Drilling & Completions

Applications of Machine Learning in Drilling with Petro.ai


 
Here’s a quick video that shows Petro.ai in action, processing real time drilling data.

I recently participated in the SPE Workshop on Applied Data Science and Novel Measurements for Unconventional Development. One thing that really stood out was the number of drilling use cases presented. Several operators and a couple service providers discussed different applications of machine learning (ML) to predict and analyze drilling operations.

In this post I want to introduce some of the challenges and practical considerations we’ve faced at Ruths.ai in applying ML to drilling use cases and also to share some of the insights we’ve gathered.

Wrangling Drilling Data

High frequency drilling data is standardized around the WITSML format, which lowers the barrier to using ML for feature extraction and prediction. However, there are still several challenges to overcome before the data can be used for any ML application.

Raw WITSML data by itself is not very useful. Not only is it a strange format for data but it lacks all context for the rig operation. The WITSML feed will have things like block height or pump rate but not anything about the hole size or drill bit being used. That information is typically stored in a well planning software like WellView or OpenWells. The raw WITSML data needs to be converted to a more useful format, cleaned, and combined with well plan. Furthermore, the raw data comes in at one-second frequency intervals and could contain errors like missing variables and dropped channels that need to be taken into account. Varying rig operations from one service provider to another (not to mention between onshore and offshore) can also make it difficult to build a generalized ML model.

In order to be useful and impactful, we need to decide on who will consume the output of the analysis in advance. The rig crew can have very different needs and priorities compared to engineers or managers in the office. Providing relevant data to the rig crew probably necessitates a real time solution that has no latency and does not dependent on connectivity. This adds complexity to the solution because it requires an edge device that can process the drilling data in real time on the rig. For use in the office, near real time is probably sufficient. Engineers can see the current rig operation, see the past 24 hours, and use the data for planning future operations and performing lookback analysis.

Saas or Paas Approaches

As the presentations at the SPE Workshop showed, operators have taken different approaches with the build vs. buy decision. Some operators have built in-house solutions, others have partnered with vendors to build custom tools, while others have purchased off-the-shelf SaaS products.

As with many analytics use cases, the proliferation of SaaS apps can be challenging to support. The value of the drilling data can be magnified if it’s made available to other technical domains. Many companies have drilling and completions in the same group, yet drilling data is siloed and inaccessible to the completions engineers. The unique approach we took with Petro.ai is to provide a PaaS (platform as a service) that can address multiple use cases, including drilling analytics.

Petro.ai is a cloud ready platform that can be deployed inside your environment. This means the cleaning, processing, and storage of drilling data happens in your private environment and no data has to be shared with third party vendors. The drilling analytics engine cleans the raw data, combines it with the well plan, and infers the rig state every second. Petro.ai predicts the rig state using a machine learning algorithm that requires very little training data and that’s been deployed on both offshore and onshore rigs. The high frequency data is classified into 14 distinct rig states.

  1. Ream Up
  2. Ream Down
  3. Connection while in slips
  4. Friction Test
  5. Tripping Out
  6. Tripping In
  7. Rotating Only/Stationary
  8. Static/Stationary
  9. Circulate Hole/Stationary
  10. Circulating and Rotating/Stationary
  11. Sliding Drilling
  12. Rotating Drilling
  13. Rotating but not Making Hole
  14. Drilling Shoe Track

We also generate eight summary well states:

  1. STS (slip to slip)
  2. TSTS (tripping slip to slip)
  3. STW (slip to weight)
  4. WTS (weight to slip)
  5. Drilling
  6. Trip In
  7. Trip Out
  8. Flat time

The above metrics are then used to create over 60 unique KPIs like on bottom ROP, mean connection times, or cost/lateral foot which are all saved inside Petro.ai. This process enables a massive reduction in data size as the one second data is converted to summary statistics which are then used for analytics. The processed data is presented to end users conveniently: in a web app, through integrations with business intelligence tools like TIBCO Spotfire or Microsoft Power BI, or through a restful API. This flexibility gives you different ways to interact with the data to best fit your needs. Drilling managers might want to access a web dashboard to see KPIs while a data analyst might want to drill down into historical data to compare trends.

Typical Use Cases

After having these drilling tools running with multiple operators, we’ve seen a couple use cases really make an impact. These use cases fall into three buckets.

  1. Near real time
  2. Lookback
  3. Other domains

By bringing in offset and historical data we create performance benchmarks which can be compared to actual performance in real time. These reports are used on location as part of morning meetings on the rig to review the last shift. The rig crew can also review real time torque and drag measurements or look at actual vs. planned AFE tracking.

Back in the office, engineers and managers leverage the analysis for design and lookback analysis. Different crews, rigs, bits, or service providers are compared according to different measures, like ROP or cost per lateral foot.

What’s really exciting is that Petro.ai not only runs all this drilling analysis, but also ingests everything from bulk seismic to completions to production data. Now these drilling metrics can be put in context and made visible to the entire asset team. This has opened the use mechanical specific energy (MSE) as a proxy for rock quality and as an input into multivariate models on completions design.

Although drilling is not the largest portion of the well AFE, it’s prime for the application of analytics due to its structured data. There are clearly areas where the application of analytics can provide both operational and design insights. This makes drilling one of the exciting core use cases of the Petro.ai platform.

If you have any questions, feel free to post below!

Leave a Reply

Your email address will not be published. Required fields are marked *