Predicting Dysfunction Vibration Events while Drilling Using LSTM Recurrent Neural Networks
Abstract The objective of the study is to build a robust Recurrent Neural Network system using Long-Short-Term-Memory (LSTM) to predict future vibrations during drilling operations. This provides a reliable solution to the complex problem of modeling several forms of vibrations encountered downhole. This accurate prediction system can be readily integrated into advisory/warning systems giving drillers the potential to save time, improve safety, and increase efficiency in drilling operations. High-frequency downhole drilling data onshore fields, obtained from a major O&G service provider, was used to train and validate the models. First, multiple classification algorithms such as Logistic Regression, KNN, Decision Trees, Random Forest were utilized to identify the presence and severity of Stickslip, Whirl, and other drill-string vibrations. LSTM-RNN was then used instead of traditional RNN intended for sequential data, to resolve the vanishing gradient problem. LSTM-RNN architecture was built to predict vibrations a)10 seconds and b) 30 seconds into the future. Results of the traditional classification models confirmed the hypothesis that dysfunctions can be successfully identified based on real-time downhole drilling data. 98% accuracy was obtained in successfully identifying torsional vibrations during drilling. A total of 101 parameters including measured and derived variables are available in the dataset. Modeling was performed with 14 features and vibrations were predicted. The RNN model was trained on data from multiple wells that encountered vibrations during drilling. The models were able to predict vibrations 10 seconds into the future with an MSE of 0.02 and 30 seconds into the future with reasonable accuracy and MSE of 0.10. Avoiding excessive vibrations will result in fewer trips by increasing longevity and reducing malfunctions of downhole electronics, the drill-string, and the BHA. Reduced NPT means drilling complex wells efficiently in less time which in turn directly translates to lower costs for the company. In addition to significant cost benefits, automated technology predicting anomalies and reacting in real-time translates to improved safety because it would now require fewer operators at risk on the rig floor. The work opens up avenues for a sophisticated advisory/warning system and effective ‘look-ahead’ drilling processes in the future.
