Inspiration

Predicting peak oil rate seems very challenging due to the complex nature of both oil wells and their environment. We were inspired to tackle this problem with a non-linear approach.

What it does

Our random forest regression model predicts the peak oil rate production given certain characteristics of an oil well, specified by Chevron.

How we built it

We visualized data to determine what features to drop and trained the data on linear and non-linear models to identify the model best suited for this problem.

Challenges we ran into

The biggest challenge we ran into was the amount of missing data in the provided dataset. As we were unable to reasonably 'fill in' the missing data, we decided to remove any datapoints that had a missing value for one of the features we decided to use. As a result, we had to carefully decide what features we wanted to include when training our model. We also ran into challenges of deciding what model architecture best fit our use case, which we solved by reasoning that the data followed a non-linear relationship.

Accomplishments that we're proud of

We are proud to achieve an RMSE of 79 and an r^2 value of 0.66 for our random forest regressor, the RMSE being quite low given the scale of peak oil production rate in the thousands.

What we learned

We learned the importance of carefully selecting features which impact the correctness of our machine learning model, and we also learned about the power of random forest regression in predicting complicated relationships that we see in the data.

What's next for Predicting Peak Oil Rate

In the future, we'd like to talk with industry experts to better understand the meaning of our features to build a model architecture more in line with domain knowledge.