2.2 Realization at a polymerization plant
2.2.1 Implementation of the proposed methodology
The implementation of the integrated process methodology consists of a pro-cess DW collected from the operating polymerization propro-cess, the correspond-ing rule-based knowledge discovery tools and the process simulator with its graphical interface.
Process Data Warehouse
In the polypropylene technology, the DCS implemented by Honeywell has the function for storing the large amount of process data through its Process His-tory Database (PHD) module. The structure of this module can be seen in Fig. 2.2.
Application programs PHD API
Copy of config data
Reference database PHD SERVER
Real-time database
Current value database Raw queue Data Queue
PHD disk archives Real-time
Data Interface (RDI) Real-time system
&
&&
& DCS
Figure 2.2: Structure of Process History Database (PHD) module.
There are two main operations:
- Data collection: Data originate from real-time system and are collected by a real-time data interface (RDI). Tag parameters for all the variables are stored in a reference database. A tag contains all important informa-tion about a process variable (name, type, unit, etc.). RDI sends data to the PHD server, which places the collected data for a tag in the raw data queue and applies data processing (smoothing, compression, etc.) to move raw data queue entries to the data queue of the tag. Data queue of the tag then holds processed data that is ready for insertion into the active logical archive les using the continuous store thread.
- Data retrieval: An application program makes a call to the PHD appli-cation programming interface (API) indicating the desired tag and time
range for data. The system checks the data queues to see if the data is still held in the queues, otherwise PHD accesses the data from the connected archive les.
Data ow goes as follows: First, the tag names of the relevant process variables are selected from all the possible tags in the plant. Process data belonging to the selected tags are accessed in PHD by the Uniformance Desktop application program (by Honeywell). While the Uniformance runs as an MS Excel add-in, the results of data queries are saved in Excel les.
The PHD system has the following features from process analysis point of view:
- It automatically saves and stores every process variable, its operational value, set point, condence, etc. that is generated in the process;
- It is extended only by basic visualization applications in order to help process engineers daily work routine;
- Because of limited storage capacities for that enormously large data, only last 6 months are available, previous data are saved on data discs and stored separately;
- Process boundaries of process variables are not accessible;
- Set points of product-types are hard to collect one-by-one;
- O-line laboratory measurements are manually uploaded.
Concluding the above statements, this structure is inapplicable for process analysis (data mining, model building, simulation) purposes.
The implemented data warehouse collects all the relevant information for these statistical and modeling applications. It was implemented in MySQL°c Database Server and contains all the major variables of production with a sample time of 15 seconds (which is the original sampling interval of the control system) of one year continuous production. In the selection procedure of major variables, plant engineers and operator personnel were involved as well in order to collect all information in DW, which has relevant contribution for further analysis purposes. In its prototype form, the implemented data warehouse consists of the following data tables:
- Measured data: selected variables of the technology measured and stored by the DCS;
- Boundaries of the variables: H, L (high and low alarm) HH, LL (high and low interlock limit) values for process variables;
- Production parameters of products: MI and operational set points, cat-alyst recipes;
- Information about Powder and Extruder silo changes: reasons and dates of changing the storage drums that have the information of product tran-sitions;
- O-line laboratory MI measurements for polymer powder and granulate with various sampling frequencies;
- Applied catalyst system.
Data Warehouse is reached via ODBC driver through MATLAB. Also, every tool and interface was developed in MATLAB and MATLAB Simulink. The models integrated to DW and detailed in the following subsections take in-put data source from the process DW. Their main simulated outin-puts can be evaluated as well by comparison to real process data stored in the DW.
