Model predictive control of a distillation unit
In a coordinated plant process, particular attention must be paid to control of individual columns.
OTMAR LORENZ, BERND-MARKUS PFEIFFER, MICHAEL SCHÜLER and UTE FORSTNER
Viewed : 3579
Phenol is an important industrial chemical and is used as an intermediate especially for the manufacturing of various plastics such as polycarbonate, epoxy resin, and nylon. The phenol distillation unit separates the main products phenol and acetone, as well as cumene, alpha-methylstyrene, water, low boilers, and high boilers. Several distillation columns at the plant are involved in an interactive process and are closely linked to each other by heat flow and mass flow. The operation of several central columns determines the mode of operation of the entire plant. One of these central columns is the column for separating phenol from low boiling components. This is the column where the new distributed control system (DCS) embedded model predictive control (MPC) function block is now used.
The production of phenol is a precisely coordinated process. Particular attention must be paid to control of the individual columns. Not only yield and quality of the end product depend on this, but also the energy efficiency of the overall process. During the migration of the plant to the process control system Simatic PCS 7, the operating team decided to use the embedded MPC10x10 function block for this control task.
In recent years, the use of MPC has significantly increased due to the potential benefits for plant operating companies. For example, an advanced process control (APC) solution enables throughput to be increased, reliability to be improved, and the use of raw materials to be optimised. This trend benefits from the fact that DCS controller hardware becomes more powerful and inexpensive. Many plants up to now use separate APC software packages that must be linked to the DCS in order to operate MPC. However, DCS such as Simatic PCS 7 are now available that include an embedded MPC function block. This offers a lot of advantages, which can be demonstrated using the example of the successful application by Ineos at its Gladbeck site. Usability has been significantly improved, which means that MPC is no longer something for experts alone.
Ineos Phenol is the world’s largest supplier of phenol and acetone and operates a plant in Gladbeck, Germany, that produces 650000 t/y of phenol.
The distillation in Gladbeck separates the main products of phenol and acetone, as well as cumene, alpha-methylstyrene, water, low boilers, and high boilers. Several distillation columns at the plant are involved in an interactive process and are closely linked to one another by thermal and mass flows. The mode of operation of several central columns determines the mode of operation of the entire plant. One of these central columns is the column for separating phenol from lighter boiling components, in which the new MPC10x10 function block is now used (see Figure 1).
Phenol production is a continuous process with several distillation columns at the end, used for the separation and purification of the phenol product. A sophisticated control concept is necessary for the charge materials of the columns – water, electricity, steam and so on – as these inputs also affect other columns and provide an important leverage for raising the efficiency of the plant.
Due to difficult process dynamics with long dead times, various interactions and external influencing factors, this control task is extremely complex and demanding. After a setpoint step, for example, it may take several hours before the column reaches a new steady state equilibrium. Without an MPC controller, however, plant operators have difficulties in conforming to the required product quality specifications with regard to phenol concentration.
In contrast to phenol concentration, the column bottom temperature reacts significantly more quickly and is easier to control. At Ineos, therefore, three manipulated variables – vapour flow, reflux and water feed – are used for the coordinated control of phenol concentration at the column head and temperature at the column bottom. Additional feed-forward disturbance compensation variables are calculated from other measurable variables in order to improve prediction quality. For the plant operator, this is a matter not only of stabilising production but also of optimising resource consumption. For example, the consumption of heating steam can be minimised while maintaining the boundary conditions for product quality.
The dynamic process model for the embedded MPC10x10 function block was derived from the existing external MPC, running on a separate PC linked via OPC to the DCS. For this purpose, artificial learning data is generated by simulation of the existing model. This data is read into the MPC configurator (which is part of the DCS engineering station) instead of measured data from the real plant. The MPC configurator uses this data to generate the required process model. In this way, a very high degree of agreement (93…96%) is achieved between the new and the existing process models.
The specification of the controller targets is also derived from the external MPC. In contrast to the previously used MPC software package, the embedded MPC10x10 uses only a single target function, in which all targets are weighted and added together. The controller targets that can be used in the MPC10x10 are:
• Controlled variable setpoints and zones
• MV move penalties
• MV target values.
The specification for controller design determines that the first controlled variable (phenol concentration) is more important than the second (temperature, see Figure 2). The MV move penalties that are suggested as default values for robust controller tuning by the MPC configurator are used as they are.
The MV target values also have different weightings: due to the associated energy costs, MV1 (steam flow) is given the highest weighting. In total, the weighting of the MV target values is considerably smaller than the weighting of the controlled variables so that the MPC clearly prioritises compliance with the control targets. Only when both controlled variables are within the specified setpoint ranges (dead zones) the controller slowly attempts to approach the MV target values (see Figure 3).
The control performance of the embedded MPC can be evaluated from the trend of manipulated and controlled variables for a period of 10 days, which includes a load variation after about three days. The two controlled variables – phenol concentration and temperature – are kept within tight limits of tolerance. The MV targets are adjusted at the load variation, the highest priority being given to the steam target value with regard to the energy efficiency of the plant.
For more than eight years, the column has been controlled using a multivariable control software package. Although this software was working properly, there have been crucial disadvantages: the algorithm of this controller required so much computing power that a separate PC had to be connected to the DCS. In addition, the software was still running in Windows XP. When the process control system was due to be migrated to a new DCS version during a plant shutdown, it was clear that a new solution had to be found. The software package of the existing external controller could only have been migrated at considerable costs. By contrast, the MPC10x10 function block belongs to the standard scope of delivery of Simatic PCS 7 since V8.1.
Add your rating:
Current Rating: 3