Predicting Trouble Stages with Geomechanical Measurements and Machine Learning: A Case Study on Southern Midland Basin Horizontal Completions

Abstract

Unexpected problems during completion create costs that can cause a well to be outside its planned AFE and even uneconomic. These problems range from merely experiencing abnormally high pressures during treatment to casing failures. The authors of this paper use machine learning methods combined with geomechanical, wellbore trajectory, and completion datasets to develop models that predict which stages will experience difficulties during completion. The operator collected geomechanical data for 26 lateral wells. Several of these wells experienced significant difficulties during completion, including casing failures. By examining the completion data for the study wells, engineers developed two objective proxies for trouble stages: a high fracture gradient cut-off and a sand-to-water ratio cut off. The geomechanical data were combined with wellbore trajectory and stage-level completion data. Machine learning techniques were then used to predict the fracture gradient at the end of treatment and the sand-to-water ratio of a stage. The data were subdivided by target horizon to ensure the model results were not skewed towards a specific horizon. The models were cross validated, hyper-parameter tuned, and then tested against a subset of the data withheld from the training algorithm (the test data subset). Multiple metrics were applied to measure the effectiveness of the models. For the fracture gradient regression, the model that minimized the root mean squared error (RMSE) was chosen. This model was compared to a prediction that each stage would have a fracture gradient equal to the average fracture gradient. The correlation between predicted and measured frac gradients was also calculated. These metrics were analyzed for both the entire data set and the test data subset. By all these measures the model performs significantly better than the control prediction. For the sand-to-water ratio metric, the model that maximized the receiver operating characteristic area under the curve (ROC-AUC or AUC) score was chosen. We chose to measure the success of the model with the precision of its prediction of outlier (trouble) stages on the test data subset. The model performs with a precision of at least 73%. Both models identify stages where casing failures occurred as trouble stages. With these initial results, the authors are confident they can predict the probability of any stage causing a significant operational disruption in these horizons. By predicting the probability of a trouble stage, mitigation plans can be implemented to reduce costs ahead of time. By further analyzing the models, engineers can identify the major drivers for the trouble stages and work to reduce them as early as the well planning stage.