Process model
The technology detailed in Section A.1.1 can be divided into several operational parts, which can be modeled separately, and the whole system can be composed by dening the connections between these individual elements:
- Catalyst blending subsystem
- Bulk polymerization subsystem (loop reactor section)
- Impact polymerization subsystem (gas phase reactor section) - Separation subsystem
- Extrusion subsystem
- Polymer grade storage subsystem
The main element of the whole technology that denes the overall per-formance is the loop reaction section. That is why the experimental tool is developed only for this section. All other parts have much less engineering inuence on product outcome, e.g. catalyst preparation and additive blend-ing in the extrusion section are based on strict recipes, other subsystems as separation and storage have no eect on product quality by normal operation.
Hence, the bulk polymerization subsystem was modeled and implemented (for the control model, the recipes of the catalyst blending were also involved) by semi-mechanistic (hybrid) models of the three loop reactors for the control model where the recipes of the catalyst blending were also involved. Where rst principle models were not available because of condentiality or unavail-ability, black box models were identied based on industrial data collected from the process data warehouse.
The main advantage of using this modeling technique is that in hybrid models, all the a priori knowledge can be built into the model, so the vari-ables have their physical meaning, hence are easily understandable for human observers like process operators. Moreover the scale of reaction kinetics did not need to be characterized experimentally because it can be included in the model as a black box element.
All the three loop reactors' models are identical, the only dierence stands in the parameters, so a general model is explained here. By type, the loop reac-tor is between the PFR (plug ow reacreac-tor) and the CSTR (continuous stirred tank reactor) reactors, but from modeling point of view it can be considered as a CSTR model. In Fig. 2.3 the response of three types of models can be seen:
for a given amount of injected indicator material 'A' at time t0. As the speed of cycling increases the output of the loop reactor gets more and more similar to the CSTR. This means that for suciently large cycling velocity (innite recycling), the loop reactor's response will be identical to the CSTR, hence the variables will be only time-dependent instead of time and location-dependent.
The detailed weight-based model of a loop reactor can be found in Section A.1.2.
Catalyst productivity calculation and accumulation rate calculations are based on parametric phenomenological equations. This is a simplication of the model in order not to have too many process parameters that need to be identied. Also considering that reaction kinetics are not fully known, this is a sucient solution for the information level of operation.
t0 cA
t0 t0
cA cA
time time time
Figure 2.3: Output response of PFR, loop reactor and CSTR models for ma-terial 'A' injection at time t0, repsectively.
The model was implemented in MATLAB Simulink. The standalone pro-cess model is a connection of three reactor models and has 17 input data taken from the Data Warehouse and simulates the production while calculates 60 outputs. Figures of Simulink screenshots can be seen in Section A.1.3 where the identied parameters of the models are listed as well (Table A.1). The pa-rameter values were tuned in order to t the model output to most productions i.e. the model corresponds to average production.
Control model
At the plant, the cornerstone of process control is Honeywell's Prot°c Con-troller, a multivariable control and optimization application for complex and highly interactive industrial processes. It is based on RMPCT°c - Robust Multi-variable Predictive Control Technology - of Honeywell Inc., which is a hierarchic special distributed control system with layered optimization of multi input - multi output systems. At process level, HPM°c's - HighPerformance Process Managers - are responsible for basic control through PID controllers.
The set points of these controllers are given by the Advanced Process Control System that co-operates with RMPCT's model based control system.
In my work, this APC performance calculation module has been imple-mented. Beyond setting the set points of the local control loops, based on the inlet feed ow rates and the specic temperatures, APC calculates the performance of the reactors, the conversion rates, the concentration of output streams. It sends the calculated process values necessary for local control and provides variables not measured directly (Fig. 2.4). The backbone of APC is the heat balance calculation of the bulk polymerization reactors and the gas phase reactor, which are the ground for all the calculated mass ow rates and concentrations in the technology.
Process Advanced Process
Control System
Local Control System
Figure 2.4: Connection of the process, local and advanced process control.
Summarizing, regulatory control is based on the performance calculations of this APC system that is responsible for steady state operation and helping the operators during process transitions. This later task has high priority because of the frequent grade transitions (large product portfolio). From this fact comes the need for an exact implementation of the local control connected to the APC performance calculation system.
