High-Quality Synthetic Dataset Generation for Enhanced Anomaly Detection of Hydrate Plugging in Oil Wells

This paper details a systematic procedure for generating a high-quality synthetic dataset designed to train anomaly detection classifiers for hydrate plugging events in oil wells. Building on a validated hydrate deposition model within a commercial multiphase flow simulator, the study augments an existing dataset— the 3W dataset—with high fidelity simulated cases. In particular, the use of multiphase flow simulations with inherent hydrate formation, transport and deposition mechanisms, is aimed at generating synthetic data to address the following points: small or insufficient number of instances in the dataset, class imbalance (number of positive versus negative instances), and the physics-based features availability, in addition to sensor data, to enhance the quality of the dataset with the purpose of training machine learning models. This work focuses on specific production scenarios as base cases upon which anomalies may develop across diverse operational conditions, including steady state production, slugging, rapid choke restriction and hydrate plugging events. The base dataset demonstrates strong agreement with field observations. Building upon those base cases, a parametric matrix of cases is generated with the multiphase flow simulator with varying conditions such as choke positions, gas lift rates, GOR and BSW. Then, a set of new cases with different (synthetic) well geometries is specially designed to further multiply the set of conditions in the dataset. The strategic augmentation of real field data with simulated cases not only improves dataset size and balance but also offers novel additive information that significantly enhances machine learning-based anomaly detection, supporting more effective risk management strategies in oil production operations.