Artificial neural networks for reservoir level detection of carbon dioxide seepage location using permanent down-hole pressure data.
Date of Graduation
Monitoring the potential seepage of CO2 is a required practice for any CO2 sequestration project. Monitoring is used in order to ensure the efficiency of the sequestration process and that the CO 2 will remain in the target formation for extended periods of time. Different techniques that exist in the literature differ based on their resolution and location of their application. Monitoring techniques can be divided into three categories: surface techniques, which measure CO2 at or above the surface, shallow depth techniques, which monitor the concentration of CO2 in soil or shallow aquifers, or deep techniques, which monitor CO2 movement at the reservoir level. The technique presented in this work can be placed in the third category. High frequency pressure data is collected from permanent down-hole gauges installed at the reservoir level. These gauges collect pressure data at high frequencies up to a measurement per second. The collected pressure data can be sent to an Artificial Neural Network (ANN). The trained ANN analyzes the pressure data received from several sensors installed in the reservoir and can detect whether seepage is occurring in the reservoir. The network compares these pressure measurements against a base model of the reservoir. A previously developed reservoir model has to be available with reliable predictive capabilities of the reservoir pressure. The network looks for changes in pressure measurements at the sensors. Once the pressure change exceeds a threshold, it starts searching for the possible location of the seepage. In the case of CO2 seepage, the ANN provides an approximate location and the amount of the CO2 seepage. An ANN trained for a heterogeneous reservoir (heterogeneity in porosity and permeability) could detect the location of the seepage with reasonable accuracy, as low as an area of 0.6 acres in a reservoir with a total area of around 579 acres. The seepage rate in such a reservoir was around 0.3% of the total stored gas per year, which was slightly above the 0.1% per year limit found in the literature. Along with the seepage detection network, two Surrogate Reservoir Models were developed for a hypothetical CO2 injection project. A Surrogate Reservoir Model (SRM) is a prototype of full-field reservoir simulation models. They are essentially ANNs, which once trained can mimic the behavior of the reservoir with change in selected input parameters. An important feature of SRMs is their fast analysis and generation of outputs in a very short time. SRMs can be used for uncertainty analysis and history matching. The developed SRM is used to provide a reasonable history match of the reservoir under study. This model then can be used for forecasting pressure distribution in the reservoir. The pressure difference between the actual field pressure distribution (recorded by the pressure sensors) and the SRM predictions at pressure sensorsâ€™ locations are then sent to the ANN trained for CO2 seepage location detection. The CO 2 seepage location detection ANN analyzes these data and predicts the location of the CO2 seepage.
Jalali, Jalal, "Artificial neural networks for reservoir level detection of carbon dioxide seepage location using permanent down-hole pressure data." (2010). Graduate Theses, Dissertations, and Problem Reports. 9095.