Artificial Intelligence in Oil & Gas Production

Solution Seeker, a Norwegian tech start-up and spin-off from the Norwegian University of Science and Technology, is developing the world’s first artificial intelligence for real-time production optimisation.
This article appeared in Vol. 15, No. 3 - 2018


Centre for Integrated Operations in the Petroleum Industry

The Solution Seeker team. In 2007, the Norwegian University of Science and Technology (NTNU) launched the Center for Integrated Operations in the Petroleum Industry (IO Center), focused on developing novel and innovative methods for oil and gas production optimisation. Among the IO Center partners was oil majors:

  • ConocoPhillips
  • Engie
  • Statoil
  • TotalBP
  • Shell
  • Eni
  • and Petrobras.

The IO Center’s research were split into short- and long-term production optimisation. Vidar Gunnerud, at the time a PhD student, led the short-term optimisation programme, which initially aimed at developing optimisation algorithms to run on top of multiphase flow and process simulators in real time. 

Vidar Gunnerud, founder and CEO of Solution Seeker. Having developed his algorithms, however, Gunnerud realised that third party simulators were both inaccurate and too slow for the purpose of real-time optimisation. In addition, they were hard to maintain. How could his advanced optimisation algorithms perform when running on these slow and inaccurate digital twins?

It was still early days in the fourth industrial revolution; 2007 was the year of the first iPhone, the year after Facebook and still two years before Uber. Few were yet to talk about big data and machine learning. All the same, Gunnerud and his professor Bjarne Foss at the Research Group at Engineering Cybernetics NTNU rejected the assumption-based simulators and asked themselves “why don’t we instead plug directly into the oil field itself, read all its data and see what is actually happening?”

During the programme five PhDs and 20 master’s projects had studied real-time optimisation, and more than 50 scientific papers had been published on the topic. Three major inventions were made; namely how to build purely data-driven models, how to set up experiments to generate new data, and how to optimise from these models in practice. Based on these inventions, two of which were patented, Vidar Gunnerud, Bjarne Foss and NTNU, through its Technology Transfer Office, set out to commercialise the technology, and on 31 January 2013 they formed Solution Seeker as an NTNU spin-off company – a company now well known in the Norwegian tech start-up scene.

Oil Production Analytics & Optimisation on the Norwegian Continental Shelf

Long-Term Collaboration

The huge quantities of raw data generated during production can be used to optimise the process. © Solution Seeker. ConocoPhillips and Engie became interested in the novel approach to production analytics and optimisation, and in 2014 they entered into a long-term collaboration with Solution Seeker to further develop these methods. The start-up was given live data access and followed the pilot fields over a couple of full year cycles, learning how to make the methods work in a truly operative environment and understanding the real challenges and pain points both in the process and the work processes which the algorithms were to support.

Where oil and gas production differs significantly from other processes (such as refineries, petrochemical and power plants) is the high number of unknowns. Reservoir, well and flow dynamics are never fully understood, and they are continuously changing as the reservoir is drained, new wells are drilled or shut in, and the production system modified. Production data is typically sparse and noisy, with a great deal of inherent uncertainty in the measurements themselves.

Furthermore, key sensors may fall out or drift over time. All this leaves the production team with the challenging task of continuously optimising and tuning the production settings in the face of uncertainty: identifying the optimal choke settings, gas lift allocation, well routing, pump speeds, etc. Typically there are tens or hundreds of production settings to be adjusted, and thousands, millions or even billions of relevant combinations to be considered. A successful real-time algorithm needs to be fast, accurate and robust to solve this stochastic optimisation problem and capture the full production potential within the limitations of the production system.

Artificial Intelligence as Service

During 2015–2017, Solution Seeker obtained proof of concept of its methods while also developing new ones. Additional oil companies joined it: first Wintershall, then Lundin and most recently Aker BP. In close collaboration with its partners, the company is developing a full-stack ‘AI as a Service’ technology, called ProductionCompass, with a three-step data pipeline: from data mining to machine learning to optimisation.

How Does Oil Production Optimisation Artificial Intelligence Work?

