Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
www.elsevier.com/locate/ces
b Manufacturing
Practice, Tata Consultancy Services, 54 B Hadapsar Industrial Estate, Pune 411 013, India
Practice, Tata Consultancy Services, Air India Building 11th Floor, Nariman Point, Mumbai 400 021, India
Received 6 May 2003; received in revised form 15 July 2003; accepted 10 September 2003
Abstract
The elitist version of nondominated sorting genetic algorithm (NSGA II) has been adapted to optimize the industrial grinding operation
of a lead-zinc ore bene5ciation plant. Two objective functions have been identi5ed in this study: (i) throughput of the grinding operation
is maximized to maximize productivity and (ii) percent passing of one of the most important size fractions is maximized to ensure smooth
8otation operation following the grinding circuit. Simultaneously, it is also ensured that the grinding product meets all other quality
requirements, to ensure least possible disturbance in the following 8otation circuit, by keeping two other size classes and percent solid of
the grinding product and recirculation load of the grinding circuit within the user speci5ed bounds (constraints). Three decision variables
used in this study are the solid ore 8owrate and two water 8owrates at two sumps, primary and secondary, each of them present in each
of the two stage classi5cation units. Nondominating (equally competitive) optimal solutions (Pareto sets) have been found out due to
con8icting requirements between the two objectives without violating any of the constraints considered for this problem. Constraints are
handled using a technique based on tournament selection operator of genetic algorithm which makes the process get rid of arbitrary tuning
requirement of penalty parameters appearing in the popular penalty function based approaches for handling constraints. One of the Pareto
points, along with some more higher level information, can be used as set points for the previously mentioned two objectives for optimal
control of the grinding circuit. Implementation of the proposed technology shows huge industrial bene5ts.
? 2003 Elsevier Ltd. All rights reserved.
Keywords: Dynamic simulation; Mathematical modeling; Multiobjective optimization; Genetic algorithm; Pareto set
1. Introduction
Grinding is one of the very important unit operations in
most of the mineral processing plants. Since grinding is a
very energy intensive process, modeling and thereby optimization of grinding operation of industrial scale has been a
continuous endeavor of the scientists and engineers. Though
modeling of grinding circuit has attained a reasonable state
of robustness (Herbst and Fuerstenau, 1973; Lynch and Rao,
1975; Herbst et al., 1983; Kinneberg and Herbst, 1984;
Rajamani and Herbst, 1984, 1991a), work on several aspects of optimization and control is yet to reach the similar
state of maturity (Lapidus and Luss, 1967; Bryson and Ho,
1969; Birch, 1972; Herbst and Rajamani, 1979; Rajamani
and Herbst, 1991b). Recently, a successful implementation
of modeling and control of an industrial leadzinc grinding
0009-2509/$ - see front matter ? 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ces.2003.09.036
386
2. Formulation
The grinding circuit is modeled 5rst. Each of the unit
operations, in this case they are rodmill, ballmill, hydrocyclones and sumps, are modeled separately and a connectivity
matrix, expressed in terms of 1 and 0, connects all of them
to simulate the whole circuit. In case of mill and sump operations, population-balance equations for solids and water
streams are considered. Each of the size class is tracked by
C1
Ore
Feed
C2
387
Final
Product
Water
Ball
Mill
Rod
Mill
Water
Water
Secondary
Sump
Primary
Sump
Other than this, there are several other locations where 5xed
amount of water is added to maintain smooth 8ow of slurry
within the circuit. These points are ballmill input (WBMF ) as
well as discharge (WBMD ) and rodmill input (WRMF ) as well
as discharge (WRMD ). These latter 8owrates are never manipulated whereas the former ones are manipulated to meet
the grinding product quality. The properties of the product
that draws attention of the grinding circuit performance are
throughput (TP ), size analysis (percentage passing of three
size classes namely coarse (PC ), mid (PM ) and 5ne (PF )
size classes), percentage of solids (PS ). Another very important parameter that is to be kept within speci5ed bound
is recirculation load (RL ). The model can predict all slurry
properties in any of the streams of the circuit given in Fig. 1
in steady state as well as dynamic mode.
