ABSTRACT

The main process objective of lime kiln operation is to produce a uniform quality lime. Additional operating objectives include minimizing fuel consumption and complying with environmental regulations. Given the long process delays and variable interaction inherent to lime kilns, these objectives are extremely difficult to achieve under manual or standard closed-loop kiln operation. A predictive control algorithm, MACSTM (Multivariable Advanced Control System) has been successfully applied to a Scandinavian kiln. A significant reduction in process variability has been achieved while simultaneously extending the operating life of the refractory. In addition to the kiln temperature profile, flue gas oxygen and CO are also explicitly regulated by the controller. By sustaining stable kiln operation at the process limits, the advanced controller is able to minimize energy use while ensuring safe combustion conditions are maintained. _____________________________________________________________________________________

BACKGROUND

The advanced control project was conducted at Stora Cell’s bleached kraft mill in Skutskär, Sweden. The kiln was delivered by FL Smidth in 1976 and has dimensions of 104 meters by 3.3 meters with a nominal production rate of 250 TPD. The kiln has recently been retrofitted with a low pressure burner, new chain section and coolers, as well as an ABB Advant DCS.

The kiln is operated over a wide range of production rates, and the lime mud feed is typically cut off for 15 minutes every two hours while the mud filters are cleaned. These continuous process disturbances combined with intermittent wet scrubber gas combustion resulted in frequent refractory damage from overheating. In addition, wide variations in the kiln temperature profile led to significant lime quality variation. The control project was primarily justified on the basis of eliminating refractory damage and the costs associated with unscheduled downtime and maintenance. Improved caustic area operations and fuel savings were additional economic incentives.

CONTROLLER DESIGN

Capstone Technology's MACS software is a model based, predictive controller. At its core, the technology utilizes dynamic models and optimization to predict and control future process behavior. The process dynamics in the optimization are described by an explicit process model derived from plant operating data. Consequently, the control software is general and can be readily applied to a variety of processes.

The first step in developing the advanced kiln controller was to perform a series of process bump tests, from which the dynamic model matrix, shown in Figure 1, was developed. Using this process matrix, the advanced control software was configured to maintain front-end temperature to an operator specified target, control back-end temperature to an artificial setpoint ten degrees less than the current process value and to use stack oxygen as a low constraint. The stack CO analysis is used as a supplement to the oxygen measurement to ensure proper combustion. The result of this configuration is a controller that adjusts fuel flow and ID fan speed in a coordinated manner to hold the front-end temperature constant while driving the back-end temperature ever lower. Naturally, as the back-end temperature is decreased, the flue gas oxygen content is also forced lower. Eventually, the low oxygen limit is reached, at which point the optimizing controller halts the back-end temperature minimization. The result, uniformly high quality lime produced with minimum energy input.

MILL RESULTS

The advanced control performance improvements are illustrated in Figures 2-4. On each graph, the vertical bars represent closed-loop control over a four month period. In contrast, the solid areas give reference data from the two months prior to the installation of the advanced controller.

The first graph shows variability of front-end temperature around the operating target. The y-axis has been normalized to indicate the fraction of temperature measurements which fall into a particular range. For example, the two tallest bars on the front-end temperature graph indicate that roughly 62% of the temperature observations were within a [-5, 5]oC range around the setpoint. Overall, front-end temperature variability has been reduced by more than 90%. Due to the strong correlation between front-end temperature and lime residual carbonate, the kiln is now capable of producing a uniformly higher quality lime.

Figure 3 illustrates the performance improvements for back-end temperature plotted as variability around the mean. As shown, the advanced controller has achieved a 50% reduction in variability as compared to the reference data. The control improvements on back-end temperature help reduce atmospheric heat losses as well as lessen operational problems such as ring formation.

Finally, closed-loop oxygen data is compared to the reference period in Figure 4. It is apparent that the advanced controller has achieved a reduction in the average flue gas oxygen content. Furthermore, the low oxygen constraint has prevented kiln operation with insufficient oxygen. As a result, significant energy savings are possible while simultaneously observing environmental restrictions.

The advanced kiln controller has now been in operation for one year with an uptime of 95%. A key benefit has been increased process availability resulting from the elimination of unplanned refractory maintenance. Overall economic benefits have provided a project payback of six months.

MODEL PREDICTIVE CONTROL

Model predictive control is a multivariable optimal control formulation which can handle hard constraints on manipulated and controlled variables1, 2. The controller uses a set of linear dynamic models representing the process to predict the effect of future control moves on the output variables (controlled and constraint). An optimization routine is used to compute a set of future control moves such that a performance index is minimized while simultaneously ensuring that the process constraints are satisfied. The set of control moves is calculated at each control interval, but only the first calculated move is implemented. The entire process is repeated at each subsequent control interval. Process measurements are used for feedback to compensate for unmeasured process disturbances.

The predictive control technique is shown graphically in Figure 5. At each control interval, the algorithm computes a series of m future control actions. This interval, t+m, is referred to as the control horizon. The dynamic process models are used to calculate the predicted process response over the prediction horizon of t+k control intervals. The optimization is carried out such that the predicted error for all output variables is minimized across the entire prediction horizon.

Figure 5: Model Predictive Control

A control interval of two minutes was selected with a prediction horizon of 120 minutes. The controller is configured with a ten interval (20 minute) control horizon, and manipulated variable changes are allowed at intervals 1, 2, 3, 5, and 10. Consequently, the control law at Skutskär can be expressed as the quadratic optimization:

where TFET, TBET, O2, and CO refer to frontend temperature, backend temperature, oxygen, and carbon monoxide, respectively. The controller is tuned by adjusting the individual weighting coefficients, alpha and beta. Each controller variable has a weighting coefficient. As a particular alpha is increased, the penalty for moving the associated manipulated variable will be greater. Similarly, as a particular beta is increased, the controller will try harder to move the controlled variable to target since this will minimize the performance index J.

Additional optimization constraints consist of the linear dynamic process models. These models are represented graphically in Table 1. Furthermore, the optimization is bounded by hard constraints on the manipulated variable magnitude and rate of change. Note as well that the oxygen and CO criterion are active only if the constraint limits are violated.

CONCLUSION

The model predictive controller proved to be very effective at stabilizing this lime kiln and the results speak for themselves. The kiln is now able to operate in a stable, economic manner and has achieved one year of operation without refractory loss. The ability to rapidly configure the controller allowed the mill to achieve and maintain high product quality in a matter of weeks. And last but not least, mill personnel including kiln operators, instrument technicians and engineers have a better understanding of kiln process interactions and a new standard for kiln performance.

ACKNOWLEDGEMENT

The authors thank Kaj Bäckman, Stora Cell AB, and Hans Lindberg, Stora Corporate Research, for originally sponsoring this work.

REFERENCES

1. Prett, D. M. and Garcia, C. E., Fundamental Process Control, Butterworth-Heinemann, 1998.

2. Sooterboek, R., Predictive Control - A Unified Approach, Prentice-Hall, 1991

• Chemicals, Petro Products, Plastics