Data Reconciliation for Flowsheets Involving Plug-flow Units.

An Application to an Oil Sands Processing Plant

L. Lachance
Algosys Inc., Québec, Canada
Email: [email protected]

D. Hodouin
Department of Mining, Metallurgical and Materials Engineering, LOOP (Laboratoire d’observation et d’optimisation des
procédés), Université Laval, Québec, Canada

A. Desbiens, E. Poulin
Department of Electrical and Computer Engineering, LOOP (Laboratoire d’observation et d’optimisation des procédés),
Université Laval, Québec, Canada

ABSTRACT
The objective of the paper is to develop a material balance calculation procedure for a production
unit involving two plants separated by a pipeline. The case study is taken from the oil sands industry.
The raw sands are first mixed with water in a conditioning plant, then transported through a pipeline
to an oil extraction plant. The combination of pipeline dimensions and nominal throughput creates a
significant variable time lag between the two plants. As usual, mass balance calculations required
for the evaluation of plant performance indicators, such as bitumen recovery, are inaccurate due to
measurement errors. In addition, in the present case study, the time lag of the pipe creates an addi-
tional source of uncertainty, which may lead to misinterpretations and wrong decisions.
Raw data are usually reconciled through steady-state mass balancing, to correct for measurement
errors. However, because of the significant pipeline dynamic lag, standard steady-state daily mass
balance may fail at giving an appropriate picture of the operations. The proposed solution to the data
reconciliation problem includes the use of a pipeline plug-flow model in the mass balance calcula-
tion. To fulfil the industrial constraint that current data reconciliation must not modify past recon-
ciled values, while still satisfying dynamic constraints, a daily one-step method, i.e. a unitary filter-
ing horizon, is recommended. The method is illustrated with simulation data, and the results com-
pared to those of standard data reconciliation methods.

Keywords: Data reconciliation; Mass balance; Oil sands

considered, the first step toward its successful

INTRODUCTION
implementation is getting a clear and reliable
The extraction of crude oil from oil sands is picture of the actual operations. Unfortunately,
an energy-intensive activity. In view of the high raw measurements provide a fuzzy picture,
processed tonnage, growing energy costs and since they are corrupted by measurement errors
environmental concerns, any optimization pro- and incomplete by the fact that there are often
ject that could reduce energy expenses, while more states to be estimated than direct mea-
maintaining production throughput objectives surements. Furthermore, their reliability is
and quality requirements, would be highly at- problematic since they do not satisfy a priori
tractive. No matter the optimization technique knowledge such as mass and energy conserva-