The complete MOOP solved is, thus, written as follows:
MaxI [I1 ; I2 ]T
I1 = T P
I2 = P M
Subject to (s.t.):
PCL 6 PC 6 PCU
PFL 6 PF 6 PFU ;
PSL 6 PS 6 PSU ;
RLL 6 RL 6 RU
L:
All equations (Appendix A)
Decision variable bounds
L
U
6 MRM 6 MRM
;
MRM
L
U
WPS
6 WPS 6 WPS
;
L
U
WSS
6 WSS 6 WSS
;
388
The two objective functions used here are con8icting in nature and so it is likely that a Pareto set of nondominating
optimal solutions is obtained. The binary version of nondominated sorting genetic algorithm II (NSGA II) which
is an adaptation of the simple genetic algorithm suited for
multiobjective optimization problems, is used to solve the
problem de5ned above. Details of this method are available
in the literature (Deb, 2001a,b).
3. Results and discussions
The model is 5rst tested for its validity. As the decision
variables considered here are three and each one of them
has upper and lower bounds, all possible combinations (23 )
among these three decision variable bounds are considered.
Now for these eight operating conditions, data are collected
from industry and simulations are also run. The comparison
between the simulations and data for all these eight cases
is presented in Fig. 2. This shows the validity as well as
the sensitivity of the model over the whole operating region
considered. The cases considered are denoted as the bounds
for the decision variables used (e.g. case (a) L-L-H means
the simulation of lower bounds for decision variable 1 and
2 and upper bound on decision variable 3). The eight cases
considered are: (a) L-L-H (b) L-H-L (c) L-L-L (d) L-H-H
(e) H-L-H (f) H-H-L (g) H-L-L (h) H-H-H.
The Pareto optimal points were obtained for the problem considered. The spread obtained in the front was very
dense. The 5nal Pareto front was quite diIerent from the best
nondominated front obtained in the initial population. This
means better fronts were evolved as generations progressed.
The elitist approach of considering both parent and child
population for selecting the better candidates for the mating
pool was found to work very well. This was supported by
the fact that subsequent generations didnot allow better 5tted chromosomes to get lost in the crowd due to the randomness involved in the selection operation of GA. Crowding
metric helped to maintain diversity in the Pareto front.
Starting with some randomly selected candidate solutions
in the zeroth generation, NSGA II was observed to take
about twenty-5ve generations to converge to the Pareto set.
After convergence, it was found to maintain the same points
in the Pareto fronts over any number of generations. The
nondominated fronts found in the zero as well as generation
number twenty 5ve, for normalized objective functions, are
given in Fig. 3. The number of distinct Pareto points found is
of the order of half of the number of populations considered
in a generation as rest of the points are basically multiple
copies of existing ones.
After formation of Pareto set, one can choose an operating point from this set using his or her intuition or some
higher level information related to requirement of the process. In the present case, the concerned industry used to
cater a varied set of needs of its clients. In case of some
clients, the quality was allowed to be compromised a little
0.9
0.9
Cumulative Distribution
Cumulative Distribution
0.8
0.7
0.6
0.5
0.2
0.4
0.6
0.8
1.2
Normalized Size
0.5
0.9
0.9
0.8
0.7
0.5
0.2
0.4
0.6
0.8
0.5
1.2
0.2
(d)
0.8
1.2
Cumulative Distribution
0.9
0.8
0.7
0.6
0.4
0.6
0.8
1.2
1.2
1.2
Normalized Size
0.9
0.8
0.7
0.6
0.2
0.4
0.6
0.8
1.2
Normalized Size
0.5
(f)
0.9
0.9
Cumulative Distribution
0.8
0.7
0.6
0.2
0.4
0.6
0.8
Normalized Size
0.5
0.6
0.7
(g)
0.4
0.8
0.5
0.2
0.6
Normalized Size
(e)
Normalized Size
(c)
Cumulative Distribution
0.7
(b)
0.6
Cumulative Distribution
0.8
0.6
Cumulative Distribution
Cumulative Distribution
(a)
389
0.8
0.7
0.6
0.2
0.4
0.6
0.8
1.2
Normalized Size
0.5
(h)
0.2
0.4
0.6
0.8
Normalized Size
390
1.05
1
0.95
0.9
0.85
0.8
Generation 0
Generation 25
0.75
0.7
0.7
0.75
0.8
0.85
0.9
0.95
1.05
1.1
Normalized Throughput
classi5cation in the primary hydrocyclone and 5ne classi5cation in the secondary hydrocyclone. Three points are
identi5ed on the global Pareto front e.g. 5rst at the maximum value of TP , second at the minimum value of TP
and the third at the in between (average) the previous two.
