Predictive Drilling Analytics: the future of drilling?

Data is used increasingly to improve drilling performance and reduce operational problems. But what if an analytical model could also serve as an “early-warning system”, in order to avoid significant Non-Productive Time (NPT) events during operations? For the past 2 years, Maersk Oil has been working on just such a technology: Predictive Drilling Analytics (PDA). We talk to Joy Sunday Oyovwevotu, Senior Drilling Engineer in charge of this project.

Joy Oyovwevotu

Drilling & Wells

Could you briefly describe the background to the project?

Joy Oyovwevotu : Maersk Oil started the Predictive Drilling Analytics (PDA) project in May 2015 with IBM to test whether data analytics techniques could be applied to drilling.


We also wanted to assess whether real-time and static drilling data can be used to predict and avoid significant Non-Productive Time (NPT) events. Could the model recognize anomalies by applying pattern-recognition to historical data and then use that knowledge to predict significant NPT events in a new well?

Total & Maersk Oil

By Michael Borrell

When the project was initiated, a series of significant NPT events had shown that human cognition alone can’t provide the required level of operational efficiency in drilling. So the PDA project was conceived to provide just such a capability. One day, cognitive analytics may become the drilling engineer’s partner in a continuous effort to improve drilling performance.


What are the main objectives of the PDA model?

J.O : We designed the PDA model to do three things. First of all, we wanted it to identify the onset of borehole instability when drilling through the overburden. The second objective was to assess the probability of pack-off during well operations. Thirdly, we wanted it to assess borehole conditions before running casing/liner and cementing operations.


Is the model purely physics-based, or is there more to it?

J.O : No, it’s a hybrid model which combines statistical analytics with drilling domain expertise, including experience-based logical relationships between well-understood drilling parameters. This approach allowed us to overcome some of the data gaps in the historical database and jump-start the model’s learning rate.


The development of this hybrid model proved to be one of the most successful aspects of the project. Although the PDA model is not purely physics-based, almost all of its components reflect different physical aspects of the drilling process, as well as relationships between established and newly-derived parameters. These physics-based or constrained parameters are then extended over a range of borehole conditions, using analytics to account for uncertainties inherent in derived correlations.


Unlike a purely physics-based model, the analytics help the model to accommodate new scenarios and provide a basis for recognizing similar circumstances in the future.


The collaboration with our drilling experts helps us distinguish between correlations and causality as the model was developed. Rigorous testing of correlations helps reduce the likelihood of mistaking spurious correlations for actual relationships thereby leading to wrong predictions and incorrect warnings.


The system was tested in May 2017. What were the results?

J.O : We extracted the historical RT data of two discovery wells. The whole idea behind such a test was to give us the opportunity to gain confidence in the PDA model and see how to improve it.


On the first well, the model did relatively well in the sense that it detected 21 out of 24 anomalies. We also ran it in parallel with other commercially-available drilling software, so we could actually compare the results and see how well the model was doing. So in 21 out of 24 cases, there was agreement between what the model predicted and what we were seeing getting from the other software.


The three cases that were not in agreement were false negatives; in other words, cases where the model suggested that nothing would happen, but something did.  It is worrying because they could cause us to miss significant NPT events. That was definitely very important because that’s the whole point of doing predictive analytics in the first place. We spent a lot of time trying to understand what happened, and concluded that the false predictions were either due to poor data from our end, or to something that the model had not seen before.


And when we got to the second well, which was the final one, the model predicted events correctly. The field trial has shown that if the system is given enough data and operational scenarios, the model can become a reliable tool for identifying risks, giving us enough warning time to take mitigation measures, thus proving its value in supporting decision-making for well operations and risk management.


The conclusion is that our model needs more data to improve; the more our model is exposed to training data, the better it will become.


What about the future?

J.O : In the future, we’d like to be able to evaluate other options that could make implementation in a production system cheaper and more cost-effective. The most practical way to reduce the implementation cost of the PDA would be to run it on multiple wells simultaneously. We believe that systems like this will one day be used throughout the industry, providing vital information to support drilling activities.