Through its data mining algorithms, ProductionCompass uses vast amounts of raw production data, capturing thousands of time series, continuously sampled and tracing them back over years. Leveraging machine learning techniques for automatic pattern recognition and classification, the production data is refined and distilled through advanced statistical analyses and compressed to high-grade information through proprietary and patented algorithms. It then builds estimation and prediction models by leveraging multiple real-time machine learning systems working in parallel. AI combines hierarchical neural networks, first principle physics, statistical models, and truly known parameters. This enables the technology to capture and analyse the dynamics of the production system and separate reservoir effects from the production control responses such as choke positions, gas lift rates, artificial lift equipment, and network routing of the wells.

Finally, leveraging the power of its predictive capabilities with up-to-date, fast and accurate models, ProductionCompass enables a continuous search for maximum with its optimisation algorithm.

Issues with Machine Learning

A well-known issue of machine learning is the exploit/explore trade-off; should one exploit the existing knowledge, or should one explore for new data in order to improve the models and in turn exploit even better models in the future?

Because ProductionCompass captures and calculates the uncertainties through the whole data pipeline, from raw data to estimates to predictions, oil companies are able to exploit the upside potential in their systems while managing downside risk through the built-in tracking and stop-loss functionality. Furthermore, Solution Seeker’s data scientists can design tailored experiments to provoke responses from the field in order to reveal dynamics and behaviour as yet unknown in the production history of the field.

Harnessing Artificial Intelligence in Oil & Gas Production

As Solution Seeker has developed and deployed its methods, new uses for the technology appear. This is part of the lean start-up philosophy adopted by Gunnerud from day one – he and his team embrace a collaborative approach, working in partnership with the oil companies, with frequent interactions, rapid development and deployment of new methods, a direct feedback loop from the end users, and a pragmatic approach to problem solving.

Problem Solving

One such case appeared when an oil company experienced severe slugging, which could not be explained by the simulators used by their third-party flow assurance advisors. Slugging was not a part of the current scope for Solution Seeker, but as it was so severe that production had to be reduced significantly, it overshadowed any other optimisation effort.

This slugging could not be explained by the simulators and the assumptions they rely on – but it was easily observable to the data mining algorithms reading the actual production data streamed directly from the field itself. By exploring and developing new machine learning algorithms, and applying them to the mined data, Solution Seeker was able to identify the problem and pin-point the troubling well, enabling the client to choke it back and regain control over the output and resume maximum production.

Maximising Hydrocarbon Production: Minimising Loss

Through its work to date, Solution Seeker and its clients have identified three main drivers for value capture: namely ensuring best practice based on existing knowledge; further optimisation based on data-driven prediction models; and freeing up engineer time to drive creative problem solving.

McKinsey recently undertook a study for a major oil company on the NCS, showing an average 10% performance difference between the best and worst performing production team on the same field. Even for the very best performing field, the difference was 5%. By default, and before any optimisation, the AI learns; the data mining and machine learning algorithms learn from all the production teams’ actions over the last few years – both those who were successful and those who were not. Hence it learns the best practice observed to date and avoids the mistakes of the past.

Based on this learning, the AI then builds prediction models. What is there between and beyond what we have observed? What is the better combination of production settings? By leveraging these prediction models into advanced mathematical optimisation algorithms, production is expected to improve from historic best practice by 2–5%.

Finally, as AI is superior at analysing production history, significant amounts of highly valuable time is freed up for the production engineer. This allows the production engineer to do what can be called creative problem solving, in which humans remain superior to machines. As an example, one of Solution Seeker’s clients has several shut-in wells, and now the production engineer can spend more time figuring out how to restart these wells and further add to the production maximisation objective.

Global Deploy of Artificial Intelligence in Oil Production

As oil companies globally start to adopt AI, Solution Seeker is preparing to scale its technology and operations. The first version of the technology will be made commercially available later this year, and the company is already in talks with operators from South and North America, to the Middle East and to Asia. With highly configurable software, powerful cloud computing and modern web technology, the company is looking forward to being able to efficiently serve clients globally.


Related Articles

Reservoir Management Worldwide

A Workflow for Success in Shales

Following a rock properties-based workflow for shale plays makes it possible to gather intelligence which not only defines where to drill and how to frac, but also better predicts the economic outcome.

Reservoir Management Europe

Getting into Hot Water

Hot water from subsurface reservoirs can be used directly to heat our houses. Data, software and knowledge from oil industry operations can help delineate this under-used resource, as is now being done in The Netherlands.