You are on page 1of 4

PROCESS AND PLANT OPTIMIZATION

Originally appeared in: June 2008, pgs 47-52. Used with permission.

SPECIALREPORT

Avoiding plant tests with dynamic simulation


A model-predictive controller based on models derived from first-principles dynamic simulation as an alternative to empirical models identified from plant test data is presented
N. ALSOP, Preemraff, Lysekil, Sweden, and J. M. FERRER, AspenTech, Barcelona, Spain

n recent years, process designs have become increasingly complex, integrated and interactive. Recycles, heat integration and minimum holdups are common design features. While such designs provide for optimal steady-state operation, they present particular challenges for dynamic control. Whereas steady-state simulation based on first principles has underpinned process designs for many years, only recently has dynamic simulation become a mature desktop technology. Typical applications of dynamic simulation are relief studies, operator training, testing DCS configurations, developing operating procedures and more recently, advanced process control.17 This article explores the use of dynamic simulation for developing and analyzing the basic and model-predictive control (MPC) scheme for a propylene/propane (PP) splitter.
Problem definition. The PP splitter, located at the back

PC-414 LC-129
1 10 20 30 40 50 60 70 80 90 100 110 120 130 131 140 150 160 170 180 182

Total reux

LC-119 FC-202 Propylene Minor reux FC-187 MISO Major reux LC-116

FC-185

Feed

FC-186

end of the FCC complex, separates a C3 stream into a top product containing 99.5% pure propylene and a bottom product containing 98.5% pure propane. To achieve the tight product specifications, the PP splitter is 90-m tall, comprises 181 trays and operates at very high reflux ratios. An energy-saving feature is the heat pump that reboils the tower bottoms by condensing hot compressed overhead vapors. A process flow diagram including the basic control scheme, as provided by the process vendor, is shown in Fig. 1. As shown in the figure, bottom level is controlled by the bottom draw rate and compressor surge vessel level is controlled by top product rate. Analyzers are present on both top and bottom product streams. Bottom product quality is controlled automatically using minor reflux whereas top product quality is controlled by operator adjustments to the major reflux. The aim of the work presented in this article was to implement MPC for the control of top and bottom qualities. In addition, the MPC application is to drive the process to an economic optimum while respecting all equipment constraints. The process economics are governed by the recovery of higher-valued propylene at the expense of energy consumption in the compressor. In the case of the PP splitter, reliable plant step test data on which to base an MPC design are very difficult to obtain for several reasons. First, the very long settling times and the frequency of daily disturbances prohibit the unit from ever reaching a true steady state. In addition, the tight quality specifications prohibit

FC-232 Propane

FIG. 1

PP splitter including basic control scheme supplied by process vendor.

steps of sufficient magnitude to achieve a signal-to-noise ratio that is acceptable for model identification. For these reasons, it was decided to develop the MPC model from data generated by a dynamic simulator instead of the actual process.
Process simulation. The starting point for any dynamic

simulation study is a sound steady-state simulation. In this case, a steady-state model of the PP splitter, heat pump and ancillary equipment was available from previous process engineering studies. The steady-state model was calibrated against averaged plant data as illustrated by the temperature profile shown in Fig. 2. Starting from the steady-state model, the dynamic model was constructed by specifying additional engineering details including pressure/flow relationships and equipment dimensions.8 In addition, all basic controllers were included in the model and configured exactly as they are in the plant. The dynamic solver

HYDROCARBON PROCESSING JUNE 2008

SPECIALREPORT

PROCESS AND PLANT OPTIMIZATION


The alternative basic configurations were tuned and tested thoroughly by dynamic simulation. Major reflux was chosen to control bottoms level since it provided superior dynamic performance. It was observed, however, that product flows and qualities responded much faster to mass balance upsets when minor reflux is used for bottoms level control, or when an internal reflux controller was provided.
Step testing and identification. The independent variables for the MPC controller comprise three manipulated variables and three feedforward (FFW) variables as listed in Table 1. A sequence of steps to the six independent variables was configured to execute automatically in the dynamic simulator. Each variable was stepped up and down by the magnitude and duration shown in the table. After each step, the top and bottom qualities were found to stabilize after 23 days as shown in Fig. 4. The total simulated time for all the steps was 32 days, which took about 2.5 real days to run in the PC. The simulated step test data were then imported directly into the identification software package, after which the dynamic model identification task was as for conventional MPC projects. The dependent variables for the MPC controller comprise two quality variables, three equipment load constraints and differential temperature as listed in Table 2. The equipment constraints include tower differential pressure, compressor motor amps and