2322
tion laws.
Data reconciliation through mass balancing
is now a current practice in many industries
(Narasimhan and Jordache, 2000; Romagnoli
and Sanchez, 2000), including in the mineral
processing and metallurgical industries (Bellec
et al., 2007), for improving operation know-
ledge and performance. The redundancy result-
ing from raw measurements combined to mass
错误!
conservation constraints allows the calculation
of reconciled estimates that are more reliable Fig. 1 Simplified representation of the plant flowsheet
than raw measurements, and obey mass and
energy conservation laws. Standard data recon- PROCESS DESCRIPTION AND
ciliation engines provide good results as long as SIMULATION
the underlying hypothesis of steady-state opera- Overview of the operations
tion is satisfied by the process, or timeaveraging
applied to raw data for cancelling hold-up vari- The conditioning plant in Fig. 1 receives ore
ations. Applications of this technique in a feed coming by trucks from the mine (stream 1).
noiseless framework (Vaclavek, 1974) and the It is a mixture of bitumen, water, and solids that
conditions of validity in the presence of mea- contain sands and rocks. The ore feed goes
surement noise (Almasy, 1990; Lachance et al., through crushers in order to break large rocks
2006; Lachance et al., 2007) are available in the into smaller ones, which are subsequently re-
literature. moved (stream 3) using a vibrating sieve.
This paper considers a particular industrial Water is then added (stream 2) to create an ap-
mass balance problem for the plant flowsheet propriate slurry for the hydrotransport through
shown in Fig. 1. It has been simplified, since the the pipeline (stream 4). The extraction plant,
actual process has series or parallel units and which receives the slurry (stream 5), uses flota-
streams which, for this paper, have been lumped tion units to remove solids (water addition
into a single node. The main difficulty is the through stream 6), thereby creating rejects
presence of a pipeline between the two plants. (stream 7). The final product (stream 8) is a
Since the pipeline has a significant and timeva- slurry with low content in solids that goes to
rying residence time relative to the mass bal- another plant, not considered here, where naph-
ance period, the usual assumptions for steady- tha is then used to remove remaining solids and
state data reconciliation do not hold on a global water.
basis. A dynamic model, involving a plug-flow Process simulator
reactor assumption for the pipeline behaviour, is
used to take hold-up variations into account. The data used below to analyse and solve the
The dynamic data reconciliation (DDR) prob- mass balance problem associated to Fig. 1 does
lem (Hodouin, Lachance and Desbiens, 2007) is not come from real plant measurements. Instead,
usually solved either on a total time horizon by a simulator is used to generate random varia-
extended Kalman filters (EKF) (Simon, 2006) tions around nominal values that are selected to
or on a fixed moving horizon (Liebman, Edgar be close to typical industrial data (Table 1). This
and Lasdon, 1992). In this paper, the DDR methodology allows the evaluation of the re-
problem formulation is modified by using a conciliation procedure accuracy, since “real”
unitary horizon (one-step method) in order to values are known. Both the conditioning and the
remove, as requested for industrial applications, extraction plants are simulated using separation
the need to update past reconciled data of the coefficients only. For example (see Table 1),
previous days. The DDR method is illustrated 1.1% of the bitumen that comes in the condi-
with simulated data and compared to steady-st- tioning plant from the ore feed (stream 1) goes
ate methods. out to the rejects (stream 2), and the remaining

2323
98.9 % goes through the pipeline (stream 4).
Table 1 Nominal values of daily cumulated masses and compositions for streams of Fig. 1 plant
Stream
Daily cumulated Mass fraction
mass
Bitumen Water Solids
j F j (t)
mj1 mj2 mj3
1 60 000 0.11 0.05 0.84
2 24 000 0.00 1.00 0.00
3 2377 0.03 0.04 0.93
4 81 623 0.08 0.33 0.59
5 81 623 0.08 0.33 0.59
6 15 000 0.00 1.00 0.00
7 85 624 0.004 0.450 0.546
8 11 000 0.56 0.31 0.13

The conditioning and extraction plant simu- pipeline is assumed to behave as a plug-flow
lators provide average values on a daily basis. reactor completely filled with incompressible
The underlying assumption is that the plants slurry of constant volume V p . This allows writ-
behave as if they were in either steady-state or ing the following volume conservation equa-
stationary regime, where during a one-day pe- tion:
riod the process variables exhibit fast centred V 4 (k)=V 5 (k)=V(k) (1)
variations around constant means. As a conse-
quence, material inventory variations can be where V j (k) is the daily cumulated volume
neglected. Mean changes are thus supposed to through stream j for the day index k (number of
occur only from day to day. This is not true for sampling periods of one day). Since the resi-
the pipeline since it exhibits slow dynamics and dence time in the pipeline never exceeds one
thus a time-varying residence time as well as a day, the cumulated volume of any day is always
change of hold-up. The pipeline process dy- high enough to empty the hold-up that has been
namics must be taken into account since its res- built during the previous day. Fig. 2 gives a
idence time is much longer than those of the visual support to the dynamic mass conserva-
plants, although being smaller than the sampling tion constraints around the pipeline that are now
period of one day (the nominal residence time described.
of the pipeline is 0.47 day). The daily cumulated mass at the pipeline
The day to day change of process variables output, F 5 (k), is obtained by writing the total
means are generated by random disturbances of mass conservation constraint around the pipe-
the conditioning plant feed, around its nominal line
values. These disturbances are simulated as ad- ) F4 (k ) + V p [ ρ 4 (k − 1) − ρ 4 (k )]
F5 (k= (2)
ditive independent zero-mean Gaussian white
noises of relative standard deviations 5%, 9% where ρ 4 (k) is the stream 4 slurry density for
and 10%, respectively for mass flowrate, bitu- day k. It can be calculated from the mass frac-
men mass fraction, and water content. The so- tion and densities ρ i (i = 1 to 3) of bitumen
lids mass fraction is then adjusted so that the (1.01), water (1.00) and solids (2.65) by the
mass fractions sum to one. following expression (valid for any stream j)
Variations of the pipeline feed (stream 4) are 3 mij (k )
induced by the disturbance of the mass flowrate ρ j (k ) = 1/ ∑ (3)
and composition of the ore feed around their i =1 ρi
nominal values. The simulation of the pipeline
In Equation (3) mij (k ) is the mass fraction of
consists in obtaining the daily mass flowrate
and composition of stream 5 knowing daily species i in stream j. The species i (i = 1 to 3)
mass flowrate and composition of stream 4. The mass conservation around the pipeline leads to

2324
the expression of the composition of stream 5 stream 5 have the same nominal values, but the
(see Fig. 2) disturbances applied on stream 1 combined to
1
(V (k ) − V p ) ρ 4 (k )m4i (k )
the pipeline dynamic model of Equations (2)
m5i (k ) =
V ( k ) ρ5 ( k )  and (4) generate differences between the flo-
(4) wrates and compositions of the pipeline input
+V p ρ 4 (k − 1)m4i (k − 1) 
and output.
As can be seen from Table 1, stream 4 and

Fig. 2 Scheme of the dynamic mass conservation around the pipeline (V and F : daily cumulated volume and mass; m : mass
fraction; V p : pipeline volume)

Simulated measurements DATA RECONCILIATION METHODS

The above plant simulator provides “real’’ Data reconciliation techniques are needed for
values for variables to which zero-mean Gaus- correcting raw data and thus obtaining coherent
sian white measurement noise is added to obtain mass balance calculations and reliable perfor-
“measured’’ values. It is assumed that the mass mance indicators, while using all the available
vector F(k) is available (in practice a volume measurements. These methods are commonly
flowmeter and a slurry density gauge would be used for steady-state data, but data reconcilia-
required to perform F measurements). It is also tion problems involving process dyna- mics,
assumed that each stream is analyzed for its unavoidably occurring in industrial practice,
mass composition and the resulting 24 mea- such as the plug-flow unit of the present case
surements at time k are included in the vector study, is not fully covered by the standard lite-
m(k). The measurement equations are then the rature in data reconciliation for process indus-
following: tries (Narasimhan and Jordache, 2000; Romag-
noli and Sanchez, 2000).
F=(k ) F (k ) + eF (k ) wh ree eF ~ N (0Σ
, F ) (5) There are various ways of solving the pro-
m=(k ) m (k ) + em (k ) wh re
e em ~ N (0Σ, m ) (6 ) duction accounting problem for the plant flow-
sheet illustrated in Fig. 1. Computer packages
where the tilde over a variable stands for its
for steady-state data reconciliation are available
measured value. The measurement errors e are in many plants. For that reason, a steady-state
zero mean normal variables of variance matric- approach, illustrated in top of Fig. 3, is to sepa-
es Σ. They are supposed to be white and uncor- rately reconcile daily mass balances of the two
related. A relative random measurement error plants, thus neglecting the presence of the pipe-
standard deviation of 5 % has been considered line. A second steady-state approach, illustrated
on each of the F j (k) and m ij (k ) variables with in the bottom part of Fig. 3, is to consider both
exception for the composition m2i (k ) and plants simultaneously thus neglecting hold-up
variations in the pipeline, considered as a
m6i (k ), i=1...3 which are here considered as steady-state node. Finally, a third approach ex-
perfectly known, because the associated streams plicitly takes account of the pipeline dynamics.
are pure water. These three data reconciliation methods are

2325
abbreviated in the followings by DRM 1, DRM tions associated to (9) express that the flowrate
2 and DRM 3. F of a stream is the sum of the species flowrates
W, since the slurry contains three and only three
2 6
components. The three other equations corres-
pond respectively to mass conservation for bi-
1 4 5 8 tumen, water and solids around the plant node.
3 7 The reconciled states xˆ (k ) are obtained by
Conditioning plant Extraction plant
minimizing, at each time k, a quadratic criterion
2 6
containing the weighted squared values and
cross products of the residuals (differences be-
1 4 5 8 tween measurements and estimated states),
3 7 while verifying the mass conservation of Equa-
Conditioning plant Extraction plant tion 9. The reconciliation problem is thus for-
mulated as:
Fig. 3 Two possible configurations for steady-state mass
arg min ( y (k ) − x (k ) ) Σ −1 ( y (k ) − x (k ) ) (10)
xˆ (k ) =
T
balancing
x (k )

Separate steady-state data reconciliation for s.t E x (k ) = 0

each plant (DRM 1) The reconciled states are analytically obtained
Since both plant configurations are identical, as:
as shown in Fig. 3, only the solution for the
conditioning plant is described below, knowing


( 

)
 I − ΣE T E ΣE T −1 E  y (k ) (11)
xˆ (k ) =

that the one for the extraction plant is similar. A This solution is then separately applied to
simple optimal linear solution can be considered each plant.
for this type of problem by introducing compo-
Simultaneous steady-state data reconciliation
nent mass flowrates variables W ji = F j m ij . A
of plants (DRM 2)
13×1 state vector x is then defined at each
When neglecting the hold-up variations oc-
time k
curring in the pipeline, this second approach
x (k )T = [ F1 (k ) F4 (k )
(7) (bottom of Fig. 3) results in a steady-state mass
W11 (k )W ji (k )W43 (k )  balance for a network of three nodes and eight
where the three component mass flowrates streams. Following the previously introduced
for water stream 2, W2i , i = 1 3 , have been method for each plant (DRM 1), the state and
measurement vectors x and y are now 26×1
omitted since their mass fractions m2i (k ) are vectors obtained by concatenating those of the
exactly known. The state vector (7) is not di- two plants. The mass conservation constraints
rectly accessible in practice, but only through are the same 12 equations (six for each plant) as
the measurement vector in DRM 1 augmented with three equations ex-
y=(k ) x (k ) + e (k ) (8) pressing that species flows in stream 4 and
where e(k) is a Gaussian zero-mean white stream 5 are equal. With these new vectors x
noise with an associated 13×13 variance matrix and y and the 15×26 full rank matrix E, a data
Σ. The matrix Σ is not diagonal, although not reconciliation problem similar to (10) can be
full, since there is correlation between the mea- solved as (11).
surement errors on any F j and its associated Dynamic data reconciliation (DRM 3)
W ji for j = 1 to 4 and i = 1 to 3.
An optimal DDR method at time k should
The basic requirement for data reconciliation update the past reconciled values at times k−1,
is that the estimates satisfy the mass conserva- k−2 and so on, since the dynamic model equa-
tion constraints, written in matrix form as tions link the present state variables to the pre-
Ex = 0 (9) vious ones. This would mean that the material
where the 6×13 full rank matrix E contains balance of the previous days would be modified
coefficients 1, −1 and 0. The first three equa- by the reconciliation with the new measured

