Auto-Navigation of Optimal Formation Pressure Testing Locations by Machine Learning Methods

ABSTRACT

Formation pressure testing provides important information for exploration and production activities. Accurate reservoir pressure measurements are necessary to help ensure that a well is drilled safely and to identify and evaluate the potential and value of the discovery. The interpretation of pressure gradients provides the reservoir compartmentalization structure of a well, oil-gas-water fluid contacts, and can indicate compositional grading, as evidenced by second order density changes. However, this is based on the assumption that the pressure testing quality is sufficient for high resolution analysis. Unfortunately, obtaining quality data from formation testing can be difficult and prolonged. Locations initially selected for formation pretesting along the wellbore are often not optimal, and the time spent conducting pressure testing on those locations is wasted. Specifically, in this 23-well study, only 57% of the locations selected were high quality, meaning that 43% of the pressure test locations were of suboptimal quality. Further, pressure testing of the suboptimal locations requires twice as much time as for high quality locations. Therefore, considerable operational time savings can be realized if low quality locations can be avoided. If only the optimal locations are chosen for pressure testing, then data quality can be significantly improved to reduce uncertainty during formation evaluation.

A multivariate machine learning method is presented that builds a statistical correlation between the formation pressure test quality index and conventional wireline logging data. Primarily, the model is constructed from triple combo log data to a pressure test quality index. The procedure begins with logging data extraction and preparation. Both conventional wireline logging data and corresponding pretest data from 20 wells in one region are obtained. Data preprocessing and missing data estimation were conducted to help ensure the sample number of the conventional wireline logs matches the sample number of the pretest. Each well contains multiple pretest data, which results in a total of 935 samples (conventional logs and pretest pairs) for machine learning model development and validation. After the dataset preparation is completed, various learning algorithms were explored with an optimal learning algorithm selected to create the final model. The model is then applied to three additional case studies for independent validation, including (1) an easy pressure testing job, (2) a typical pressure testing job, and (3) a difficult pressure testing job. The time savings and quality improvement are shown for each. The novelty of the new machine learning model lies in the ability to predict the quality index log for the formation pretest based on previous conventional wireline logging data and to guide the wireline engineer to select the locations along the wellbore at which to conduct the pretest. This method reduces the number of unsuccessful pretest locations along the wellbore from 43% on average to between 15 and 5% for the three case studies, which results in significant time savings.