Available online at www.sciencedirect.

com

Powder Technology 186 (2008) 31 39 www.elsevier.com/locate/powtec

Constrained model predictive control in ball mill grinding process

Xi-song Chen , Qi Li, Shu-min Fei
School of Automation, Southeast University, Nanjing, Jiangsu Province, 210096, China Received 27 October 2006; received in revised form 12 July 2007; accepted 26 October 2007 Available online 7 November 2007

Abstract Stable control of grinding process is of great importance for improvements of operation efficiency, the recovery of the valuable minerals, and significant reductions of production costs in concentration plants. Decoupled multi-loop PID controllers are usually carried out to manage to eliminate the effects of interactions among the control loops, but they generally become sluggish due to imperfect process models and a close control of the process is usually impossible in real practice. Based on its inherent decoupling scheme, model predictive control (MPC) is employed to handle such highly interacting system. For high quality requirements, a three-input three-output model of the grinding process is constructed. Constrained dynamic matrix control (DMC) is applied in an iron ore concentration plant, and operation of the process close to their optimum operating conditions is achieved. Some practical problems about the application of MPC in grinding process are presented and discussed in detail. 2007 Elsevier B.V. All rights reserved.
Keywords: Model predictive control; Dynamic matrix control; PID control; Ball mill; Grinding process

1. Introduction Ball mill grinding is a fundamental operation process, and in many respects the most important unit operation in a mineral processing plant. Grinding process represents almost half of the total operating costs associated with the mining operation, and the product particle size greatly influences the recovery rate of the valuable minerals and the volume of tailing discharge in the subsequent processes. Low qualified rate of product particle size can cause unacceptable economic loss and could be harmful for pollution control. For effective concentration or subsequent mineral liberation, grinding process has to maintain the following outputs (or controlled variables) stable, mainly including the product particle size distribution, circulating load and mill solids concentration, etc [13]. The control of grinding processes is a challenging problem due to two basic reasons. First, a large number of variables involved have strong couplings among them. If regulating one

controlled variable is attempted with one manipulated variable, the other controlled variables will, more likely than not, be influenced in an undesired fashion. Second, the existence of large time delays, time-varying parameters and nonlinearities are some of the other difficulties encountered by the control engineers. Many research papers have been published on the development of control strategies for grinding process. Traditionally, the grinding processes are controlled by multiple single-input single-output (multi-SISO) PID controllers which are usually decoupled to avoid multivariable interactions [1]. Fig. 1 shows a widely used multi-SISO (two-input two-output) system with two decouplers, namely D12(s) and D21(s). D12 s D21 s W12 s W11 s W21 s W22 s 1 2

Corresponding author. Tel.: +86 25 83792724; fax: +86 25 84545542. E-mail addresses: chenxisong@263.net (X. Chen), liq_kj@js.gov.cn (Q. Li), smfei@seu.edu.cn (S. Fei). 0032-5910/$ - see front matter 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.powtec.2007.10.026

where Wij(s)(i = 1,2; j = 1,2) are process models. One difficulty of this kind of control lies in the choice of the proper inputoutput pairing. When manipulated variables are not properly selected, interactions between controlled and

32

X. Chen et al. / Powder Technology 186 (2008) 3139

Fig. 1. Multi-SISO system with decouplers.

manipulated variables can result in undesirable control loop interactions, leading to poor control performances [1]. Another difficulty is that static decoupling (based on steady-state gains) is used much more often than dynamic decoupling in practice because the dynamic version may not be physically realizable [4,12]. Therefore, a close control of the process based on conventional multi-SISO PID controllers is usually impossible in real practice for a long time, and the controllers generally become sluggish especially when model mismatch occurs due to great disturbances, such as the changes in ore hardness and fresh ore feed size, etc. Although robust control [5], adaptive control [6], neural control [7] and multivariable control [8] are studied either in simulation level or in experimental level, very few of them currently operate in industry. Moreover, in most of the literatures [1,68], two-input two-output model is usually employed for control study. However, two controlled variables are the minimum requirement for a basic grinding process. For the process with high quality requirements, it is often not adequate to meet the production needs if only two controlled variables are considered. Presently MPC has become the most widely used multivariable control algorithm in the chemical process industries and in other areas [9]. Unlike many other multivariable control strategies, MPC was conceived primarily by industry. Dynamic matrix control (DMC) [10] and model algorithmic control (MAC) [11] are the two powerful MPC multivariable control