Circuit behavior (cumulative size classi5cation plots) at all
the prime locations of the grinding circuit are presented in
Figs. 4 (maximum TP ), 5 (minimum TP ) and 6 (average
TP ). These locations are (a) rodmill discharge, (b) ballmill
discharge, (c) primary cyclone over8ow, (d) secondary
cyclone over8ow, (e) primary cyclone under8ow, (f) secondary cyclone under8ow, (g) mill discharge sump and
(h) secondary cyclone discharge sump. These 5gures show
that classi5cation becomes coarser as one moves from lowest
TP to highest TP in the global Pareto set. Any kind of
desired results that happens to be in between these three
cases (maximum TP , minimum TP and average TP ) can
be obtained from the other points present in the global
Pareto set. The biggest advantage of this kind of MOOP
formulation exercise is this provides a wide range on feasible solutions where from a user can choose the best one of
his or her choice.
The NSGA II parameters used to obtain the Pareto set
for the present study are as follows: Population size=80,
number of decision variables=3, string length for each decision variables=20, maximum number of generations=50,
crossover probability=0.9 and mutation probability=0.001.
The eIect of varying several GA parameters on the Pareto
solution was studied next. If the number of population is decreased from 80 to 50, without changing any other parameters, all the points generated randomly in the 5rst generation
were coincidentally nondominating in nature. As the population number decreases, convergence of initially generated
inferior fronts to Pareto set front was relatively faster (requirement of less number of generations for convergence)
but the spread of the Pareto set obtained was relatively poor.
Increase in the value of string length does not provide better
fronts. The eIect of varying the value of crossover probability, from 0.9 to 0.7, again led to no signi5cant changes in
the optimal solutions though number of generations taken to
converge to the 5nal Pareto was more (in case of 0.7) than
earlier (in case of 0.9). The eIect of a change in the value of
mutation probability, from 0.001 to 0.01, neither aIect the
spread as well as number of points nor imparts any improvement in the Pareto set. The variation of string length also
did not provide any better fronts, but if the string length is
decreased, the number and thereby the spread of the Pareto
fronts becomes relatively less dense. Once Pareto front is
achieved, any increase in maximum number of generations
successfully maintains the diversity in the Pareto front from
generation to generation.
This kind of MOOP formulation can be put into MPC
framework to provide more meaningful solution for controlling the process. The Pareto points can be used as a tool
for decision making and providing meaningful setpoints to
the grinding control problem solved by MPC (described
in the introduction). By solving the MOOP formulation,
a user can have multiple feasible solutions. It is not to be
forgotten that by solving SOOP formulation of MPC, one
can have only one solution at a time. Even the quality of
solution supplied by MOOP formulation of MPC is better
than that of SOOP formulation of MPC. Many solutions
that appear as unattainable or infeasible (for SOOP formulation of MPC) become feasible when MPC solves the
MOOP formulation. A rule based decision making system
(based on process experience) can be put one layer above
MPC based (MOOP formulation) supervisory controller to
decide which solution (out of several solutions provided in
the global Pareto front) to download to the process as set
point.
The proposed MOOP formulation when coupled with
MPC framework and implemented in the plant conditions,
the direct bene5ts are more in terms of signi5cant improvement in throughput (an increase of the order of 25%)
without compromising the product quality much. There has
been a remarkable improvement in zinc recovery (an increase of nearly 3% against an initial estimate of 1%) after
the installation of the controller compared to the prior state.
