A Transformer-based approach for anomaly detection in intelligent well completions

Abstract

This study introduces a novel methodology and makes case studies for anomaly detection in multivariate oil production time-series data, utilizing a supervised Transformer algorithm to identify spurious events related to interval control valves (ICVs) in intelligent well completions (IWC). Transformer algorithms present significant advantages in time-series anomaly detection, primarily due to their ability to handle data drift and capture complex patterns effectively. Their self-attention mechanism allows these models to adapt to shifts in data distribution over time, ensuring resilience against changes that can occur in time-series data. Additionally, Transformers excel at identifying intricate temporal dependencies and long-range interactions, which are often challenging for traditional models. Field tests conducted in the ultradeep water subsea wells of the Santos Basin further validate the model’s capability for early anomaly identification of ICVs, minimizing non-productive time and safeguarding well integrity. The model achieved an accuracy of 0.954 4, a balanced accuracy of 0.969 4 and an F1-Score of 0.957 4, representing significant improvements over previous literature models.