2326
values of the present day. This might not be since only one variable (V(k)) will be implicated
acceptable from the viewpoint of industrial ac- into the NLP method. First, a value of V(k) is
counting. A preferred solution is to perform a selected, then the minimization with respect to
sub-optimal reconciliation, i.e. a one-step DDR F and W is performed, subject to the linear con-
method that assumes that, at time k, the past straints lumped into E(k)x(k)=A(k) xˆ (k − 1)
values are kept at their previously reconciled where the matrices E and A are now time vary-
values. This strategy is applied to the dynamic ing since they contains respectively reconciled
part of the oil sands processing plants, i.e. to the values of V at time k and k−1. It has the follow-
pipeline. ing analytical solution since the set of con-
The state and measurement vectors x and y of straints is linear:
size 26, as well as the criterion to be minimized,
y ( k ) − Σ E ( k ) T ( E ( k ) ΣE ( k ) T )
−1
are the same as in DRM 2 (see in Equation (10) xˆ (k ) =
the type of criterion to be minimized). The mass (14)
conservation constraints include both the ( E (k ) y (k ) − A(k ) xˆ (k − 1) )
steady-state mass conservation equations of the The NLP method varies the V(k) value until the
conditioning and extraction plants and the dy- minimal value of the criterion is delivered. The
namic mass conservation equations. It is com- DDR solution given by Equation (14) and the
posed of the same equations as DRM 1, aug- V(k) value at convergence are sub-optimal since
mented with: the reconciled values obtained at time k -1 are
not re-evaluated.
3
W4i (k ) 3
W5i (k )
ρi
∑
= ∑
=
ρi
V (k ) (12)
RESULTS AND DISCUSSION
=i 1 = i 1

(1 − V p / V (k ))W (k ) − W (k )
4
i
5
i Assuming that there are no leaks and no
(13) measurement bias, daily mass flowrate or com-
=
−V pWˆˆ4i (k − 1) / V (k − 1) for i =
1, 2 position estimates calculated either with raw
where V(k) is the cumulated volume that goes data or reconciled data obtained by the three
through the pipeline during day k. Equation (12) reconciliation methods are unbiased estimators
is a consequence of the assumption that the of their true values. Table 2 compares the rela-
pipeline hold-up volume is constant. Equation tive standard deviations of the process variables
(13) expresses the conservation of any pair of (those implying the bitumen phase only), using
the three components around the pipeline node, raw data or data reconciled by methods 1, 2,
assuming that past values are kept at their pre- and 3.
vious estimates. The third component conserva- Table 2 shows that DRM 1 is able to signifi-
tion equation is useless, since it would be re- cantly remove measurement errors on measured
dundant with Equations (12). Equation (13) is variables, since all estimates have accuracies
derived from Equation (4) where the densities lower than or equal to 5%. DRM 2 removes
and the products Fm are replaced by their cor- additional variance to some variables, because
responding expressions F/V and W. Since the of the additional constraints that link stream 4
conservation of the species 1 to 3 is enforced and stream 5, but adds noise to some other ones,
via Equations (12) and (13), the constraint of because of the distortion created by the en-
overall mass conservation of Equation (2) is forcement of steady-state behaviour for the
warranted. pipeline. Based only on the estimation error
Equation (13) is non-linear since it contains standard deviations for measured variables,
ratios of V and W. The minimization of the re- DRM 3 does not show any advantage over the
conciliation criterion of Equation (10) with re- two steady-state methods for most process
spect to the 27 variables (26 in x plus V(k)) variables. This is logical since the conditioning
subject to the 12 equations of DRM 1 plus the and extraction plants are steady-state. However
four equations (12) and (13) could be solved by DRM 3 is a better method, when looking at
Non-Linear Programming (NLP) methods, but performance indicators more sensitive to the
for simplification a hierarchical optimization pipeline dynamics, as is discussed below.
method consisting in two steps is preferred,

2327
 Table 2 Relative standard deviations of estimation errors for selected process variables/%
Measured variables Raw data DRM 1 DRM 2 DRM 3
F1 5 3.5 3.8 3.5
F3 5 5.0 5.3 5.1
F4 5 3.0 3.0 2.7
Daily cumulated F5 5 3.3 3.1 2.7
mass
F7 5 3.1 2.8 2.4
F8 5 4.1 4.3 4.2
m11 5 3.5 6.1 4.7
1
m 3 5 5.0 7.8 6.4
m14 5 3.6 6.8 5.2
1
m 5 5 3.9 3.4 3.6
Bitumen concentration
m71 5 5.0 5.1 6.5
m81 5 2.4 2.4 2.9
Performance indicators
Recovery Cond. Plant
10.2 0.1 0.1 0.1
F 4 m1 4 / F 1 m1 1
Recovery Extr. Plant
10.2 0.35 0.35 0.31
F 8 m1 8 / F 5 m1 5
Overall Recovery
10.2 6.6 7.2 3.9
F 8 m1 8 / F 1 m1 1