strategies. While these algorithms differ in certain details, the main ideas behind them are very similar. MPC schemes are popular in multiple variable processes because: (1) MPC can predict the future behaviors of a process with large time delays, (2) it has no variables paring problems in multivariable systems, and (3) it can easily handle constraints imposed on both the manipulated and controlled variables. MPC also has the prominent features such as relatively easy to tune, and less sensitive to model mismatch compared with many other model based advanced control strategies [12]. This paper presents an application of multivariable model predictive control scheme to a ball mill grinding process. The rest of the paper is organized as follows: A three-input threeoutput model of grinding process is constructed in Section 2. After a brief description of DMC scheme in Section 3, Section 4 presents an industrial application with constrained DMC schemes in an iron concentration plant. Some practical problems about the application of MPC in grinding process are presented and discussed in Section 5. Conclusions are given in Section 6. 2. Process description 2.1. Grinding process The grinding process studied in this paper operates in a closed loop as shown in Fig. 2, including a ball mill and a spiral classifier. The feed, iron ore (from primary crusher, size 14 mm) is fed into the ball mill by swaying feeders. The tumbling action of the balls within the revolving mill crushes the feed to finer sizes. The slurry containing the fine product is discharged from the mill to a classifier. The slurry is separated into two streams: an overflow stream containing the finer particles and the circulating stream containing the larger particles. The overflow is the desired product. The circulating is recycled back to the ball mill for regrinding. The product particle size is specified as 72% passing 200 mesh screens. As mentioned above, simpler control variables' pairing was studied by earlier researchers, and the circulating load and the product size were usually taken as controlled variables.

Fig. 2. Process diagram of grinding process.

X. Chen et al. / Powder Technology 186 (2008) 3139

33

tion based model is quiet desirable by performing a plant test. The most common plant test is to make a step change in the manipulated input and observe the measured process output response. Then a model is developed to provide the best match between the model output and the observed plant output [4,12]. By far the most commonly used model, for control system design purpose, is the first-order plus time-delay model or second-order plus time-delay model, while the former can be represented by the transfer function relationship
Fig. 3. The relationships among the variables in grinding process.

ys

Kp ess u s Tp s 1

However, in the process studied, in order to keep these two controlled variables around their set-points, mill solids concentration is another key intermediate controlled variable that should be also stably controlled. For example, the change of ore hardness or feed size can be first identified by the variation of mill solids concentration, thus some control strategy of controlling mill solids concentration by manipulating of the fresh solids feed can be carried out in advance before the particle size has been influenced. Too high mill solids concentration could eventually lead to blockage in the mill, while too low solids concentration undoubtedly causes inefficiency in grinding [2]. To sum up, the controlled variables in this process are the product particle size, mill solids concentration and the circulating load. Three manipulated variables are available: the fresh ore feed rate, mill feed water flowrate and dilution water flowrate. The principal disturbances affecting the system are the ore hardness, feed size, etc. The relationships among the variables in grinding process can be shown in Fig. 3. 2.2. Process models For many processes it is a time consuming work to develop fundamental process models, while developing a transfer func-

where Kp is process gain, is dead time, Tp is process time constant. For the purpose of control study and industrial application many experiments have been conducted in the process around the normal operating point to show whether the increase in a particular manipulated variable increases, decreases or has a complex relationship with the controlled variables and whether the action is slow or fast. Fig. 4, for instance, illustrates the responses of product particle size to the step changes in fresh ore feed rate and dilution water flowrate respectively. As shown in Fig. 4(1), the step increase in fresh ore feed rate + 5 t/h at the normal operating point causes coarser product particle size or a decrease in terms of the percentage passing 200 mesh (%-200 mesh). The response model can be approximated as follows: Gs Kp ess ys us Tp s 1 4

As shown in Fig. 4(2), the step increase in the dilution water flowrate +6 m3/h initially causes a finer classification. However, this condition generates an increase of the solid circulating load to the mill feed. Then, the discharge from the mill slowly becomes coarser and consequently the overflow particle size

Fig. 4. Responses of particle size to step disturbance in the manipulated variables: (1) feed rate + 5 t/h, and (2) dilution water flowrate +6 m3/h.