parameters were configured with the right number of flash calculations per tray residence time and the tray weir height was adjusted to reflect the plant tray holdup.
Controller conditioning. The first candidate MPC structure

considered was that in which both minor and major refluxes are used as manipulated variables (MVs). The primary controlled variables (CVs) are top quality (mole % propane in propylene) and bottom quality (mole % propylene in propane). Using the steady-state simulation described, the 22 matrix of gain relationships between the 2 MVs and 2 CVs was constructed. From a condition analysis of this matrix, it was very quickly concluded that this matrix was very nearly collinear on the basis that both minor and major refluxes act equally on cutpoint. This means that aggressive moves in both refluxes will be made by any such controller utilizing this MV pair in trying to keep top and bottom qualities on specification.
Basic controller configuration. Alternative basic control

configurations were considered to improve matrix conditioning. In the first configuration (Fig. 3), bottom level is cascaded to the major reflux controller. In this case the bottoms flow is made available as the cutpoint MV and minor reflux is the fractionation MV. In the second configuration, bottoms level is cascaded to the minor reflux controller.

28 26 Temperature, C 24 22 20 18 16 14 12
FC-185 Total internal reux LC-129
1 10 20 30 40 50 60 70 80 90 100 110 120 130 131 140 150 160 170 180 182

PC-414

LC-119 FC-202 Propylene

Minor reux Major reux FC-187 FC-186 LC-116

50

100 Tray number

150

200

FIG. 2

Simulated temperature profile and actual process data (diamonds).

Feed

FC-232

Propane

FIG. 3

Revised basic control configuration.

TABLE 1. Independent variables


Description MV1 MV2 MV3 FFW1 FIG. 4 Step test profiles in the dynamic simulator. FFW2 FFW3 Bottom flowrate Minor reflux rate Compressor discharge pressure Feed flowrate Atmospheric temperature Feed temperature Step move 0.1 m3/h 0.3 m3/h 0.2 barg 0.25 m3/h 15C 1C Step duration 4 days 1 day 1 day 5 days 1 day 4 days

HYDROCARBON PROCESSING JUNE 2008

PROCESS AND PLANT OPTIMIZATION


reboiler approach temperature. The differential temperature CV is used to establish a prediction error FFW as explained later. Dynamic models between all dependent and independent variables were identified from the noise-free simulation data. Finite-impulse-response models for various time horizons were observed to overlap, an example of which is illustrated in Fig. 5. As shown the total time to steady state for top quality following a feed rate change is on the order of 4,200 minutes.
MPC model matrix. The complete dynamic 66 model

SPECIALREPORT

has a significant effect on predicting both the top and bottom qualities. Variation in feed quality is captured instead by use of the prediction error measurement of differential temperature between two points in the column. The points were chosen based on both speed and magnitude of dynamic response to feed quality changes, as demonstrated in the dynamic simulation. The steady-state model was also used to demonstrate that the linearity of the model in the composition region of interest could be improved somewhat by the use of logarithmic transforms.
Tuning and commissioning. Tuning MPC controller dynamic parameters was performed by extensive simulation in the MPC simulation environment. Tuning factors were established that provided an appropriate trade-off between controller speed and robustness. Steady-state optimization cost factors for each MV were obtained from the steady-state model by calculating the difference in profitability between the base and MV offset cases. Mar-

matrix is shown in Fig. 6. Note that the sixth step test variable, feed temperature, was dropped since noise in this instrument

TABLE 2. Dependent variables


Description CV1 CV2 CV3 CV4 FIG. 5 FIR models showing effect of feed rate on top quality. CV5 CV6 Top quality (propane in propylene) Bottom quality (propylene in propane) Column differential pressure Compressor amps Reboiler approach temperature Column differential temperature Units mole % mole % barg amp C C

FIG. 6

MPC model matrix.