Bitumen recoveries are important perfor- 100 %, when using experimental data as illus-
mance indicators, and Table 2 shows that their trated in Fig. 4. The values calculated with the
relative standard deviations have been strongly reconciled data do not show this problem. As it
reduced by the reconciliation procedures. The should be, DRM 3 gives performances similar
nominal bitumen recoveries at conditioning and to those of DRM 1 and DRM 2 for individual
extraction plants are respectively 98.9 % and plant recoveries. However the overall recovery,
95.0 % for an overall recovery of 94.5 %. Since which embeds the pipeline, i.e. the only node
the bitumen content of the rejects is very low in exhibiting dynamics, is significantly more ac-
each plant, the recovery is quite high and meas- curate than those obtained through DRM 1 and
urement noise may lead to values higher than DRM 2.
1.4
Bitumen recovery

1.2
recovery

0.8

0 10 20 30 40 50 60 70 80 90 100
Days

1.2
Bitumen recovery

1.1
Bitumen

1
0.9
0.8
0.7

0 10 20 30 40 50 60 70 80 90 100
Days

Fig. 4 Bitumen recovery for the conditioning plant (top) and extraction plant (bottom); (x): calculated from raw data, (o):
reconciled by DRM 1

2328
 Table 3 shows some additional results on larger than it should be considering the pipeline
plant imbalances (residuals of steady-state mass dynamics (790 compared to 446). Reconciled
conservation equations) and bitumen accumula- values of DRM 1 lead to an estimation error that
tion in the pipeline ( F4 m14 − F5 m51 ), for compari- is of the same magnitude as the rate of accumu-
son of the methods. The raw data gives strongly lation (relative standard deviation of 90%).
incoherent data since the plant imbalances ex- DRM 2 gives an exactly 100% relative standard
hibit largely non null values. The reconciled deviation, as it could have been expected since
values with the two steady state methods are it enforces zero rate of accumulation. However,
now coherent and lead to zero imbalances for DRM 3 is able to improve the rate of accumula-
the individual plants. Table 3 also shows that tion accuracy (65% relative standard deviation),
the standard deviation of the pipeline daily ac- since it takes account of the pipeline dynamics
cumulation estimated with the raw data is twice
Table 3 Standard deviations of “real’’ daily bitumen plant imbalances and pipeline accumulation, and standard devia-
tions of their estimation errors (t).

Process Standard deviation of “real’’ Estimation error standard deviation

variable value Raw data DRM 1 DRM 2 DRM 3
Cond. Plant
0 655 0 0 0
imbalance
Extr. Plant
0 642 0 0 0
imbalance
Pipeline
446 790 417 446 304
accumulation

Another important output of mass balance remains bounded, as it must be, whereas the
calculation in a dynamic context is the estima- estimated inventory is drifting away from the
tion of the equipment hold-ups. The bitumen true value. DRM 2 produces a constant rather
inventory I(k) in the pipeline can be calculated than drifting inventory. Obviously this has the
by integration of the accumulation rate from its advantage of allowing perfectly balanced ac-
initial value, using I (k )= I (k − 1) + F4 m14 − F5 m51 . counting reports, but the drawback of exhibiting
The estimation error progressively increases and random errors larger than the measurement ones.
gives unacceptable values that would bring Using DRM 3 it is possible to use an alternative
problems for production accounting as illus- estimation of the bitumen inventory, according
trated in Fig. 5 for DRM 1, where it is assumed to W 4 (k)(V p /V(k)). Fig. 6 shows that the inven-
that the initial bitumen inventory in the pipeline tory variation is now close to the simulated in-
is perfectly known. The real bitumen inventory ventory.

