag and drop An example of an aerator schedule is shown in Figure 2.
■ The objectives are each given a cost that reflects their priority
The model implemented in this case predicts the state of the process Project procedure
for the next 43 hours and calculates the best operating strategy over this The project proceeded in several steps. First, there were two site visits
period by minimising the cost of operation over the time horizon. In the where ABB and Iluka engineers cooperated to understand the mass
same way that an expert chess player tries to foresee the development of balance of the process and the rules on how it should be operated.
the game and then makes the initial move that is most likely to lead to ABB developed an initial model and tested using data collected during
victory, the MPC calculates a time series of setpoints likely to lead to the the visits. A high level of communication between the engineers was
lowest cost outcome and then the first step of the setpoint vectors are essential to ensure that the model closely matched the reality. At the end
implemented. The process then evolves and the model is recalculated so of this step the initial model was used to train Iluka engineers on the use
that if the expectation does not match the reality a new winning strategy
is applied in real time. Figure 1: MLD model for "Standard" production
The form of the model in MLD is given by,
LINE-A MLD System ILUKA_STANDARD

RI Bin Aerators

y $
ri_overflow_cost
y $
y u a_mature_cost
ri_bin_level u_st_dis
u
Where: u_st_cha
¬ x represents the state of the process, for example a ri_bin
ce_delay
discharge_control distime_delay1b

tank level
¬ u are the inputs to the model some of which can be u y $ u
y $
u_al_outflow al_stop_cost u_dr_outflow
drying_cost
controlled (changed) by the model and some of which
are uncontrolled. An example of a u would be a u
u_cyclo_on wet_separation y
dryer feedrate setpoint acid_leaching drft_level

¬ y are variables used for control decisions, for example y

drying
y alft_level y $
costs cyft_level dr_stop_cost
¬ E represent the constraints on the process. These
Cycloning Acid Leach Drier
constraints may be ‘hard’ or ‘soft’. Hard constraints
must be respected or the model will fail; soft

32 International Mining JANUARY 2008

NARNGULU 18/12/07 9:38 am Page 5

PROCESS DEVELOPMENT

of the Expert Optimizer system facilitating their participation in the Figure 2: Predicted Aerator schedule for 24 hours
commissioning.
Working in co-operation with the process control department at
Narngulu, two ABB engineers commissioned the strategy over a five week intangible benefit of freeing the operators to take an overview of the
period. The system was then left running for about three months. During process instead of always concentrating on the nitty-gritty. Operator
this time the site engineers observed the system, made modifications Murray said “I never turn it off, it does a great job. It makes life here [in
when required and determined where the strategy should be refined. The the control room] so much easier”. Operator Ian said “I was always
ABB engineers returned to site for a two week visit to refine the strategy, trying to see what would happen tomorrow, it does it so much better
ironing out any bugs detected since the first visit. than I could”. IM
Operating the system following the rules above provides several
benefits:
■ Keeping the RI bins at low levels ensures that
the aerators are fed with fresh, hot material. This
speeds up reaction time and gives a better quality
product
■ Discharging aerators as soon as they are mature
reduces power consumption and improves quality
by not over processing the RI
■ More continuous running of the SREP leach
system decreases the acid and lime consumption,
and improves quality
■ Running the dryers at a steady rate to maximise
the use of waste heat for drying and minimises
the use of gas
■ Preparing the system for maintenance so that a
queue of undischarged aerators does not develop
whilst downstream equipment is unavailable.
After six months of operating the results are as
show in the table:
The annual total savings are A$550,000 in
materials and A$696,000 pa in power, giving a Aerators at Iluka. Expert Optimizer plans their operation
project payback of less than six months. As well as in order to maximise throughput and reduce cost
these obvious tangible benefits the system has the

34 International Mining JANUARY 2008