Machine Learning Provides Effective Leak Detection in Carbon Sequestration Projects

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 201552, “Leak Detection in Carbon Sequestration Projects Using Machine Learning Methods: Cranfield Site, Mississippi, USA,” by Saurabh Sinha, SPE, University of Oklahoma and Los Alamos National Laboratory; Rafael Pires De Lima, Geological Survey of Brazil; and Youzuo Lin, Los Alamos National Laboratory, et al., prepared for the 2020 SPE Annual Technical Conference and Exhibition, originally scheduled to be held in Denver, 5–7 October. The paper has not been peer reviewed. Saline aquifers and depleted hydrocarbon reservoirs with good seals located in tectonically stable zones make an excellent storage formation option for geological carbon sequestration.Ensuring that carbon dioxide (CO2) does not leak from these reservoirs is the key to any successful carbon capture and storage (CCS) project. In the complete paper, the authors demonstrate automated leakage detection in CCS projects using pressure data obtained from the Cranfield reservoir in Mississippi in the US. Results indicate that even simple deep-learning architectures such as multilayer feed-forward neural networks (MFNNs) can identify a leak using pressure data. Introduction Several methods that use different types of data currently are available to detect leaks. Although some of the methods are a direct indicator of CO2 presence, they cannot provide an early warning for the leaks, thus delaying remedial measures. An ideal process for the identification of leakages requires constant and repetitive comparisons of different data. Machine-learning (ML) techniques are ideally suited for this task. In this work, the authors demonstrate the use of ML techniques such as linear model, random forest, and MFNN on time-series signals obtained from a pressure-pulse test. The methodology uses the time-series data instead of 2D images or 3D voxels, thus providing a computational advantage. The authors write that an ML algorithm can distinguish between a pressure signal corresponding to a leak vs. the pressure signal corresponding to a baseline nonleak case. The trained models can then be used as an early-warning system to flag anomalous data to then be analyzed by a human interpreter. Background A pressure-pulse test uses at least two wells: an injection well and a monitoring well. The reservoir is then shocked by a series of predetermined cycles of injection and shut-ins (i.e., a pulse). The response then is recorded at the monitoring well with a pressure gauge that measures the target formation pressure. The test may be repeated with different pulses to understand the reservoir properties better. A harmonic pulse is preferred over a square wave because it allows for spectral decomposition of the pulse to analyze the reservoir response at different frequencies. Three wells are used in the study: F1, F2, and F3. Well F1 is the injector well, where alternative cycles of injection of CO2 and shut-in are carried out. Well F2 is the monitor well, which remains shut in for the duration of the test and where the pressure is monitored with the use of a pressure gauge. An artificial leak is simulated in the test by opening a surface valve at Well F3.