APPLICATION OF SUPERVISED CLASSIFICATION AND TIME-SERIES MODELS ON PREDICTION OF UNDESIRABLE EVENTS IN OFFSHORE OIL PRODUCTION

Abstract

AbstractOffshore oil production is faced with critical challenges due to rare but high-impact undesirable events that disrupt well operations. The early prediction of events such as sudden water breakthroughs, valve failures, flow instabilities, and blockages is considered essential for preventing production losses, environmental incidents, and safety hazards. Traditional physics-based models are often limited in handling the complex, multivariate nature of these problems, whereas supervised machine learning and time-series modeling are increasingly applied to provide data-driven solutions for detecting subtle patterns that occur before failure. In this thesis, the application of supervised classification algorithms and temporal models has been investigated for the purpose of predicting undesirable events in offshore oil wells. A publicly available dataset (3W), which contains over 50 million sensor readings from an offshore field with eight documented types of production anomalies, has been used as a benchmark. A detailed study of the production challenges represented in the data has been carried out, including anomalies such as Abrupt Increase of BSW (Basic Sediment and Water), Spurious DHSV closures, Severe Slugging, Flow Instability, Rapid Productivity Loss, Quick Choke Restrictions, Scaling, and Hydrate formation, along with their operational relevance. The structure of the dataset, the origins of the sensor measurements, and the labeling of instances have been reviewed in order to inform feature engineering and model development strategies. Several supervised learning methods – including k-Nearest Neighbors (KNN), Decision Trees (DT), Random Forests (RF), and Artificial Neural Networks (ANN) – have been implemented and compared. Additional classifiers such as Support Vector Machines (SVM) and ensemble boosting methods have also been used to evaluate classification accuracy and robustness. In parallel, time-series modeling techniques have been applied to capture the temporal dependencies present in sensor data. Approaches such as recurrent neural networks (e.g. LSTM autoencoders) have been examined for their ability to predict the early stages of events based on time-dependent sensor information. To support the development of realistic time-series models, simulated datasets have been transformed using autoencoder-based distortion techniques. These distortions have been learned from real well signals and applied to simulated data in order to preserve the dynamic characteristics of operational behavior while introducing realistic anomalies. Recurrent neural network models, including Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures, have been applied under both multiclass classification and regression-based frameworks. In particular, a regression-based approach has been used to forecast class transitions by treating normal (class 0), transient (class 101), and anomaly (class 1) states as sequential time steps—representing past, current, and future behavior. This structure has been designed to simulate the natural progression of events in the wells. Instead of directly classifying the well state, the model has been trained to predict a numerical value representing the next probable class, which has then been post-processed and mapped back to discrete class labels. By adopting this method, a forward-looking prediction of system behavior has been made possible, offering the potential for earlier anomaly detection. The findings highlight the advantages of integrating supervised classification with time-series modeling and realistic data transformation. The combination of these methods has contributed to a more complete understanding of offshore well conditions and provided a practical foundation for building intelligent systems capable of monitoring operations and supporting decision-making in oil production environments.