To produce pulp (for paper production or as market pulp) is a complicated process with many integrated stages that impact the final quality. The raw material, i.e. pulp logs and wood chips from saw mills, differ in quality and properties due to species, age classes, moisture content, storage time, etc. This article will focus on the bleaching plant, which is at the end of the pulping process and thus is the last opportunity to ensure that the quality of the pulp is correct. The main aim of the bleaching process is to increase pulp brightness by removing or modifying the light absorbing substances that remain in the pulp after cooking and washing. Chemical treatments are applied in sequential stages to achieve the right brightness while maintaining the pulp strength.

Maintaining stable pulp quality is a difficult task and usually requires manual control by experienced operators. Even though the operators are very skilled, they tend to add more chemicals than they need while overcompensating for earlier decisions. This often leads to oscillating behavior, which in turn increases chemical costs and environmental impact.

Competition on the world pulp market is fierce. It is therefore crucial that the pulp production is cost-effective. The chemical cost of bleaching is normally between 17 and 22 euros per ton. In general, this cost can be marginally lowered by controlling each step in the best possible fashion. However, most implemented process control systems can't effectively reduce the marginal cost due to the complexity of the process. The experience from the operations research-based system described in this article shows a saving potential of 10 percent compared to manual control.

History Behind the Project


The online process control system OptCab was developed in collaboration between Linköping University, Billerud's paper mill Skärblacka and the software company Eurocon AB. OptCab is a system that enables online response to determine chemical charges in the bleaching process. The story of OptCab began in 2000 when Staffan Jacobson, a process control engineer at Skärblacka, initiated a pre-study together with Patrik Flisberg and Mikael Rönnqvist to review possibilities to improve the bleaching control.

A first research prototype was developed in 2001. Process controllers P.O. Amren, Bertil Lindström and Stefan Nilsson provided critical information regarding chemical processes and practical aspects of the process. The prototype was based on a combination of user-built procedures to develop approximations and standard modules for modeling and optimization. It was only used to test data offline. In 2002, disadvantages with the first prototype were identified, and further development of modules took place. The first, brief online test was held in May 2002.

Supported by development manager Margareta Öhrn and production manager Curt Johansson, the project continued, and more online testing was carried out between August 2002 and January 2003. In 2003, the system — still a prototype with only small possibilities for operators/process control staff to change parameter settings — was continuously tested and upgraded. From the end of 2003 and through the first half of 2004, the system was completely redesigned. At that point, the company Eurocon AB took overall responsibility for the project and designed a user-interface while all of the O.R. modules were redeveloped. The last version of OptCab was installed and put in operation at the end of 2004.

Process Control Operations in Action


Billerud's paper mill in Skärblacka has three pulp lines, four paper machines and a pulp drier. The mill has 700 employees and produces sack and kraft paper with an annual production capacity of 250,000 tons. Furthermore, it produces fluting and market pulp with an annual production capacity of 90,000 tons and 70,000 tons, respectively.

The pulp production at Skärblacka starts by cooking and washing the wood chips, followed by a pre-bleaching step and a second washing. The bleaching process (see Figure 1) is the last step in the pulp production and is therefore significant to secure the pulp quality. As mentioned earlier, the main goal with the bleaching steps is to increase the brightness of the pulp by removing or modifying the lignin that still remains in the pulp after cooking and washing. Chemical treatments are applied in sequential steps to reach the right brightness while the strength of the pulp is kept as unaffected as possible.

Figure 1: The four main steps (with added type of chemical) in the bleaching process.

Long response times are a well-known problem at paper mills; the bleaching process is no exception. It takes about nine hours to complete the bleaching. At each step, the pulp slowly moves through large tanks while the chemical reactions take place. The time between a chemical charge and the measured result can be up to five hours. Many properties such as brightness, pH value, kappa number, temperature, flow speed and concentration are measured before and after each step. Some properties are only measured every 30 minutes, which, together with potential sensor errors, makes information quality-sensitive. There are limits on brightness after each step in the process. The operator can manually choose chemical charges, and this requires experienced operators. Also, because of the long time span of the process, decisions made by one operator may impact late in the next operator's shift.

