Hydrocarbon Processing - October 2021 - 19

Digital
E. EIDE
Kongsberg Digital, Oslo, Norway
Hybrid modeling: Unlocking the value of data
for oil industry assets
We live in a world increasingly driven by data scenarios for
advanced analytics, predictive modeling and machine learning
(ML), but how reliable is this for large oil and gas assets
where safety and reliability are of paramount importance?
Can the industry rely purely on machines to anticipate its
needs, or does it need to stay with the more traditional physical
modeling to determine safety, lifetime and maintenance
of assets? The author believes the answer lies in a combination
of the two.
There is no denying the value of data. Historically, there
was not enough data to support ML algorithms, which improve
with greater volumes of historical data.
However, the falling cost and size of electronics, distributed
intelligence, easy connectivity and the Internet of Things
(IoT) have contributed to an exponential rise in the amount
of data available about assets and processes. Termed 'big
data,' this opens the possibility of deeper insights into the
performance of assets.
However, pieces of data on their own do not add value-
this only comes with context and domain knowledge to enable
the interpretation of data and extract value from it. If
done correctly, the benefits can be huge, including reduced
costs, optimized capacity and less downtime.
Artificial intelligence (AI) technology is only an enabler
to handle data. However, it must offer and require business
input and domain knowledge for an organization to harness
the reward. In addition, it needs to be combined with crosscollaboration
of data scientists and software engineers who
can build a good, data-driven solution. Put these together
correctly, and only the imagination limits the use cases for
the oil and gas industry. For example, steps to increase the uptime
and performance of an oil rig by detecting component
fatigue and identifying areas of predictive maintenance could
save companies millions of dollars a year.
How can safety be ensured? One of the challenges for
heavy-asset industries is to manage safety-critical applications.
How do you know you have the required data quality?
With the complexity of constantly changing assets and
evolving processes, a pure data-driven model has difficulties
in analyzing and predicting time series for safety-critical applications.
It may also not include enough of the operating
space to really understand the underlying physics, especially
with the complexity of incorporating energy and mass conservation,
or it may not have enough historical data to learn
the insights needed. This could result in biased decisions,
missed opportunities, outdated decisions, and, critically, put
equipment, the environment and people at risk. This is where
high-fidelity simulators using physical models can help.
High-fidelity simulators are tried and proven and, as they
are based on first-principle physics, ensure system modeling
in the full domain of operations. However, they are computationally
expensive, making frequent predictions and real-time
optimization difficult. Solving these physics-based models
can also be complicated and time-consuming, and they may
suffer from small differences between the plant and model as
components wear or hidden parameters change.
The problem is that both pure data-driven models and
high-fidelity simulators on their own have too much uncertainty
to provide reliable, timely predictive and prescriptive
analysis in safety-critical applications. The answer lies in a
hybrid of the two.
Hybrid modeling: Unlocking the value of data. Although
pure data-driven models and high-fidelity simulators
of the physical world each have shortcomings alone, they
complement each other perfectly. Data-driven models can
learn statistical dependencies and process data a magnitude
faster than simulators. Conversely, physical models offer robustness
and transparency to ensure crucial, underlying factors
are not missed.
Correctly engineered and with enough information from
current operations, physics-based models and simulators
can help the user understand complex processes and predict
future events. This method is widely used and well-proven
across multiple applications, including predicting the orbits
of rockets or the behavior of nano-sized objects in modern
electronics. ML algorithms are also designed to predict outcomes,
which depend on the problem being addressed.
Fundamentally, there are two different scenarios for predictive
modeling. In the first, there is a lot of experimental
data on the behavior of the system but no direct theoretical
knowledge. This makes it impossible to formulate a mathematical
model of the system. Using ML algorithms, however,
underlying patterns and correlations can be identified, and
predictions made.
In the second predictive modeling scenario, the problem
being addressed can be described using a physics-based
model. This does not mean that a ML approach will not
work. Using artificial neural networks designed to emulate
the human brain and given enough examples of how the
physical system behaves, the ML model will learn physics in
the same way humans learn and, over time, it too will make
accurate predictions.
Hydrocarbon Processing | OCTOBER 2021 19
Hydrocarbon Processing - October 2021

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