HYDROCARBON PROCESSING JUNE 2008

SPECIALREPORT

PROCESS AND PLANT OPTIMIZATION


Fig. 8 shows the top and bottom quality trends before and after implementing the controller. The controller was turned off for about a week due to upsets in the FCC unit which is upstream of the PP splitter. As shown, top and bottom quality control are resumed immediately after the controller is placed online again. The rigorous dynamic model was subsequently used as the engine of an operator training simulator (OTS). It was connected to an emulation of the control system and a replica of the operator console, so operator training was provided for startup, shutdown and no-APC operation scenarios.
LITERATURE CITED Butler, D. L., R. A. Cameron, L. D. Eckelman and G. K. McMillan, Virtual Plants for Hands-on Advanced Process Control, ISA Conference, Houston Texas, September 2001. 2 Mather, U. and R. J. Conroy, Successful multivariable control without plant tests, Hydrocarbon Processing, June, 2003. 3 Rugeri, G., F. Trivella and W. van Wassenhove, Dynamic Modelling in Advanced Process Control, AspenWorld, Orlando, Florida, October 2004. 4 Trivella, F. and G. Marchetti, Integration for innovation, Hydrocarbon Engineering, November 2004. 5 Mantelli, V., M. Racheli, R. Bordieri, N. Aloi, F. Trivella and A. Masiello, Integration of Dynamic Simulation and APC: a CDU/VDU Case Study, ERTC Conference, Budapest, Hungary, May 2005. 6 Fournet, A. Use of HYSYS Dynamics as a tool to develop process models for APC applications, Aspentech User Group Meeting, Frankfurt, Germany, April 2007. 7 Valappil, J, D. Messersmith, S. Wale and V. Mehrotra, Lifecycle Dynamic Modelling in the Design and Testing of Advanced Process Control, IEEE Advanced Process Control Applications for Industry Workshop, Vancouver, Canada, May 2006. 8 Gonzalez, R. and J. M. Ferrer, Analyzing the value of first-principle dynamic simulation, Hydrocarbon Processing, September 2006.
1

FIG. 7

MPC interface.

100 100 90 90 80 80 70 70 60 60 50 50 40 40 30 30 20 20 10 10 0 0

Commissioning complete

Controller off due to FCC upset

23-05-05 10:14:49

31-07-05 04:14:49

07-10-05 22:14:49

15-12-05 15:14:49

22-02-06 09:14:49 Nicholas Alsop is the APC manager at Preemraff, a complex


220,000 bpd refinery located on the west coast of Sweden. He has specialized in the field of APC for 14 years, having held positions at Conoco in the UK, BHP and Caltex in Australia. Dr. Alsops first degree is in chemical engineering from the University of Sydney, and his PhD in process control was obtained from the Imperial College, London.

FIG. 8

Top (green) and bottom (red) product quality.

ket prices for products and energy were used in establishing the profitability function. Finally, the MPC controller was built and connection to the DCS was established, the interface for which is shown in Fig. 7. Due to the extensive simulation effort in the MPC offline environment, and the fact that the underlying rigorous model of the PP splitter provided a sound basis on which to build the controller, next to no additional controller tuning was necessary during commissioning.

J.M. Ferrer Almazn holds an MSc degree in electrical engineering (University of Zaragoza, Spain). He worked four years for Dow Chemical at its Tarragona site as a process control engineer, three years at EDS in automatic warehouse systems and seven years at Aspentech developing applications of dynamic simulators in advanced control, ESD design, controllability studies, DCS checkout and operator training simulators.

Article copyright 2008 by Gulf Publishing Company. All rights reserved. Printed in U.S.A. Not to be distributed in electronic or printed form, or posted on a Website, without express written permission of copyright holder.

You might also like