Yokogawa APC Implementation on Vacuum Distillation Column, in Turkey

Additional authors of this article: Tupras process control experts Berna HaktanırGörkem Oğur, and Duygu Aydın


The Platform for Advanced Process Control and Estimation by Yokogawa is a new state-of-the-art multivariable model based control technology, jointly developed with Shell, and released in Q4 2015. A number of applications have already been implemented at several sites around the world.

A pilot project with Platform for Advanced Process Control and Estimation was carried out by Yokogawa and the Turkish Petroleum Refineries Corporation (Tupras) during March and April earlier this year on a Vacuum Distillation column at the Izmit Refinery. Similar tests have been conducted by Tupras on different process units with Aspen DMC+ and Honeywell RMPCT in the past. Having used ExaSMOC and ExaRQE by Yokogawa for 8 years already, Tupras was familiar with Yokogawa Advanced Solutions, and was eager to test and evaluate the new platform in a real world situation. It is important to note that a key objective of the evaluation was to run the automated step testing function 'Auto-Step', as this was the first time Auto-Step had been used on a real process unit.

Application Overview

The 5C101 Vacuum Column is a part of Plant 5 Crude Unit which was built in 1971, with 12500 m3/day crude and 7000 m3/day vacuum capacity. The column processes atmospheric residuum coming from both Plant 5 and another crude unit.

Beforehand, the unit utilized a SMOC controller, but since the column was being revamped and changes were implemented in the base layer strategy, the controller was turned off and waiting for model renovation. In other words, the trial was not just a migration, but an APC implementation from scratch. The study was completed in 2 months.

The Vacuum Unit application has 5 Manipulated Variables. The control objective is to maintain the LVGO and HVGO product qualities within prescribed specification limits and manage relevant process constraints. The application has an Economic Function designed to push the product properties to their maximum limit as well as to maximize the separation between LVGO and HVGO. This effectively maximizes the production of valuable distillates at the expense of the bottoms vacuum residue. The LVGO and HVGO product qualities are inferred properties with estimation updates based on Laboratory test results. A schematic drawing of the process and MV’s, CV’S is given below.


Figure 1: A schematic drawing of the process


Project Steps

The Pilot project was executed according to the four following steps:

  • Week 1 and 2: Preliminary plant tests and model identification. Two different options were tested (LVGO tray temperature controller in open loop and in closed loop) to help decide on the preferred controller structure. The closed loop option was selected as the operators are familiar with it and for reasons of column operation stability
  • Week 3: Parallel activities: Controller off-line design and run-time software installation followed by run-time system functionality testing with the help of a simple DCS based simulation
  • Week 4: Controller implementation, setting of parameters, closing of loops and fine-tuning based on observation of the closed loop performance
  • Week 5: first Auto-Step run

It should be noted that the inferred properties had been developed previously. The same correlations were re-used and implemented in the controller.


1. Auto-Step

'Auto-Step' is Platform for Advanced Process Control and Estimation's automated step testing functionality. It is embedded in the controller, and gives the user the option to switch the controller mode from Control (i.e. normal closed loop) to Auto-Step. In Auto-Step mode, Manipulated Variables (MV’s) are stepped according to user specified guidelines (maximum step size, minimum switching period), while protecting the specified process constraints. The objective of Auto-Step is to provide an effective way to generate test data to re-estimate the controller internal model accurately. It is useful for a first application implementation and also to maintain the reliability of APC models over the years with reduced engineering effort.

Figure 2: Auto-Step moves of MV’s and CV behavior

Auto-Step was run in multi-MV stepping mode. It was run 3 times for a total of about 6 days. Three to five MV’s were stepped depending on operational requirements. In particular when running the vacuum column in Bitumen mode the Operator prefers to keep the furnace outlet temperature stable, so it is excluded from Auto-Step. A key consideration in running Auto-Step is to set the control limits wide enough to have room to move within the process constraints. When getting close to a process constraint, or in case of a disturbance causing a violation, Auto-Step will typically switch to its “Control” mode for constraint protection. This protects the process but in principle is less favorable in terms of data quality for identification.

Also, in the second run, auto-step is tried at optimization mode. In optimization mode, auto-step testing prefers smaller step sizes. It moves all the CV’s to the limits due to optimization directions and makes step changes nearby these borders which is beneficial for unit operations and quality measures. In the third run of auto step testing, to increase the step sizes and auto-step flexibility, economic function were turned off.

The data generated during the periods in Auto-Step was used for re-identifying the model. Most model parameters were consistent with the preliminary model; pressure effects on the LVGO and HVGO product qualities were updated based on the Auto-Step new model.

In the graph below, PCT (Pressure Compensated Temperature) trend of HVGO, which is used for HVGO T-95% estimation is shown. Blue line shows actual field data and the red line is model estimation.

The graph is divided into four time zones with updated model results due to test data.

  • First zone represents initial model results with the highest mismatch.
  • Second zone is the result of initial data and first auto-step test run data and
  • The third and the fourth zones are also results of cumulative data with each auto-step run. (8 days in auto step and initial step data)

The graph shows that despite slight improvements in the second and third zones, fourth zone with cumulative data of all auto-step test + initial test data has the best results. It is also note mentioning that increasing the test time leads to better models.