34

X. Chen et al. / Powder Technology 186 (2008) 3139

becomes coarser until the operation stabilizes at a particle size slightly lower than initial. This response model can be approximated by a combination of two second-order plus time delay models: Gs Kp1 es1 s Kp2 es2 s Tp1 s 1 Tp2 s 1 Tp1 s 1 Tp2 s 1 Kp1 es1 s Kp2 es2 s Tp1 s 1 Tp2 s 1

system needs 2 decouplers to solve the problems of loop interactions, while for a three-input three-output model total 6 decouplers are required which will bring great difficulties to the industrial implementation [4]. Moreover, static decoupling is often used which will degrade the control performance in real practice, while MPC scheme is much suitable for this case. 3. MPC control algorithms The underlying philosophy of DMC scheme [10] consists of predictive model, reference trajectory, feedback correction and rolling optimization as shown in Fig. 5. DMC can be stated as follows: at any sampling instant, given a reasonably accurate predictive model and the desired future closed loop behavior (or a reference trajectory), plan the set of future control moves in such a way that the predictive output is as close to the reference trajectory as possible without any violations in the operating constraints. DMC scheme is based on finite step response models which are obtained by making a unit step input change to a process operating at steady state as mentioned above. The step response model is the vector of step response coefficients, a a1 a2 a3 N aN T 7

Two important rules should be complied with when conducting the experiment in the industrial process: 1) The step changes in ore feed rate or water flowrates should be limited to the ranges of 5% to 15% of the normal values. If the step change is too small, the measured output may not change enough to develop a reliable model. On the contrary, if the step change is too big, the output variable may change too much which will cause economic lost. Also, if the step change is too large, nonlinear effects may dominate. That is, the operating condition may become significantly different than the desired condition [4,12]. 2) Adequate times of the same experiment, usually more than 5, are needed to guarantee the accuracy of the results [4]. Similarly, the responses of particle size to mill feed water as well as the responses of the other two controlled variables to the three manipulated variables are obtained in the same way respectively, and they are omitted here for space. Let Ff, Wm and Wd denote manipulated variables, namely fresh ore feed rate, mill feed water and dilution water respectively, and Sp, Dm, Fc controlled variables, namely particle size, mill solids concentration and circulating load. The grinding process model can be written as: Y GU
2 3 2 3 2 3 2

where ai represents the step response coefficient for the ith sample time. N is the model length which should be long enough so that the coefficient values are relatively constant, i.e., the process is close to a new steady state. Mathematically, the control problem is posed as a minimization of a quadratic objective function of the future prediction errors subject to constraints on future manipulated variables and controlled variables. Let the process dynamics of the MIMO system be represented by the following convolution model: ym k 1 y0 k 1 ADuk 8

Sp s u1 s Ff s y1 s where Y 4 y2 s 5 4 Dm s 5; U 4 u2 s 5 4 Wm s 5; y3 s Fc s u3 s Wd s 2 3 G11 s G12 s G13 s and G 4 G21 s G22 s G23 s 5A total of 9 open loop G31 s G32 s G33 s

