In the second part of this two-part article, Lisa J. Graham, Seeq Corp., provides use cases of machine learning in action in the oil and gas industry.
Part one is available to read here: https://www.oilfieldtechnology.com/digital-oilfield/04052020/harnessing-machine-learning--part-one/.
In developing an understanding of how this new age of analytics can drive continuous improvement, the focus must be on the idea of reinforcing decision support, not simply execution. Decision support focuses on situations, systems, and people to provide insight, assessment, and experience. Most importantly, decision support is highly collaborative and enables action.
The following use case studies illustrate this decision support concept in action. Predictive analytics enable SMEs to identify what will likely happen, based on current-state data, enabling corrective action to be taken in time to either change, avoid, or plan for the outcome.
Use case: Instantaneous shut-in pressure analysis to optimise well production
For a given process or asset, standard relationships and first principle-based equations form the basis of SME insight into operations. Application of ML techniques can benefit from incorporation of these known relationships and drive understanding of the model results.
Optimising well spacing and perforations down the pipe is a complex process for upstream oil and gas companies. This is because wells interact with the reservoir and each other, affecting how they will perform when put into production. A key metric is the calculation of instantaneous shut-in pressure (ISIP), which is particularly important to determine which adjustments will result in a more productive well. Analysis is hindered when calculation of this metric is done individually for each well in different spreadsheets or not completed for all wells in a timely manner.
The solution provides SMEs with an application so they can quickly and consistently calculate the ISIP across hundreds of wells. Figure 1 shows an example ISIP calculation performed in Seeq.
Figure 1. An example ISIP calculation performed in Seeq, with results captured in Journal for later reference (left pane).
SMEs used their understanding of the process to first create logic to identify the well stages, and to then perform the ISIP calculation. Once complete, the same analytics application enabled the production team to quickly and easily scale its analysis to hundreds of wells. The results were viewed by the SMEs and formed the basis for further analytics and ML.
Use case: controlling finished product properties
Refining and chemical operations are dynamic and require constant adjustment based on changes in product demand, feed slate, catalyst deactivation, and asset/equipment availability and condition. SMEs must therefore use an advanced analytics application providing them with the ability to develop predictive models and refine them as conditions change.
Controlling finished product properties is critical for maximising product quality, and as a result, profitability. Many companies control finished product properties based on feedback from lab results received some hours after material is produced. If the results show the product is not on-specification, then significant margin loss associated with product downgrades is realised. Alternatively, if the product is above specification, extra materials and energy were unnecessarily consumed. Controlling the finished product properties is further complicated on units which continually switch between making different grades of product.
The solution is a predictive model to accurately forecast product properties based on conditions in an upstream portion of the process. The example in Figure 2 shows prediction of product viscosity for a polymerisation process.
Figure 2. Example of modelling to predict product viscosity for a polymerisation process based on upstream conditions in a multi-product facility. Grade transitions are identified by the solid bars (capsules) along the top of the workbench.
A model is created and boundaries are assigned to monitor deviation from the target. The SME can easily refine and update the model as equipment and feed changes occur, and as new products are introduced. Using the model for near-real time quality control rather than traditional feedback methods reduced product margin losses by >US$1 million annually for a unit with an average production rate around 40 000 lb/hr and a variable margin difference of US$100/mT between on-specification and off-specification product.
Use case: implementing a collaborative preventative maintenance analytics strategy
The final example comes from Devon Energy, a leading independent oil and natural gas exploration and production company in North America. Devon Energy has implemented advanced data historian and analytics technologies to transform significant amounts of data into actionable insights for the support of data-driven decisions2,3. The technologies are used by many groups including completion, drilling, and production—and the initiative required close cooperation between the company’s operations technology (OT) and information technology (IT) groups.
Analytics were used to develop a risk-based maintenance strategy by leveraging techniques to determine the condition of in-service and predict failures, providing optimal prioritisation and scheduling of maintenance activities across a fleet of assets. SMEs used the same analytics application to interact with the data, build models, and apply ML algorithms.
During operation, assets such as pumps may experience a fluctuating set of operating conditions. Depending on the specific service, these conditions could include frequent on/off cycling, and running at various speeds as throughput requirements change or adjustments are required due to variations in feed properties. These changing conditions make equipment degradation extremely difficult to observe from simple visual examination of process data, but advanced analytics algorithms can be used to find hidden trends in the process data and illuminate upcoming failures (Figure 4 trend view).
Figure 3. Using Seeq, advanced analytics algorithms were used to find hidden trends in the process data illuminating an upcoming failure (upper trend view). An increase in the amperage signal alongside an unchanging voltage signal (lower scatterplot view) led the SME to conclude failure was pending2.
Success was driven by empowering the SME to easily access the data, apply the analytics algorithm, and then cross check results. While the algorithm revealed an increase in amperage over time for a common set of conditions, the observation was not actionable until the SME used her process knowledge to cross check the results with other data.
Applied analytical methods must keep pace with constant variations in operating conditions. ML applications must therefore be a synergistic exercise between human and machine, including:4
- The ability to interact with the data and correlations to further apply SME knowledge and intuitions based on experiences.
- Quick access to a data-rich environment where ML algorithms can be applied to delve into the data in a non-biased manner to draw out currently unrecognised correlations.
- Opportunities to explore the processes with colleagues to further eliminate biases and heuristics.
The feedback from this experience provides continuous improvement opportunities to drive improved outcomes. When predictive analytics are combined with contextual data, SMEs can make informed decisions to drive business value. Understanding trade-offs, the SME can incorporate the different data sets along with business objectives to arrive at the best outcome.
Engineering skill sets will continue to grow as technology advancements provide more tools. This is the time not to divide and conquer, but to instead align best practices to harness the best of humans and machines in concert.
1. D. Faggella, “What is Machine Learning?”, Emerj, 21 December 2018. https://emerj.com/ai-glossary-terms/what-is-machine-learning/
2. J. Abel, “Devon Energy Uses Real-time Data and Advanced Analytics to Make Better Decisions”, 16 July 2018.
3. D. Koetter-Manson, “Finding Hidden Value in Real-Time Data”, Devon Energy, ARC Industry Forum, February 2019.
4. L. Graham, “The Human-Machine Matchup: Combining human reasoning with the ML functionality found in advanced analytics applications improves pharma processes”, Pharma Manufacturing, April 2019.
Figures all courtesy of Seeq
The author wishes to thank Krista Novstrup and Brian Parsonnet from Seeq for their invaluable contributions to this paper.
Part one of this article is available to read here: https://www.oilfieldtechnology.com/digital-oilfield/04052020/harnessing-machine-learning--part-one/.
Read the article online at: https://www.oilfieldtechnology.com/digital-oilfield/05052020/harnessing-machine-learning--part-two/