It is important to note that this increase in recovery has been
achieved without aIecting the product grade. This increase
in recovery impacted the return on investment hugely and
brought that down from 1.5 years (calculated initially with
an target improvement of zinc recovery of 1%) to little less
than half a year. The indirect bene5ts of the same on the
plant operating system are listed below:
A sustained increase in current throughput levels, besides
consistently meeting grinding circuit quality requirements vis-Qa-vis product size distribution and percentage
1
0.9
0.8
0.8
0.7
0.6
0.5
0.4
0.3
0.4
0.3
0.1
0
0.2
0.4
0.6
0.8
1.2
Normalized Size
(b)
1
0.9
0.8
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.4
0.6
0.8
0.6
0.8
1.2
1.2
1.2
1.2
0.6
0.5
0.4
0.3
0.2
0.2
(d)
1
0.9
0.8
0.8
Cumulative Distribution
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.4
0.6
0.8
Normalized Size
0.9
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.2
(e)
0.4
0.6
0.8
1.2
0.2
(f)
Normalized Size
1
0.9
0.8
0.8
Cumulative Distribution
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.4
0.6
0.8
Normalized Size
0.9
(g)
0.4
0.7
1.2
Normalized Size
0.2
0.1
0
(c)
Normalized Size
0.9
Cumulative Distribution
0.5
0.2
0.1
Cumulative Distribution
0.6
0.1
(a)
391
0.7
0.2
Cumulative Distribution
Cumulative Distribution
1
0.9
Cumulative Distribution
Cumulative Distribution
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.2
0.4
0.6
0.8
1.2
Normalized Size
(h)
0.2
0.4
0.6
0.8
Normalized Size
Fig. 4. Cumulative distribution curve for normalized size for various streams ((a) rodmill discharge, (b) ballmill discharge, (c) primary cyclone over8ow,
(d) secondary cyclone over8ow, (e) primary cyclone under8ow, (f) secondary cyclone under8ow, (g) mill discharge sump and (h) secondary cyclone
discharge sump) in the grinding circuit for maximum throughput condition in Pareto set.
extremely diRcult and time consuming) without affecting any key process variables. The controller
smoothly executes the step change given to the
process by keeping all other key process variables
unchanged.
0.9
0.9
0.8
0.8
Cumulative Distribution
Cumulative Distribution
392
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.6
0.8
0.4
0.3
1.2
1
0.9
0.8
0.8
Cumulative Distribution
0.7
0.6
0.5
0.4
0.3
0.6
0.8
1.2
(d)
1
0.9
0.8
0.8
0.5
0.4
0.3
0.2
0.1
0
(e)
0.4
0.6
0.8
1.2
1.2
1.2
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.4
0.6
0.8
1.2
(f)
1
0.9
0.8
0.8
Cumulative Distribution
0.7
0.6
0.5
0.4
0.3
0.6
0.8
1.2
0.8
0.4
0.3
0.1
Normalized Size
0.6
0.5
0.2
0.4
0.4
0.6
0.1
0.2
0.2
0.7
0.2
Normalized Size
0.9
(g)
0.2
0.1
0
Normalized Size
Normalized Size
0.6
1.2
0.3
0.9
0.7
0.4
0.1
Normalized Size
0.8
0.5
0.2
0.4
0.6
0.6
0.1
0.2
0.4
0.7
0.2
0.2
Normalized Size
0.9
(c)
(b)
Cumulative Distribution
Cumulative Distribution
0.4
Normalized Size
Cumulative Distribution
0.5
0.1
0
(a)
Cumulative Distribution
0.6
0.2
0.1
0
0.7
(h)
0.2
0.4
0.6
0.8
Normalized Size
Fig. 5. Cumulative distribution curve for normalized size for various streams ((a) rodmill discharge, (b) ballmill discharge, (c) primary cyclone over8ow,
(d) secondary cyclone over8ow, (e) primary cyclone under8ow, (f) secondary cyclone under8ow, (g) mill discharge sump (h) secondary cyclone
discharge sump) in the grinding circuit for minimum throughput condition in Pareto set.
levels
These
troller
totally
OFF.
