We have developed a Petroleum Analytics Learning Machine & trade (PALM) that forecasts production as the attribute list of data available from geology and geophysics, to rock physics, reservoir modeling, drilling, hydraulic fracturing, grows over time until first oil and gas production. PALM provides a multi-dimensional analytics system that inputs all available data into an ensemble of machine learning and statistical algorithms that compute first the Importance Weights of each attribute to overall field production, and then convolves these weights with actual attribute values for each well to predict its ultimate production of oil, gas, and water, all before first oil and gas is produced to the surface. The data are made available via a Systems Integration Database (SID) that combines all available attributes. We implemented and tested PALM using data from 156 horizontal wells drilled from 2009 through 2013 in the wet gas area of the Marcellus shale. We were able to identify 72 attributes and data from 1,842 fracs common to all of these wells. We derived condensate, gas and water production predictions from rock properties, logging, hydraulic fracturing, landing zones, and stress attributes. We then tested for accuracy by building Tornado plots first using only completions data attributes, then adding landing zones, stress, rock properties and reservoir modeling. When predicting top versus bottom quartiles of production, PALM progressed from 78% accuracy using only firac attributes to 97% accuracy in identifying the good producers from the bad producers utilizing all available data. We then used Area Under the Curve (AUC) of a Receiver Operator Characteristics (ROC) plot to predict that the Importance Weights would average 90% accuracy for the 156 wells. Using the 72 attribute Importance Weights convolved with attribute values from each well, we then determined that the actual accuracy when compared to the real production was 94% with plus/minus 6% error. We conclude that the error rate can be effectively minimized so that the Importance Weights result in useful predictions for future production, before first oil and gas are produced to the surface.
Petroleum Analytics Learning Machine to Forecast Production in the Wet Gas Marcellus Shale
- Share Icon Share
- Search Site
Anderson, Roger N., Xie, Boyi, Wu, Leon, Kressner, Arthur A., Frantz, Joseph H., Ockree, Matthew A., and Kenneth G. Brown. "Petroleum Analytics Learning Machine to Forecast Production in the Wet Gas Marcellus Shale." Paper presented at the SPE/AAPG/SEG Unconventional Resources Technology Conference, San Antonio, Texas, USA, August 2016. doi: https://doi.org/10.15530/URTEC-2016-2426612
Download citation file: