Automatic Formation Pressure Test Classification Through Machine Learning - The Solution for a Clean and Reliable Database

Abstract Objectives/Scope Wireline formation test tools provide formation pressure throughout every phase of field development. It gives also fluid mobility, formation fluid type and connectivity. Classification of pressure tests and interpretation of data are crucial to using data effectively and efficiently. In certain cases, especially in low mobility reservoirs, the classification and interpretation depend on field/reservoir engineers’ expertise and experience. These interpretations are subjective, and thus need to be reviewed. In this paper, we proposed a machine learning workflow to classify all pressure data, including the difficult cases. The workflow was able to distinguish good quality formation pressure measurements from tight / invalid points for the vast majority of the dataset. It flagged pressure test which were ambiguous and needed further operator review. Methods, Procedures, Process The proposed workflow includes a first step of pressure test index identification, a second step of pretests visualization and a third step of automatic classification. In the index identification phase, given a pressure test station job, the starting and ending time of pressure buildup is identified. In the visualization phase, the buildup pressure curves are represented in a 2-dimensional space. The closer they are in the 2-dimensional space, the shape of the buildup curves are more similar. An ambiguous zone is defined based on the visualization, where the mix of two or more classes appears. The pressure tests in the ambiguous zone are added to the quality check list. In the automatic classification phase, a machine learning algorithm based on K-Nearest Neighbour (KNN) and Light Grading Boosting Machine (LGBM) is applied to perform automatic classification. The pressure tests on which the algorithm make incorrect prediction are added to the quality check list. After the pressure tests in the quality check list are reviewed, the visualization phase and the automatic classification phase are repeated until the classification result reaches 100% accuracy, or no new pressure test can be added to the check list. Results, Observations, Conclusions A case study was performed on an extensive formation pressure dataset from PDO a total of 379 pressure tests. The pressures were classified both in the traditional way by an analyst, as well as through the machine learning workflow. The vast majority of the dataset was classified correctly, and only 12 pressure tests were added to the quality check list for manual interpretation. The pressure test interpretation quality check time was reduced by 90%. Novel/Additive Information Once the pressure test interpretations are quality checked, future work includes designing a machine learning model to perform highly reliable pressure test quality classification and real-time pressure test quality identification. The realtime identification supports early decision making for reservoir engineers on whether they can get a stabilized pressure in reasonable operation time.