Hydrocarbon Processing - October 2021 - 20

Digital
By combining a physics-based model with ML in a hybrid
modeling scheme, the best of both approaches is achieved.
The result is a system that uses data-driven modeling, which
is enhanced by the first-principal, physics-based models. This
hybrid solution offers the robustness and quality of a highfidelity
simulator with the speed of response of a ML solution.
Enhancing data-driven models and physical simulators.
Historical records can often lack the volume of data over sufficient
periods to get robust predictions from ML algorithms.
However, all potential operating scenarios can be created and
taught to the data-driven model by simulating data.
Equally, data collected in real time and analyzed by the
data-driven model can improve simulators to create a better
match with reality and incorporate behavior outside the
scope of the physical model. For example, perceptual quantizer
curves can change slowly over time or experience a sudden
step change if the well configuration on the other side of
the pipeline changes. Models can be better matched to 'what
if' scenarios using real-world data.
Advantages of hybrid modelling. Using a hybrid model
offers several benefits. This approach means the modeled
state is updated using real-time data, and data for the ML algorithm
to ingest can be generated around typical operating
conditions using the simulator. The ML algorithms offer the
speed required to optimize the value of real-time data and can
be used to optimize the system. Once the ML model makes
a fast prediction, this can be checked in the simulator, and
the solution fed back to the operator. The resulting solution
offers the speed and versatility of a data-driven model and
incorporates the transparency and physical constraints of the
system provided by the simulator, providing verification and
trust for safety-critical applications.
Takeaway. In a data-driven world, the value contained within
is continually increasing and can offer tangible benefits to
the oil and gas industry. The value of this data decreases over
time, so a fast response will deliver greater advantages. However,
this cannot be at the expense of safety, and modeling
systems must provide transparency and robustness to ensure
decisions are not biased and risks to people, assets and processes
are not increased.
On their own, both pure data-driven models and physical
simulators have shortcomings. However, a hybrid modelingbased
solution offers an exciting new prospect to reap the
benefits of both approaches to deliver an enhanced, reactive,
robust modeling solution in safety-critical, heavy industrial
applications.
EIVIND EIDE is a Senior Engineer at Kongsberg Digital. Based in Oslo, Norway,
he has solid experience in applied ML, hybrid ML and software development
for industrial applications. He leads a team developing mathematical modeling
powered SaaS products for the energy industry. Mr. Eide earned an MS degree
in information engineering from the University of Cambridge.
2021 FALL REFINING
AND EQUIPMENT
STANDARDS MEETING
NOVEMBER 15-18 | ORLANDO, FLORIDA
Join API for the 2021 Fall Refining and
Equipment Standards Meeting to learn about
the latest research and development on
refining process equipment, including future
R&D efforts.
This meeting will be held in a hybrid format,
in-person and virtually. It will be open to
anyone who wishes to participate.
Register at www.API.org/fallrefining
Copyright 2021 - American Petroleum Institute, all rights reserved. API and the API logo are
either trademarks or registered trademarks of API in the United States and/or other countries.
20 OCTOBER 2021 | HydrocarbonProcessing.com

http://www.API.org/fallrefining http://www.HydrocarbonProcessing.com

Hydrocarbon Processing - October 2021

Table of Contents for the Digital Edition of Hydrocarbon Processing - October 2021