Pre-processing noisy production data using outlier detection methods reduces uncertainty in identifying flow regimes

This study presents machine-learning-based outlier-removal techniques as a preprocessing step for noisy production data (rate and pressure). Applying outlier detection (OD) methods helps reduce uncertainty in flow-regime identification and preserve the unique profiles of these regimes. These methods are essential for addressing anomalies in production data and preventing inaccuracies in the extraction of reservoir properties. Flow-regime identification involves constructing diagnostic tools such as the rate-normalized pressure (RNP) plot, which is commonly used in rate-transient analysis (RTA). In this study, we simulated a synthetic log-log RNP profile for a fractured, low-permeability gas reservoir and grouped the RNP data by flow regime. We then added noise and outliers to selected RNP data groups across seven noise-presence scenarios using additive white Gaussian noise (AWGN) with three noise-contamination levels, followed by the OD process. We evaluated 20 OD methods using performance metrics including mean absolute error (MAE), accuracy, precision, recall, and F1 score. The F1 score combines precision and recall to assess the effectiveness of OD methods. Although a method may perform well in one scenario and less effectively in another, the Gaussian mixture model (GMM) consistently outperformed the other methods, achieving an average F1 score of 0.9. The optimal input parameters for GMM were a threshold factor between 0.8 and 1.1 and a component count between 7 and 10. We also performed sensitivity calculations for these input parameters using field data. In the synthetic cases, GMM produced cleaner RNP profiles and achieved an accuracy of 93.99%. This work expands previous studies of OD methods applied to RNP data by evaluating 20 methods across various noisy data scenarios stratified by flow regime. Our findings highlight robust OD methods that can reduce subjectivity and time consumption while improving the precision of flow-regime profiles, leading to more accurate production forecasting and EUR estimation.