Anomaly detection for geological carbon sequestration monitoring

Abstract

Geological carbon sequestration (GCS) is a method to reduce the emissions of CO₂ into the atmosphere. During GCS operations CO₂ is captured from the atmosphere or industrial activities and stored in geological formations for permanent storage. Monitoring is an important element of GCS because it ensures that the stored CO₂ remains safely contained in the intended formation during the long term. Additionally, monitoring wells can help to detect CO₂ leaks, prompt remediation actions, and provide valuable information to optimize storage by monitoring the behavior of the CO₂ over time. In this work, we propose a method for GCS anomaly detection based on an LSTM Autoencoder Neural Network and Isolation Forest. The LSTM-Autoencoder uses the monitor Bottomhole Pressure (BHP) response while CO₂ is being injected into a geological structure. To account for the subsurface uncertainty, multiple subsurface model realizations are created, and using reservoir simulation, the multiple monitor BHP are generated to capture the subsurface uncertainty. Anomaly BHP points are detected using the residuals of the LSTM-Autoencoder and Isolation Forest. Additionally, an anomaly score based on the subsurface uncertainty is proposed. Finally, the method robustness is evaluated using point outliers, level shift outliers, and transient shift outliers as anomaly BHP signals. Early detection of abnormal BHP pressure signals can indicate the presence of subsurface fractures, faults, or leaks. Consequently, the correct detection of anomaly points in the pressure signals is of great importance.