6 3

responses, Gij(s) (1 i 3, 1 j 3), are obtained as shown in Table 1. It's obvious that strong interactions exist between the manipulated and controlled variables. As for conventional multiloop PID controllers as shown in Fig. 1, a two-input two-output
Table 1 Transfer functions matrix of grinding process Process variables Sp, particle size (%-74 m) Dm, mill solids concentration (%solids) Fc, circulating load (t/h) Ff, feed rate (t/h) G11 s 2:85e 95s1
1:57 G21 s 130s1
45s

where ym(k + 1) is the model output vector, ym(k + 1) = [ym(k + 1) ym(k + 1) ym(k + P)] T ; u(k) is the control action, u(k) = [u(k) u(k + 1) u(k + M 1)]T; y0(k + 1) is the initial output vector without control action, y0(k + 1) = [y0(k + 1) y0(k + 2) y0(k + P)] T, and A(k) is the unit step response coefficient matrix, (P M). P is the prediction horizon, and M is the control horizon. The values of the unit step response coefficients are obtained by normalizing the transient responses of the circuit for step changes in inputs by dividing the step response values by the respective steady-state gains. Due to the measurement noise, and the unmodelled load disturbance effects as well as the plant-model mismatch, the

Wm, mill feed water flowrate (m3/h) G12 s 140s1210s1 3:46 G22 s 130s1
1:87e60s 11:09e260s 0:38e G32 s 130s1270s1
100s

Wd, dilution water flowrate (m3/h)

8:32e10s 11:07e220s G13 s 21s147s1 1:55e90s G23 s 85s1190s1

G31 s 1:14e 140s1

50s

G33 s 3:59e 180s1

10s

X. Chen et al. / Powder Technology 186 (2008) 3139

35

Fig. 5. Diagram of DMC.

predictive output should be corrected by the real output y(k) to realize close-loop control. yp k 1 ym k 1 g0 yk ym k ym k 1 g0 e k 9

where yp(k + 1) is the predictive output after correction, g0 is the coefficient matrix usually defined as g0 = [1 1 1], and e(k) is the error by feedback correction, e(k) = y(k) ym(k). Note that u(k) represents the future manipulated input moves to be decided by the optimization procedure. Thus, the DMC problem can be stated as h iT h i J k yr k 1 yp k 1 Q yr k 1 yp k 1 Duk T RDuk 10 where y r(k + 1) is the expected vector, y r(k + 1) = [y r(k + 1) y r (k + 2) y r(k + P)] T; Q is the error weighting factor vector, Q = diag[q1 q2 qp] T, and R is the input weighting factor vector, R = diag[r1 r2 rM] T. Magnitude and rate constraints on the manipulated variables and the controlled variables are expressed as followed: Manipulated variable constraints: umin V uk V umax ; k 0; 1; N M 1 k 0; 1; N M 1 11 12 13

over the prediction horizon and a future manipulated variable trajectory is calculated that minimizes the objective function satisfying all the constraints. Only the first move is applied to the plant and this step is repeated for next sampling instant. The desired closed loop behavior and MPC tuning parameters have a straight forward relationship. If tighter control of some specific controlled variables is desired, it can be achieved by choosing a corresponding higher weight in the factor Q. Also excess variations in manipulated variables' moves can be suppressed by an appropriate choice of R. 4. Industrial application The system based on this control strategy was put into operation in an iron ore concentration plant in 2004. Table 2 shows the principal specifications of the grinding process. The schematic diagram of DMC applied in the ball mill grinding process is shown in Fig. 6. All the signals coming from the sensors of the grinding process are collected in a SCADA (Supervisory Control And Data Acquisition) system. This system is based on hierarchical concepts, organized in functional and geographical groups. All

Manipulated variable rate constraints: Dumin V uk uk 1 V Dumax ; Controlled variables constraints: ymin V yk V ymax ; k 0; 1; N P 1

MPC scheme is implemented in a receding horizon framework. At any sampling instant, the optimization problem is formulated
Table 2 Main technical parameters of grinding process Particle size of fresh ore feed Ball mill Spiral classifier Overflow particle size Circulating load Mill solids concentration 14 mm Grid 2700 3600 mm 2400 mm 72 2% ( 74 m or 200mesh) 140180 t/h 68 78%

Fig. 6. The schematic diagram of DMC applied in the ball mill grinding process.

36

X. Chen et al. / Powder Technology 186 (2008) 3139

the signals received are saved in a database. Also, control strategies can be programmed on it and it has two displays directly connected to the grinding console. Fig. 7 shows the run results of the controlled variables and the manipulated variables with DMC scheme in the grinding process. This figure employed here is particularly for the illustration of the changing procedure from one stable point to another operating point when great disturbances occur to the process, such as an increase of ore hardness or feed size, etc. In the first half of the time observed, the three controlled variables are all maintained within the allowable values as shown in Fig. 7(1)(3), these being the typical run results of the control strategy in this concentration plant.

