A machine learning model for predicting multi-stage horizontal well production
Ilia Chaikine, Ian D. Gates
Abstract
In this study, a hybrid convolutional-recurrent neural network (c-RNN) is evaluated for making predictions of the five-year cumulative production profiles in multistage hydraulically fractured wells. The model was trained by using a combinations of completion parameters, rock mechanical properties, and well spacing and completion order for each stage of 74 wells in the Montney Formation in Alberta. The prediction accuracy of the various combinations was measured by using the mean average percent error and mean absolute error generated through the leave-one-out method. The best combination of inputs was found to be the rock mechanical properties surrounding each perforation cluster, the proppant amount used for every stage, and the spacing and completion order of neighboring wells. The novelty of this study is that the input variables used are at the stage level rather than the average of the entire well. The accuracy of the model was found to increase exponentially as the production of multiple wells was aggregated. The approach yields insights for planning new well drills in fields with existing development since it provides the ability to run multiple field development scenarios without having to spend capital.