All calculations are based on conservation of mass, momentum or energy.
Typically, momentum balances do not provide useful control information, and are ignored. The basic conservation law was employed:
Accumulation = Input + Output + Generation.
For energy, in a general form:
4Esystem =4U +4Ekinetic+4Epotential =±Q±W, (2.1) where E stands for energy, U for internal energy, Q for heat added to or removed from the system and W for work done on or removed from system.
Ignoring potential and kinetic energy, and assuming that internal energy can be represented by enthalpy, the energy balance for polypropylene produc-tion looks like as follows:
mcPdT
dt =Qf eed+Qamb+Qjw−Wpump+Qrxn. (2.2) Rearranging of Eq. 2.2 and substituting Qrxn =P R·∆Hrxn, one gets the production rate of propylene polymerization (PR):
P R= (−mcPdTdt −Qf eed−Qamb−Qjw+Wpump)
∆Hrxn . (2.3)
The performance calculations of APC are made in strict, logical order:
- Feed calculations (mass and energy balance inputs) - Generation terms (production rate)
- Solids and Euent calculations (mass and energy balance outputs) - Concentration calculations (internal calculations)
- Misc. calculations (e.g. heat transfer coecient)
Results of APC calculations are sent to local control level, where PID algorithms and ratio control are implemented. With the assumptions that in cascade controllers the slave controllers work perfectly (ideal follow-up control without time delay, thus no PID identication needed), the following local control loops were implemented and tuned (Fig. 2.5):
- Production rate control by catalyst inlet ow rate. It is the master PI controller of the catalyst feed to the rst reactor slave controller. Set point of this slave controller is dened in 'g/h', which is output of the master controller. This output is controlled by production rate error ('t/h') derived from: (i) 60% of total production set point and (ii) the production rate of the rst loop reactor calculated by APC;
- Reactor density control by propylene monomer inlet ow rate: master PI controller, which calculates the set point deviation from the density calculated by the process model (APC originally measures densities) and manipulates the set points of the monomer feed slave controllers in 'kg/h';
- Productivity control by hydrogen inlet concentration: master ratio troller (not PID), based on the monomer feed and hydrogen feed con-centration set points. It manipulates the set point of the slave hydrogen feed controllers in 'kg/h';
- Reactor temperature control: master PI controller, which calculates the set point deviation from the reactor temperatures calculated by the pro-cess model (APC originally measures temperatures). It manipulates the set points of the cooling water inlet temperature slave controllers in 'oC';
- Catalyst composition ratio control (not PID) by inlet ow rates of cat-alyst components: master ratio controllers, based on the monomer feed and necessary TEAL and Donor concentration set points. It manipulates the set point of the slave TEAL and Donor feed controllers in 'kg/h';
PP F
R200
R201 R202
CAT Don.
TEAL
PP PP
F F
CWR CWS CWR
CWS
Prop.
Hydr.
Prop.
TC TC
TC TC
FC
DC DC
FC
FC
FC Hydr.
Figure 2.5: Structure of local control.
The above control loops are in connection with each other. Plant operators can intervene through every point of the loop reaction process, and grade transitions are managed by them as well. In practice, they manually tune the set point of the following local control loops:
- Hydrogen feed concentration.
- Reactor temperature.
- Catalyst feed rate.
These facts show that product grade transitions could be optimized as well, while product control could mean product quality control if there were an ad-equate on-line model for that. Currently, operators' work is based on o-line laboratory measurements, so eects can be analyzed only hours after interven-tions, hence as a product model, semi-mechanistic MI soft-sensor model was needed. Its implementation can be found in Section 2.2.1.
The control model consists of two main parts: (i) the APC performance calculation model and (ii) the local control model (see Figure 2.4). Without
Controller Parameters Controller Parameters CAT feed K = 4 R200 reactor temperature K = 15
TI = 9 TI = 10
R201 slurry density K = 3 R201 reactor temperature K = 15
TI = 5 TI = 10
R202 slurry density K = 3 R202 reactor temperature K = 15
TI = 5 TI = 10
Table 2.1: Parameters of the local PI controllers.
attaching the local control system, the model is applicable to re-simulate pro-ductions with inputs taken from the data warehouse or simulate the eect of e.g. a control error.
For the local control model, PI controllers were identied. Their parame-ters can be found in Table 2.1. As the concept of local control at this point was to stabilize the model system and there were no information about ex-act structure of local controllers, a simple identication algorithm, the direct synthesis method [16], was applied to identify PI controllers and its resulting parameters were netuned afterwards.
Product (inferential) model
The product model of the integrated model system depends also on the ap-plication. It can be structured as any of the model types (white, black, semi-mechanistic, linguistic) depending on our priori knowledge and the necessary information level of the model, and can cover any product related attribute.
Here, a semi-mechanistic extension is presented, which is able to predict the melt index based on process variables of the technology, which is demonstrated in Section 2.3.3.
MI is measured o-line in the laboratory. Its dependence on process vari-ables is dicult to obtain by rst principle models because of the complex un-derlying mechanism [155]. In such a situation, semi-mechanistic models gave good results [156, 157, 158] with estimating the unknown parts of rst princi-ple models by black-box techniques. A comparison of mechanistic, empirical and neural network models can be found in [159] for MI estimation in a HDPE plant where a collection of property estimation papers are cited with dierent approaches for dierent polymerization technologies. In polypropylene quality
estimation only empirical, black-box or simplied modeling approaches can be found in the literature [160, 161, 162]. In the case under study, the product model uses a semi-mechanistic modeling approach and can be considered as a polypropylene MI soft-sensor.
Fig. 2.6 shows the inner sight of the process model in Fig. 2.1. It gives an example how the original model can be extended for product related modeling purposes by hybrid modeling. If product quality control needs the estimation of a non-measurable variable, a soft-sensor application has to be developed.
Soft-sensors are on-line hardware measurements with attached software algo-rithms that calculate non-measurable information about the technology based on either rst principle or black box or their combination as semi-mechanistic modeling [156, 157, 158]. The semi-mechanistic Process model (RSM) is ex-tended by a black-box regression model to estimate the instantaneous product attribute value (BBM) and by a rst principle model of mixing (FPM) to simu-late the product attribute dynamics. These two components build the product model. The input and output variables are denoted as U and Y as the stored and simulated PV's, while xi means the ith state variable (i[1m]).
RSM
BBM FPM
Output
x1
Y x2
xi
U
xm-1
xm
Model of reactor system
Figure 2.6: mechanistic process model (RSM) extension by a semi-mechanistic product model.
The semi mechanistic model for product quality estimation was also im-plemented in MATLAB Simulink°c environment. The measured input of the soft-sensor are the input variables of the APC calculation model. The other two elements of the software part of the sensor are a black box model for esti-mating instantaneously produced MI and the model of mixing the current MI with the cumulative MI.
For the black-box modeling part, Self Organizing Map was applied to select the appropriate variables that have eect on Melt Index [163]. The SOM facilitates visual understanding of processes so that several variables and their
interactions may be inspected simultaneously. Once a SOM model is trained by transition-free data, its regression potential can be exploited in order to predict product quality from state variables. Therefore, it can work as part of an online product quality estimator because it calculates the instantaneously produced polymer MI based on the selected state variables of the system. More details on the SOM algorithm can be found in A.1.4
To model the polymer MI that leaves the ith reactor of the cascade, the following dierential equation of mixing was applied:
d(mP P,iMIiζ)
dt =RoutP P,i−1MIi−1ζ −RoutP P,iMIiζ +P Rcurr,iMIcurr,iζ (2.4) In Eq. 2.4, ζ is the coecient of non-linear mixing with a value of −0.296 [155]. The above mentioned product quality model was attached to the original Process model and applied for MI estimation in Section 2.3.3.