8000

7000
Bitumen inventory/tons
Bitumen inventory (tons)

6000

5000

4000

3000

2000
0 5 10 15 20 25 30 35 40 45 50
Days

Fig. 5 Comparison between simulated (stairs) value and method-1 estimate (continuous) of bitumen inventory

2329
 4000

3800

3600

3400

Bitumen inventory (tons)

Bitumen inventory/tons 3200

3000

2800

2600

2400

2200

2000
0 5 10 15 20 25 30 35 40 45 50
Days

Fig. 6 Comparison between simulated (stairs) value and method-3 estimate (continuous) of bitumen inventory

Bellec, S, Hodouin, D, Bazin, C and Duchesne C, 2007.

CONCLUSIONS Multi-level data reconciliation - application to a gold ore
A dynamic data reconciliation method is processing plant, in Proceedings of 12th International Fed-
proposed to include plug-flow units, while satis- eration of Automatic Control (IFAC) Symposium on Auto-
mation in Mining, Mineral and Metal Processing, Québec,
fying the usual industrial constraint that the past
Canada, pp 39-44.
reconciled values must not be updated by the
Hodouin, D, Lachance, L and Desbiens, A, 2007. Dynamic data
current reconciliation. It is assessed using a
reconciliation: from the full model observer to the stationary
simple simulator of oil sands processing instal- on-line mass conservation filter – the flotation case, in Pro-
lation. The conclusions are that this method, ceedings of 12th International Federation of Automatic
compared to steady-state reconciliation, im- Control (IFAC) Symposium on Automation in Mining, Min-
proves the reliability of plant mass balance, eral and Metal Processing, Québec, Canada, pp 87-92.
particularly when material inventories and re- Lachance, L, Poulin, E, Hodouin, D and Desbiens, A, 2006.
coveries are to be estimated. The proposed illu- Performance of steady-state and stationary data reconcilia-
stration assumes that the variability of the tion as a function of process and disturbance dynamics, in
process variables is more related to measure- Proceedings of Mineral Process Modeling, Simulation and
ment errors than to true process dynamics. Thus, Control, Sudbury, Canada, pp 393-408.
the proposed method advantages would be more Lachance, L, Poulin, E, Hodouin, D and Desbiens, A, 2007.
significant if the process dynamics were the Tuning stationary observers: application to a flotation unit
dominant phenomenon, as is usually the case. simulator, in Proceedings of 12th International Federation of
The inclusion of a dynamic model in the Automatic Control (IFAC) Symposium on Automation in
Mining, Mineral and Metal Processing, Québec, Canada, pp
framework of mass balancing is an issue that
357-362.
will surely retain attention in the future, now
Liebman, M J, Edgar, T F and Lasdon, L S, 1992. Efficient data
that there is growing awareness in the industry
reconciliation and estimation for dynamic processes using
about mine to mill, and even beyond, monitor- nonlinear techniques, Computers & Chemical Engineering,
ing tools. 16(10/11):963-986.
Narasimhan, S and Jordache, C, 2000. Data Reconciliation and
Acknowledgements The authors would like to thank Gross Error Detection: An Intelligent Use of Process Data,
Syncrude Canada for providing the case study which 405 p (Gulf Publishing Company).
allowed the development of this data reconciliation
Romagnoli, J A and Sanchez, M C, 2000. Data Processing and
through mass balancing technique, and NSERC Canada
Reconciliation for Chemical Process Operations. 270 p
for providing the funding.
(Academic Press).
REFERENCES Simon, D, 2006. Optimal State Estimation, 520 p (Wiley).
Vaclavek, V, 1974. How long to take the balance periods at
Almasy, G A, 1990. Principles of dynamic balancing, AIChE balance calculations of quasistationary continuous processes,
Journal, 36(9): 1321-1330. Chemical Engineering Science, 29(12):2307-2313.

2330