Greg: Measurements are an essential window into the online operation needed for process control and optimization. The view can be blurred or missing because there are limitations besides accuracy and noise in measurement performance. Most notable are the “5Rs” (rangeability, reliability, repeatability, resolution and response time). Of course, the best solution is to use sensor and transmitter designs and installations that minimize these limitations.
For flow measurement, magnetic flowmeters and especially Coriolis meters offer accuracy and 5Rs that are an order of magnitude better than orifice and vortex meters. For temperature, better 5Rs correspond to tight-fitted and spring-loaded resistance temperature detectors (RTD) in tapered thermowell and a head-mounted, smart transmitter with sufficient immersion in a stream with 1-10 fps velocity. For pH, better 5Rs translate to shrouded, high-performance, glass-bulb measurement electrodes and compatible liquid, preferably flowing junction reference electrodes with sufficient immersion in a stream with 5-8 fps velocity. Since pH 5Rs are particularly challenging, three measurements with middle-signal selection (MSS) are essential because it inherently ignores—to a significant degree—the electrode with the worst 5Rs. MSS has been used for other measurements particularly when they can cause a shutdown. MSS applied to all critical measurements in a large continuous chemical plant stressed to give a production rate more than double the original design saved 10 to 20 million dollars a year by increasing onstream time from 5 months to 5 years.
The use of an online inferential measurement in MSS can improve performance and diagnostics, besides reducing hardware and maintenance costs. This is especially true in pH systems, when electrode life expectancy is a matter of weeks, which is a common occurrence. Even the best pH electrode life with good 5Rs in industrial processes rarely exceeds 10 months.
Models can detect and predict measurement problems. They can improve the measurement 5Rs. Models can also increase process knowledge provided by additional measurements. To open our minds to the opportunities that models offer to improve process performance with increased measurement capability and intelligence, we continue our conversation with José María Ferrer, who has more than 25 years of experience in dynamic simulation and control of hydrocarbon processes, and presently serves as a senior advisor at Inprocess Technology & Consulting Group.
José, how are models used for online inferential measurements (inferentials)?
José: Simulation models can be used to infer any measured (for fault detection) or unmeasured process variables (pressure, temperature, flow, level) or property (density, heating value, composition). This link, “Ensuring a successful FPSO start-up with Inprocess Digital Twins,” contains a case study of an online digital twin of an offshore platform, which provided key variables such as residence times in separators, gas Wobbe Index, and oil Reid vapor pressure.
One of the more interesting applications is inferentials for key compositions in distillation columns. There are three approaches:
1. Binary mixture in equilibrium: At 10 Kg/cm2, pure propane boils at 30°C and pure i-butane boils at 70°C. If you have a tray with a mixture of propane and i-butane with the pressure at 10 Kg/cm2 and the temperature transmitter read 50 °C, can you guess the percentage of the mixture? Instead of a guess, a concept employing simple mathematics can compute a binary mixture composition.
The same concept is used with non-linear curves of bubble and dew point (calculated offline by simulation) to determine the composition of any quasi-binary mixture if pressure and temperature is measured at the tray. It will result in a simple formula to be coded in the DCS. More details are available at “ACOWUG 2016 - Getting help from process simulation Three cases studies”.
2. Non-binary mixture in equilibrium: When there are more significant components in the mixture, the previous simple approach can’t be used. The solution is building a well-calibrated, steady-state distillation column model, and performing a multivariable case study for all the potential conditions of the column offline. From there, you can derive a valid correlation. This can be coded in the distributed control system (DCS) as a multivariable large formula, a lookup large matrix, or a machine-learning (ML) soft sensor. This case study, “Machine learning soft-sensors trained with Digital Twin” offers more information.
3. Online dynamic model: When the other two approaches are not valid, an online digital twin is used. That is a dynamic simulation model of the column being fed second by second by plant data for the boundaries of the model and setpoint of controllers. Then, it performs a second-by-second calculation of any profile composition, including replication of online analyzers measurements. There are other interesting profiles such as internal flows, temperatures, and percentage of flooding. The presentation “ACOWUG 2022-Developing inferentials based on Digital Twins - Distillation Cases Studies” contains several examples.
Greg: An obvious role of online inferential measurements is verifying, supplementing and replacing online or at-line analyzers. Online inferential measurements without response time are used for control to minimize loop dead time, particularly from the sample transportation delay and cycle time of analyzers. All online, inferential measurements need to be corrected. The measurement response time is included when corrected by analyzer results.
José, how does an Inferential measurement bias update work and how do you tune it?
José: Having a good inferential is critical for any control or model-predictive control (MPC) application. Have you tried to drive a car by only looking at the road behind you? This is what online analyzers give you, not the qualities of the product right now, but many minutes ago.
Having a good inferential (either explicit formula or online model) isn’t 100% perfect. They usually have a variable bias (offset) since there could be some other, small unmeasured disturbances or inaccuracies in the formula or online model. The bias update mechanism is a simple routine that continuously calculates this bias based on the real online analyzer measurement (or laboratory analysis) and continuously corrects the value given by the formula or model.
Typically, there are about six parameters to adjust in this bias update calculation, and how to tune those parameters to obtain the best bias update usually isn’t well explained. I made my own method to tune them and implemented them in an interactive Excel spreadsheet, which uses historical data. We teach this method and many other things in our three-day “Simulation for Process Control Engineers” training course. This is a link to the “Excel file and method.
Greg: How can models be used for automatic fault detection?
José: Simulation models represent the “ideality” of the process, where all simulated instruments work perfectly, all control valves work according to their valve characteristic, all rotating equipment obey the manufacturing performances test curves, all heat exchangers work as they should, and so on. With the online, dynamic model connected to plant data, it could reveal an incipient fault of the instrument, analyzer, control valve or equipment. Most of the faults are detected when the model is built and don’t match some plant data. We frequently find that certain instruments weren’t measuring as they should.
I remember a 230-tray C3Splitter in Belgium. My beautiful dynamic model fit very well with all plant data and composition, but not with all temperatures of the column. I spent some time reviewing multiple property packages and if any of them matched the temperatures. Finally, I decided to put on the boots and helmet, and climb the 110-m tall column. I stopped every five meters to take a temperature reading on the metal wall of the column since it was uninsulated. The view at the top was spectacular. Analyzing the data, I discovered all temperature sensors at that plant were incorrectly calibrated with 4° offsets. After correcting them, the column matched the model perfectly.
In this presentation there’s more details about using simulation models in that area.
Greg: I provided a perspective on the types and value of models for metrics and inferential measurements in my Control Talk columns “Top of the bottom line” and “At the IIoT crossroads”.