0.9
0.9
0.8
0.8
Cumulative Distribution
Cumulative Distribution
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.4
0.6
0.8
1.2
Normalized Size
0.4
0.3
1
0.9
0.8
0.8
0.7
0.6
0.5
0.4
0.3
(c)
0.8
1.2
1.2
1.2
1.2
0.5
0.4
0.3
0.1
0
0.2
0.4
0.6
0.8
1.2
(d)
0.8
0.8
Cumulative Distribution
0.7
0.6
0.5
0.4
0.3
0.6
0.8
1.2
0.3
(f)
1
0.9
0.8
0.8
Cumulative Distribution
0.6
0.5
0.4
0.3
0.2
0.2
0.4
0.6
0.8
Normalized Size
0.9
0.7
0.8
0.4
0.1
Normalized Size
0.6
0.5
0.2
0.4
0.4
0.6
0.1
0.2
0.2
0.7
0.2
Normalized Size
0.9
0.7
0.6
0.5
0.4
0.3
0.2
0.1
(g)
0.6
0.6
0.4
0.7
0.9
(e)
0.2
0.2
Normalized Size
Normalized Size
0.9
0.1
Cumulative Distribution
0.5
(b)
0.2
Cumulative Distribution
0.6
0.1
0
Cumulative Distribution
Cumulative Distribution
(a)
0.7
0.2
0.1
0
393
0.1
0
0.2
0.4
0.6
0.8
1.2
Normalized Size
(h)
0.2
0.4
0.6
0.8
Normalized Size
Fig. 6. Cumulative distribution curve for normalized size for various streams ((a) rodmill discharge, (b) ballmill discharge, (c) primary cyclone over8ow,
(d) secondary cyclone over8ow, (e) primary cyclone under8ow, (f) secondary cyclone under8ow, (g) mill discharge sump (h) secondary cyclone
discharge sump) in the grinding circuit for average throughput condition in Pareto set.
4. Conclusions
A MOOP is solved for an industrial leadzinc grinding operation. One objective function considered is to
394
w
Q
RL
RLL , RU
L
ore density
water density
volume 8owrate (slurry)
recirculation load of the grinding circuit
upper and lower bounds for recirculation load
of the grinding circuit
grinding circuit product throughput
slurry volume
volume 8owrate (water)
water 8owrate at ball mill discharge
water 8owrate at ball mill feed
water 8owrate at primary cyclone
upper and lower bounds for water 8owrate at
primary sump
water 8owrate at rodmill discharge
water 8owrate at rodmill feed
water 8owrate at secondary cyclone
upper and lower bounds for water 8owrate at
secondary sump
TP
V
W
WBMD
WBMF
WPS
L
U
WPS
, WPS
WRMD
WRMF
WSS
L
U
WSS
, WSS
Subscripts
6
m
mf
mp
of
s
sf
sp
uf
Appendix A
A.1. Rod mill equations
Mmf = M6
(A.1)
(A.2)
(A.3)
(A.4)
(A.5)
(A.6)
(A.7)
mass fraction of
(A.8)
dHm (i)=dt = Mmf mmf (i) Mmp mmp (i) S(i)Hm mmp (i)
+
i1
395
(A.9)
j=1
Rf = ((K1 nc Ds + K2 nc )=Wsp ) + K3
bypass fraction;
Qmp = Qmf
(A.10)
(A.13)
Wuf = Rf Wsp
(A.16)
(A.17)
water in under8ow
+ K8 (Qsp =nc ) + K9
1:5725
SI is the cyclone sharpness index
ln(SI )
E(i) = (i)(1:0 Rf ) + Rf corrected eRciency for
ith size:
(A.31)
dHm (i)=dt = Mmf mmf (i) Mmp mmp (i) S(i)Hm mmp (i)
(A.18)
Hs = Csp Vs
under8ow solids;
(A.32)
(A.34)
(A.30)
n=
Muf =
j=1
hydrocyclone d50;
(A.29)
(A.28)
(A.14)
(A.15)
(A.27)
(A.12)
mass fraction of
i1
(A.25)
(A.20)
(A.22)
(A.23)
(A.24)
(A.35)
(A.36)
396
(A.37)
(A.38)
(A.39)
(A.40)
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A Fast and Elitist
Multiobjective Genetic Algorithms. IEEE Transactions on Evolutionary
Computation 6 (2), 182197.
Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization and
Machine Learning. Addision-Wesley, Reading, MA.
Herbst, J.A., Fuerstenau, D.W., 1973. Mathematical Simulation of Dry
Ball Milling Using Speci5c Power Information. Transactions of the
AIME 254, 343.
Herbst, J.A., Rajamani, K., 1979. Evaluation of Optimizing Control
Strategies of Closed Circuit Grinding. Proceedings of the 13th
International Mineral Processing Congress, Warsaw.
Herbst, J.A., Siddique, M., Rajamani, K., Sanchez, E., 1983. Population
based approach to ball mill scale-up: bench and pilot scale
investigations. Transactions of the AIME 272, 19451954.
Holland, J.H., 1965. Adaptation in Natural and Arti5cial Systems.
University of Michigan Press, Ann Arbor, MI.
Kasat, R.B., Kunzru, D., Saraf, D.N., Gupta, S.K., 2002. Multiobjective
optimization of industrial FCC unit using elitist non-dominated sorting
genetic algorithm. Industrial Engineering Chemistry Research 41,
47654776.
Kinneberg, D.J., Herbst, J.A., 1984. Comparison of Models for the
Simulation of Open Circuit Ball Mill Grinding. International Journal
of Mineral Processing.
Lapidus, L., Luss, R., 1967. Optimal Control of Engineering Process.
Blaisdell, Waltham, MA.
Lynch, A.J., Rao, T.C., 1975. Modeling and Scale up of Hydrocyclone
Classi5ers, Proceedings of the 11th International Mineral Processing
Congress, Cagliari, pp. 245 269.
Mitra, K., Deb, K., Gupta, S.K., 1998. Multiobjective dynamic
optimization of an industrial nylon 6 semibatch reactor using genetic
algorithm. Journal of Applied Polymer Science 69, 6987.
Momaya, P., Mitra, K., Sistu, P., Gopinath, R., 2003. Advanced process
control of ore bene5ciation in mineral processing plants. Proceedings
of the International Symposium on Process Systems Engineering and
Control 03, Jan 3 4, IIT, Bombay, pp. 126 130.
Nandasana, A.D., Ray, A.K., Gupta, S.K., 2003. Applications of the
non-dominated sorting genetic algorithm (NSGA) in chemical reaction
engineering. International Journal of Chemical Reactor Engineering 1
(R2), 116.
Petzold, L.R., 1983. A Description of DASSL: a diIerential/algebraic
system solver. In: Stepleman, R.S., et al. (Eds.), Scienti5c Computing.
North-Holland, Amsterdam, pp. 6568.
Rajamani, R.K., Herbst, J.A., 1984. Simultaneous estimation of selection
and breakage functions from batch and continuous grinding data.
Transactions of the Institution Mining and Metallurgy C93, C74C85.
Rajamani, R.K., Herbst, J.A., 1991a. Optimal control of a ball mill
grinding circuitI. Grinding circuit modeling and dynamic simulation.
Chemical Engineering Science 46 (3), 861870.
Rajamani, R.K., Herbst, J.A., 1991b. Optimal control of a ball
mill grinding circuitII. Feedback and optimal control. Chemical
Engineering Science 46 (3), 871879.
Rajesh, J.K., Gupta, S.K., Rangaiah, G.P., Ray, A.K., 2000. Multiobjective
optimization of steam reformer performance using genetic algorithm.
Industrial and Engineering Chemistry Research 39, 706717.
Rajesh, J.K., Gupta, S.K., Rangaiah, G.P., Ray, A.K., 2001. Multiobjective
optimization of industrial hydrogen plants. Chemical Engineering
Science 56, 9991010.
Ravi, G., Gupta, S.K., Viswanathan, S., Ray, M.B., 2002. Optimization of
venturi scrubbers using genetic algorithm. Industrial and Engineering
Chemistry Research 41, 29883002.
Ray, W.H., 1989. Advanced Process Control. Butterworths, New York.