Application of Machine Learning to Study Frac-Hit

Yan, Mingyu (New Mexico Institute of Mining and Technology) | Rahnema, Hamid (New Mexico Institute of Mining and Technology) | Shabani, Babak (Oklahoma State University)



The effect of frac-hit among the stimulated horizontal wells located in the northwest of the State of New Mexico are identified by addressing how to predict whether or not a planned well caused frac-hit for older wells nearby, and in case of the frac-hit occurrence, how to predict the degree of impact. The machine learning method is used to find the relationship between well parameters such as distance and age difference, and frac-hit occurrence and the degree of impact. Determining the probability of frac-hit occurrence is considered as a classification problem, and random forest method is used to predict the occurrence of the frac-hit. Predicting the impact of the frac-hit is considered as a regression problem, and two machine learning methods, gradient boosting and adaptive boosting (AdaBoost), are used to solve this problem. In the pool of data, the data are randomly assigned to train and test set for unbiased machine learning.

The data of the training set are put into the random forest classifier to find whether the distance, age, age difference, and bearing have any impact on the occurrence of the frac-hit. Among these four factors, the bearing has the most significant impact, which means that the weight of bearing in classification process is higher than the other parameters, followed by the distance as the second important factor. Applying the trained random forest classifier on the test set data gives 78% correct outcomes compared to the actual frac-hit data in the test set.

Considering the change of oil production due to frac-hit as the indicator to measure the degree of impact in gradient boosting and AdaBoost algorithm shows that the bearing between wells is not an influential parameter in the regression problem compared to the classification problem. In other words, if the well has already experienced the frac-hit, the importance of bearing decreases, and the distance, age difference, and age of the wells become more prominent factors. The analysis shows that the average error between the actual data and the predicted results by gradient boosting and AdaBoost is about 40%.

The results of this paper can be used by the hydraulic fracturing operators to pre-determine the frac-hit probability and its impact on existing offset wells. It can also help to refine well design strategies to minimize the risk of potential well interferences.

  Country: North America > United States (1.00)
  Industry: Energy > Oil & Gas > Upstream (1.00)