In order to illustrate the regulating procedures when great disturbances occur to the control system, the responses of the three manipulated variables are also presented. As shown in Fig. 7, when increased ore hardness or feed size is first detected by the increasing mill solids concentration (see Fig. 7(1)), the fresh ore feed rate is then decreased (see Fig. 7(4)) and mill feed water is increased (see Fig. 7(5)). Since the disturbances are too strong at this time, fresh ore feed rate and mill feed water are all manipulated to be near their constraints. Coarser product particle size is then measured (see Fig. 7(2)) and dilution water is increased quickly to reject this variation (see Fig. 7(6)). These procedures will inevitably lead to high circulating load as shown in Fig. 7(3).

Fig. 7. The run results of grinding process with DMC scheme: (1) mill solids concentration, (2) particle size, (3) circulating load, (4) fresh ore feed rate, (5) mill feed water flowrate and (6) dilution water flowrate.

X. Chen et al. / Powder Technology 186 (2008) 3139

37

Fig. 7 (continued ).

It is evident that the product particle size, as shown in Fig. 7 (2), is sometimes beyond the allowable values when severe disturbance happens to the control system. However, the control results have been much improved compared with those by PID control. As shown in Table 3, the qualified rate of product particle size during testing time are compared between the grinding circuit with DMC and that with PID in the same concentration plant in 2004. The two grinding circuits run in parallel with the same process. As can be seen, the qualified rate of product particle size has been improved from about 88% to more than 95% during the three months' testing time. That is, the operation of the process close to their optimum operating conditions has been achieved. Although great improvement has been achieved by DMC strategy compared with that by PID control, there still exits space for the qualified rate to increase as discussed below. 5. Problems and discussions 5.1. Tuning work and constraints problem Several parameters should be properly tuned in DMC since they affect the close-loop performance, such as model length, prediction horizon, control horizon, manipulated input weighting, etc.

The following tuning and weighting factors are used while applying the MPC scheme to control the grinding process: Prediction horizon P = 20, control horizon, M = 5, weights on errors Q = diag[1 0.5 0.7]T and weights on the manipulated variables R = diag[1 0.6 0.5]T. It is important to make certain that the model length is long enough to capture the steady state change, and it is also crucial to tune the prediction horizon to be long enough. Generally, control horizon is much shorter than prediction horizon, yielding more robust performance. Compared with multi-SISO PID controllers, the tuning work is straightforward. For example, because the particle size should be more tightly controlled compared with the other two controlled variables, the weight on product particle size (y1) is set to a larger value of 1 than those of the others. The manipulated variables are limited to acceptable values established by the normal operation of the plant for safety or economic reasons. The constraints introduced on the manipulated variables in the grinding process are shown in Table 4. The minimum values of the water flowrates are set to avoid blockage, while the maximum values are limited by the saturations of control valves. Employing too harder constraints on the manipulated variables, however, may cause considerably sluggish response of the controlled variables. During the early time of industrial

Table 3 Comparison of running results between DMC and PID during testing time September Qualified rate of product particle size (%) Average fresh ore feed rate (t/h) DMC PID DMC PID 95.5 87.1 71.2 71.1 October 96.2 88.5 69.7 69.6 November 95.9 88.2 70.8 70.9

Table 4 Constraints on the manipulated variables in the grinding process Feed rate (t/h) Minimum value Maximum value Maximum deviation rate 65 80 3 Ball mill feed water flowrate (m3/h) 2 15 5 Dilution water flowrate (m3/h) 5 50 10

38

X. Chen et al. / Powder Technology 186 (2008) 3139

commissioning, feed rate is constrained between 68 t/h and 74 t/ h for a relatively stable mill throughput, the responses of particle size in 2 h are shown in Table 5. It can be seen that the results of particle size are much worse than that in Fig. 7, since the particle size is out of the allowable limits in nearly 25% of the time observed. This is usually because the operating point drifts from the original one due to great disturbances. 5.2. Model mismatch problem The reason why the qualified rate of product particle size cannot reach 100% in real practice is mainly due to model mismatches caused by the great disturbances, such as the changes in feed particle size and ore hardness, etc. Although the tuning procedure (for example, choosing a longer prediction horizon) makes the controller insensitive to plantmodel mismatches, stable control of grinding circuits is still occasionally interrupted by great disturbances as shown in Fig. 7. These disturbances may reduce the mill throughput drastically by causing the mass flowrate of the classifier overflow and particle size distribution to experience continuous perturbations. From the point of view of process control, the variation causes the changes in model parameters and leads to severe nonlinearities because any process models are constructed around a particular operating point. However, conventional DMC scheme employs an explicit model to predict the future behavior of the process over an extended prediction horizon. Although the performance of MPC is proved satisfactory under moderate variations of model parameters, if there is large modeling error of the controller, the closed-loop response is getting poorer, even the stability cannot be guaranteed [13]. The control performance may be further improved by two ways: One is to screen the feed size and ensure it below 14 mm. However, it has no effects on the variation in ore hardness. Another way is to test some nonlinear MPC (NMPC) strategies which are hot topics discussed at present in various literatures. These will not be discussed in this paper.

5.3. Mill throughput control In principle, grinding processes mainly aim to achieve maximum throughput at a specified product particle size. It can be interpreted as that the fresh ore feed rate should be increased as much as possible while maintaining product particle size stable. It seems that the fresh ore feed rate should be considered as a controlled variable. However, the fresh ore feed rate has to be taken as a manipulated variable because it is the only final determinative variable affecting product particle size. As shown in Fig. 4(2), although dilution water has a very fast short-term influence on product particle size, it has little influence on the final particle size since particle size stabilizes at a level slightly lower than initial. Mill feed water also has the similar response according to the response models listed in Table 1. In general, fresh ore feed rate has the long-term effects on product particle size. In order to solve this problem, fresh ore feed rate is taken as a manipulated variable while moderate constraints are imposed on it. As shown in Table 4, the normal value of fresh feed rate in the grinding circuit is 70 t/h, the reason that the low limit is set as 65 t/h, not less, is mainly for economic reasons in case of too low mill throughput. In addition, the weight on fresh ore feed rate (u1) is set to 1 which is larger than those of the other two manipulated variables because feed rate can't be allowed to move too freely in order to maintain a relatively stable and high mill throughput. The high limit of feed rate, however, is set as 80 t/h in case of blockage or overload of the mill and classifier. As a manipulated variable, ore feed rate will be decreased when ore hardness is high for example, leading to a low mill throughput. On the contrary, when ore hardness is low, MPC will automatically increase fresh ore feed rate, which will remedy, from another point of view, the decrease of mill throughput. This procedure has been demonstrated by the almost same average fresh ore feed rate under DMC and that under PID control as shown in Table 3. That means the control of product particle size is not at a cost of reduced mill throughput. 5.4. Operational failures by sensors and actuators During the testing time, the DMC scheme must be switched off to manual mode when any measurement errors happen or sensors (such as on line particle size analyzer, density meters, etc) are in maintenance or in calibration. Some measures have been taken to solve these kinds of problems. When a measurement error (such as abrupt too low flowrate) is detected the control system will exert a frozen action to keep the present output and give alarms immediately. As to the most important variable, namely product particle size, a soft sensor is developed in the control system to infer current product particle size when maintenance or calibration work is being carried out on this equipment. Other problems concern the actuators applied in the control system, such as fresh ore feeders and control valves, etc. As for ore feeders, there are four or five swaying feeders for every grinding circuit in this plant. Two feeders work in normal condition while the left are for redundancy. If ore blockage or

Table 5 The response of particle size during 2 hours in early time Time interval (min) Overflow particle size (%-74 m) Fresh feed rate (t/h) 07 716 1621 2127 2739 3945 4550 5059 5968 6882 8291 91101 101113 113120 72 73 72 71 70 69 69 68 69 70 71 72 72 73 70 70 71 70 70 69 68 68 68 68 68 69 70 71

