Geologic Success but Economic Failure: Uncovering Hidden Problems Using Recursive Partitioning

Jeffrey M. Yarus¹, R. Mohan Srivastava², and Richard L. Chambers^3
1 Quantitative Geosciences, Houston, TX
² FSS Canada, Toronto, ON
³ Quantitative Geosciences, LLP, Broken Arrow, OK

Production from six wells producing from fractured carbonates was subjected to a data mining analysis due to their suspected anomalously poor production performance. The objective was to determine if the observed production behavior was statistically unlikely through a comparison with other wells having similar geologic and engineering properties. Further, if the behavior of these wells was found to be anomalous, the study hoped to identify the cause.

Geologic and engineering data were collected from 419 wells in the study area and subjected to recursive partitioning (RP), a multivariate statistical technique. RP creates a tree of questions that classify a training data set into several small subsets. In this case, the training set consists of all the data except from the suspect wells. Once the tree is created, it can be used for prediction by taking a new observation, in this case each suspect well, classifying it according to the same set of questions, and using the behavior of the samples in the same node to predict its behavior. In this way, RP extracts from a training data set the “closet cousins” most similar to the suspect wells. In this study, the tree of questions relate to the criteria that influence production. The resulting closest cousins will be the set of wells, regardless of their geographic location, which share the same criteria. A comparison is then made between the suspect wells and their closest cousins in order to determine the probable or improbable nature of production from the wells in question.