O.R. Methodology


The main idea with OptCab is to combine flexible process models with robust optimization. The solution approach is first to determine approximate functions for the process models (describing behavior of bleaching steps). This is used in a second step where the overall control problem is formulated and solved: how much chemicals should be added in each step for the pulp (plug) entering the process.

A process model (one for each bleaching step) is determined dynamically; i.e., it is continuously updated due to varying production and pulp quality in each bleaching step. For each bleaching step, input, output and control (chemical charges) variables are defined. An approximate function describes the relationship between each output and all input and control variables. The structure of the approximate functions can be based on theoretical and empirical results for the bleaching process. In order to determine the process model, a set of parameters in the approximate functions must be computed. This is an identification problem, and we formulate it as a least-square optimization problem with the parameters as decision variables. Extra constraints on practical restrictions are added to the model.

The process model is changed dynamically, and a typical update interval is 15 minutes. The data is based on measurements over the last week, with higher priority given to the most recent measurements. An important tool for the identification is a plug flow model. This is needed in order to relate correct input with the right time delayed output (some stages take several hours to complete). It is based on pulp tracking, assuming a pure plug flow behavior to account for varying delays in the different stages of the process as a result of changes in production rate, tower levels and pulp concentration.

With the information gained in the process models, a control problem that integrates all bleaching steps is formulated. In this model we use the chemical charges as control variables and input/output properties as unknown state variables. As constraints we use:

• the approximate functions (defined in the process models),
  
  
• linking constraints for input/output variables (between steps),
  
  
• lower and upper bounds on control variables, and
  
  
• lower and upper bounds on state variables.

Beside these constraints, there are soft constraints describing that the difference in charges between consecutive iterations should be small. The objective is to minimize the cost of chemicals. The bound constraints on, for example, brightness may be too limiting, and no feasible solution can be found. Since robustness is critical, we have modified the objective by adding penalized slack variables for such constraints. The same approach is also done for the soft constraints. The control problem is a general non-linear (and non-convex) problem that is solved every five minutes. The size is relatively small, and solutions are found within a number of seconds.

When the pulp reaches the bleaching process the control problem is solved, generating control values for the pulp throughout the bleach process (all four steps). However, the approximated functions are not exact, and there may be changes in the process that were not known when the control problem was solved. For example, the production rate might change, the actual chemical charges may differ compared to the proposed optimal charges, the temperature in a tower may vary, etc. Therefore, when the pulp (pulp plug) reaches each of the other steps the control problem is solved again, this time with fixed measured values for previous steps, and possibly more accurate data for the remainder of the process. The support tool thus becomes more robust and stable since the new control will adjust to any changes that may happen in the process. The implication of this is that four control problems are solved every five minutes, each providing a control for a single bleaching step.

The OptCab System


The structure of the OptCab system is illustrated in Figure 2. A user-interface for the operators is used to change parameters such as limits and targets on brightness and limits on chemical charges. A distributed control system (DCS) is used to handle all data communication. Separate modules handle the different steps in the solution approach: plug flow model, identification problem and control problem.

Figure 2: Overview of the main components in the OptCab system.

The data and overall program handling is developed in Visual Basic, and the optimization routines are based on open source codes. When designing the modules, three practical aspects were important.

First, we wanted a robust model that always generates solutions that are feasible (or near feasible), don't oscillate and are not too different from those produced by manual control. The latter is extremely important for the operators. If an operator finds the solution to be very different to the way he or she would have controlled the process, particularly in the first tests, he or she might mistrust the solution and override it. With the new user interface, operators have the flexibility to change parameters so it becomes more cost-effective as they become more confident with the system.

The importance of a robust model soon became obvious in the early testing. A sensor for a chemical charge had been calibrated during a stop, resulting in inaccurate approximate functions since the identification problem was solved with historical data from before the calibration. A major change in the value level of the sensor after the calibration left an excessive chemical charge that was not detected until an alarm went off indicating dangerous levels of gas in the plant. Since then, maximum values of some residual levels are included in the program.

Second, we wanted to include practical constraints in the identification problem. These constraints are based on empirical and theoretical results. One example is that the brightness will not decrease with increased chemical charge. This property can be expressed by restrictions on the derivative between brightness and chemical charge. Figure 3 provides an example describing the high quality of the approximations.

Figure 3: Comparison between real and approximate brightness during a period of one week.

Third, we wanted to pre-process the data before each identification problem. Some data is faulty and is removed due to incorrect sensor measurements. Some errors are easy to identify because the data is, for example, outside the sensor's operating area. Others can be removed by identifying data that is too far away from the approximations. This is done by solving the identification problem twice. Each identification problem is relatively small and is solved within a number of seconds.

Experiences and Results


OptCab has been in operation for about two years with a practical usage level of about 98 percent. The remaining proportion relates to planned maintenance and other operators' controlled overrides. It is difficult to measure the actual cost savings, as there is a continuous development in other parts of the process as well. In order to measure and analyze the direct impact, the system was turned off and the operators were running the manual system during three weeks. OptCab was running in parallel during the manual period. The chemical costs under OptCab were 10 percent lower than the costs suggested by the operators (Figure 4). Instances when OptCab has a higher value than the manual system relates to situations when the manual solution failed to meet the constraint on brightness. Other measurements and experiences have validated this number.

Figure 4: Comparison of relative cost (vertical axis) between manual operation (yellow line) and automatic (black line) during three weeks.

Introducing an automatic black-box system is difficult. In the manual system operators controlled the process, and there was a tendency to use a safe margin against, for example, the final brightness. This margin is, of course, very individual. OptCab is an optimization-based system and will find solutions that are close or on the bounds. In situations when the properties of the pulp change quickly, OptCab reacts instantly and makes changes in the control. In the beginning when the system was new, some operators became cautious, turned off the system and used manual control. After some time, the operators became more confident with the system and let it work online without any interference. Feedback from operators has been critical in understanding the process and developing the user-interface.

OptCab is designed to minimize the chemical costs subject to a set of restrictions. However, there are other positive effects. One is that the final brightness is more stable and follows more precisely the target value. This is important for the paper quality in later stages. A second is that the operators now can spend time on other issues instead of continuously monitoring the process control. A third aspect is the improved environmental impact by lower chemical usage.

Quality in data is critical for this type of automatic system. At Skärblacka we have been fortunate to have access to high-quality sensors for pulp brightness — the most important property measured. We have noticed that OptCab can be used offline to identify faulty sensors. Currently, we are developing an automatic module to monitor all sensors in order to set off warnings if they are likely to be faulty. Further work also includes the development of a general modeling tool to configure OptCab for other bleaching lines.



Patrik Flisberg and Mikael Rönnqvist work at the Division of Optimization, Linköping University. Stefan Nilsson works with process development at Billerud Skärblacka AB. The authors acknowledge the contributions of the project team and operators at Skärblacka and the company Eurocon AB with their work in the last year to develop OptCab into a full industrial system. Further information about OptCab can be found at www.eurocon.se.

Table of Contents

OR/MS Today Home Page


OR/MS Today copyright © 2005 by the Institute for Operations Research and the Management Sciences. All rights reserved.


Lionheart Publishing, Inc.
506 Roswell Rd., Suite 220, Marietta, GA 30060 USA
Phone: 770-431-0867 | Fax: 770-432-6969
E-mail: lpi@lionhrtpub.com
URL: http://www.lionhrtpub.com


Web Site © Copyright 2005 by Lionheart Publishing, Inc. All rights reserved.