X. Chen et al. / Powder Technology 186 (2008) 3139

39

mechanical problems occur to a swaying feeder, a redundant feeder will be automatically switched on to reject this kind of operational failures. Since recycled water is utilized in the concentration plant and the water is usually not clean enough, a water filter equipment is adopted to filter the impurities in case the control valves are blocked. Moreover, V-shape ball valves are also employed in the water pipes in case of blockage since they have better shearing function compared with direct-acting valves. It should be pointed out that the upper limits constrained on the mill feed water and dilution water flowrates are sometimes higher than the current maximum real output. That is, although the control valve has been totally opened the mill feed water cannot reach the high limit because the water pressure is lower than normal value. The fluctuations of water pressure cause main disturbance to water flowrate and then leads to a poor control performance of product particle size. These kinds of problems can be solved by enlarging water pipes and installing a water pressure control system, such as variable-voltage variable-frequency (VVVF) inverters. 6. Conclusions Model predictive control is employed to handle the highly interacting multivariable system of grinding process. A threeinput three-output model of grinding process is constructed for the high quality requirements of the process studied. Constrained dynamic matrix control is applied in an iron ore concentration plant. Compared with conventional decoupled multi-loop PID controllers, the run results of DMC scheme demonstrate its better performances in process control. DMC schemes can handle manipulated variable constraints explicitly and operation of the circuits close to their optimum operating conditions is achieved. The run results show that the proposed DMC scheme is an effective way to control ball mill grinding process. Acknowledgements The authors are indebted to many collaborators who have contributed to the work on the experimental work and system

commissioning in industry. In particular, the staffs in Fudao industrial control network company and the engineers of the concentration plant are grateful acknowledged for their valuable assistance. Funding for process modeling and control research is provided by project 8508002025 and 8508002049, Southeast University. References
[1] A. Pomerleau, D. Hodouin, A. Desbiens, E. Gagnon, A survey of grinding circuit control methods: from decentralized PID controllers to multivariable predictive controllers, Powder Technology 108 (2) (2000) 103115. [2] X.S. Chen, L.L. Wang, C.H. Wang, J.S. Zhang, Automatic control system design for ball mill in concentration process, Electrical Automation 26 (4) (2004) 6364. [3] V.R. Radhakrishnan, Model based supervisory control of a ball mill grinding circuit, Journal of Process Control (1999) 195211. [4] X.S. Chen, M.L. Wang, Process control system, Science Press, Beijing, China, 2005. [5] O. Galan, G.W. Barton, J.A. Romagndi, Robust control of a SAG mill, Powder Technology 124 (2002) 264271. [6] K. Najim, D. Hodouin, A. Desbiens, Adaptive control: state of the art and application to a grinding process, Powder Technology 82 (1) (1995) 5968. [7] A.V.E. Conradle, C. Aldrich, Neurocontrol of a ball mill grinding circuit using evolutionary reinforcement learning, Minerals Engineering 14 (2001) 12771294. [8] M. Duarte, F. Sepulveda, A. Castillo, A. Contreras, V. Lazcano, P. Gimenez, L. Castelli, A comparative experimental study of five multivariable control strategies applied to a grinding plant, Powder Technology (1999) 128. [9] J.B. Rawlings, Tutorial overview of model predictive control, IEEE Control Systems 20 (3) (2000) 3852. [10] C.R. Cutler, B.L. Ramaker, Dynamic matrix control a computer control algorithm, AIChE national meeting, Houston, TX, 1980. [11] J. Richalet, A. Rault, J.L. Testud, J. Papon, Model predictive heuristic control: application to industrial processes, Automatica 14 (1978) 413. [12] B.W. Bequette, Process control: modeling, design, and simulation, Prentice Hall, New Jersey, USA, 2002. [13] Z. Zhao, X.H. Xia, J.C. Wang, J. Gu, Y.H. Jin, Nonlinear dynamic matrix control based on multiple operating models, Journal of Process Control 13 (2003) 4156.