Figure 3: Process-Model estimation mismatch was removed after use of auto-step test data


2. Economic Function

Economic Function (EF) is the platforms optimization tool.

Figure 4: Economic Function Configuration Interface

Many linear or quadratic economic functions can be built and used at the same time, with different priorities, these economic functions may have higher priorities than CV’s if desired. 

In the pilot study, this feature helped in some situations where feed maximization is targeted for sake of violating some CV’s up to certain levels. Multiple economic functions can be defined operator can make more than one EF Active in run time.

An example of one of the economic functions configured in pilot study for optimization of MV’s can be seen in the EF interface given in figure 4. 

3. DCS Configuration

One of the typical challenges implementing APC is the often complex integration process between APC and DCS; the development team place a large emphasis on eliminating this risk, the platform allows easy configuration of APC-DCS communication infrastructure according to DCS type. User selects DCS company/type and required point extensions are created automatically. This feature provides fast implementation to DCS and shortens online commissioning period.

In the pilot study, the watchdog application was handled with a watch dog template. MV parameters such as PID Names, Shed Mode and Remote mode are entered in the template. Different limits for move up and move down can be set and all the given limits were tested in the study. It was not possible to violate the limits even if some false values were entered via OPC servers proving the reliability of the application.

Another advantage is that the Operator Interface can be embedded as a DCS graphics page if the DCS is also Yokogawa (Centum VP R5 or above). Otherwise any computer can be used including DCS HMI’s given it has a Windows 7 operating system.

4. Intermediate Variables (POV’s) – Cascade Correction

From a modeling point of view the Platform's 'Design Time' function uses same identification type with AIDA (SMOC) which allows use of POVs -- a feature that helps reflection of process characteristics to model better.

For example, if there is a model between inlet temperature to reactor WABT and thus product quality, users can have a model from inlet temperature to WABT and from WABT to product quality. If WABT is changing due to some disturbances, controller will adjust the inlet temperature to keep the WABT steady before any effect is seen in the product quality. This feature also helps in slow dynamic loops as set point of a PID is often the MV for a MPC controller. Disturbances affecting the process variable of the PID will also affect the CV’s. A correction action often results in more disturbing the operation. An operator can see that and takes no action because he/she knows that this effect will be tolerated via PID loop itself.

In the controller design phase, by applying “cascade correction” option for slow response loops with just one click, set point and the PV of the PID loop is linked and any disturbance tolerated by PID loops will be perceived as zero gain disturbances. This feature is particularly useful in furnace outlet temperature PID’s. With the change of the fuel quality outlet temperature can drop 2-3 degrees which has an effect in the lower section of the column. MPC trying to adjust the quality will increase the outlet temperature set point further which is already higher than its process variable, resulting an oscillatory effect in the temperature and the thus quality. Cascade correction option which uses POV technology fixes this problem fast and effortlessly. In the pilot study LVGO temperature controller and furnace outlet temperature is modeled with cascade correction option as:

TIC.SV --> (Cascade correction) TIC.PV  --> CVs

5. Other Important Features

5.1          Design Time

Platform for Advanced Process Control and Estimation Design Time is like a combination of RQE, AIDA and SMOC, which allows user to build a controller from inferential design phase to modelling and controller building phase in one single platform without switching to different programs. This prevents complications related with switching from one program to another thus shortens controller design period.

5.2          Staged Mode

In Runtime, the platform has ‘Staged Mode’ in which it collects data from the process, calculates MV moves but does not implement them. Observing calculated moves, control engineer can make required revisions of tuning parameters before implementation. Thus the controller can be tested online beforehand and then commissioned without compromising process safety.

5.3          Runtime Interface

Figure 5: Run time Interface (MV move prediction, CVprediction)

Platform for Advanced Process Control and Estimation Runtime has an improved user interface. It has an embedded data historian which allows configuration of trends of desired MV’s and CV’s. Trend configuration tool is found user friendly.  In addition to CV prediction, calculated MV moves in prediction horizon can also be monitored. Also dynamic relationships are made visible to the operator, i.e. all MV’s related to a CV and their move plan, limiting factors are shown. This helps operator to understand controller behavior and creates ownership of the application by the operators.

Pilot Highlights

In a nutshell, highlights were:

  • Very fast implementation: The controller was running in closed loop continuously within two weeks after initial step test
  • Controller performance recognized as very satisfactory by client
  • Auto-Step was run 3 times for a total of about 6 days; the data generated was used for re-identifying the model. Most model parameters were consistent with the preliminary model; pressure effects were updated based on the new Auto-Step data
  • Compared to SMOC, an improved runtime interface, controller behavior is easily understood by the operators which creates ownership in a short time
  • Single platform for inferential design, modeling and controller design eases and speeds up controller implementation
  • DCS communication infrastructure is easy to configure for any type of DCS
  • Models make uses of intermediate variables (POVs) resulting in robust control
  • Strong optimization: More than one economic function can be implemented as well as EF priorities that can be set higher than CV’s
  • CV high and low limits can be configured separately allowing user to give higher priority for certain limits