Using Data-Driven Predictive Analytics to Estimate Downhole Temperatures while Drilling

Abstract

Abstract The advent of complex drilling programs has propelled the oil and gas industry to explore avenues that help expand the existing realm of drilling technologies. Predictive data-driven analytics is one such domain that has driven massive interest, primarily because of its successful implementation in various other industries. The application of data-driven models to successfully predict the downhole environment should be the next step change in drilling operations to help optimize complex drilling programs. This study aims to use predictive data-driven modelling to predict an important drilling parameter—downhole temperature during any ongoing drilling activity. The proposed method uses an insightful selection of available surface data to train the data driven-model to predict the target downhole temperatures measured during a segment of the drilling activity. Model training and deployment are realized in two modes: individual and cumulative. In the individual mode, the downhole temperature of the current segment is predicted by the model built using the surface data obtained from the previous segment only. In the cumulative mode, the model is built using the surface data for all previous segments drilled up to the current segment. The trained model is then employed to predict the downhole temperature while drilling the next successive segment using that chosen set of surface data. The downhole temperature is predicted by a regression model implemented by a support vector machine classifier using a sequential minimal optimization algorithm. The predictive capability of the data-driven approach was demonstrated through its implementation in the drilling of two different wells using field case studies. The temperature measured using downhole tools during a given drilling segment was used to train the model, and the downhole temperatures to be encountered while drilling the next segment were estimated using surface parameters. The performance of the model was evaluated by a squared correlation coefficient and mean squared error for the predicted temperature and the actual value measured using downhole tools. In addition, the effect of model parameters on prediction and an analysis of the importance of different input attributes also were explored to aid in successfully choosing the appropriate drilling parameters for future implementation. The proposed data-driven modelling approach was able to successfully predict the downhole temperatures while drilling any well, using surface parameters for only the particular well under consideration. This method provides a unique fit-for-purpose perspective on data-driven analytics, without requiring a vast amount of data sets from a large collection of wells for prediction accuracy. On the contrary, it works on a well-by-well approach. Effective use of data-driven models would help predict the desired downhole parameters without the need of complex downhole tools and reduce costs of operation. In addition, they also can be implemented to estimate downhole phenomena that were inexplicable while drilling when using existing tools and techniques.