Drilling is critical in the development of oil and gas resources. Drilling efficiently is a critical driver to reducing economic loss as well as reducing the environmental impact from such drilling.
One of the key components of being able to drill efficiently is understanding the rate of penetration or ROP. Onshore drilling rates range from $200,000 to $310,000 per day while offshore rates could reach up to $800,000 per day. Maximizing your ROP reduces the number of days required to drill which reduces the overall cost while also reducing the indirect impact to the land and environment as a whole.
Data
A large volume of drilling data is commonly collected by service companies during the drilling process. In addition, petrophysical logs are often recorded as well. The rate at which a drill can cut through a material such as rock is a function of the drill and material itself. For example, drilling through sandstone would result in a different expected ROP than drilling through limestone. Predicting the soil and rock consistency can be highly valuable to the drilling process. This will be addressed in a future post, but for purposes of this article, we will focus on accurately predicting the ROP based solely on drilling data.
When looking at the features of drilling data, the ROP of the drill is a direct function of drilling attributes such as the surface torque, the weight on the bit, weight on hook, as well as information about the drill mud and bore hole such as pump time, stand pipe pressure, mud flow in. While physics can help determine the data points or features that should directly impact ROP, machine learning is what allows us to analyze patterns within these data points.
Machine Learning
In this particular use case, we are seeking to predict a value or specifically the value of ROP. As such, in machine learning, this would utilize a regression algorithm.
Taking asingle run worth of drilling data from the Volve drilling data set, we extracted the key features above as well as the known ROP to learn from. We uploaded the data to Elipsa choosing the ROP column as the values to learn, and the remaining data points as the features or data to learn from.
Elipsa’s no-code platform runs through an automated data science experiment performing tasks such as data splitting, normalization, etc. and then builds a series of machine learning models on the well data. Testing data is then automatically run against the model and resulting models are sorted based on the predictive results on this unseen data.
In our example, a Decision Tree algorithm proved to be most accurate at predicting the ROP values. The results show how close predictions were inside specific thresholds as well as showing which data points are most important to driving the success of the final model.
Making predictions against this new model is as simple as an API call, allowing your existing data systems to become AI enabled.
Taking a new wellsite log, we run the data against the model to predict ROP. As you can see from the results, the majority of the predicted results are inline with actual.
ROP Optimization
As we have seen, through machine learning, we are able to accurately predict the drills ROP based on drilling and mud log data. However, those logs also include the current ROP. So, you may be asking where the value is in predicting something that is already known in real-time.
To extract the most value from these predictions, we are not looking solely to predict the ROP based on current data values but to understand how the ROP will change as values change. This focus then enables us to understand what changes can be made to maximize ROP.
Certain data points used by our model are a function of the environment surrounding the bit. While those give a great deal of information helpful in predicting the ROP, we cannot change them. However, if we can take the current data points along with data points that we can change, such as Torque, we can optimize the drill configuration to result in the highest ROP for the current conditions.
The result is now not simply a prediction but a controlled change to help drill faster based on real-time conditions.
Summary
Well drilling is one of the most expensive and environmentally taxing aspects of the upstream oil and gas industry. Machine learning can help companies to increate their drilling rates by taking in real-time information about mud logs and drilling in order to optimize the drilling configuration. The result can have a staggering impact on the economic success and environmental impact of an oil well.
Elipsa is a leader in delivering fully automated, scalable AI solutions for industrial IoT applications. Elipsa’s Intelligent Monitoring Platform seamlessly deploys across any workflow on the edge or in the cloud, increasing the availability and output of critical equipment. Elipsa’s self-training AI models and bolt-on approach enable AI deployments that are simple, fast, and flexible. To learn more, please visit https://www.elipsa.ai