Shale heterogeneities often impede the development of steam chamber in many steam-assisted gravity drainage (SAGD) projects. Unfortunately, static data alone is generally insufficient for inferring the corresponding distribution of shale barriers. This study presents a novel data-driven modeling workflow, which integrates deep learning (DL) and data analytics techniques to analyze production profiles from horizontal well pairs and temperature profiles from vertical observation wells, for the inference of shale barrier characteristics.
Field data gathered from several Athabasca oil sands projects are extracted to build a set of synthetic SAGD models, where the geometries, proportions and spatial distribution of shale barriers are modeled stochastically. Numerical flow simulation is performed on each realization; the corresponding production/injection time-series data, as well as temperature profiles from one vertical observation well, are recorded. A large dataset is assembled for the development of data-driven models: wavelet analysis and other data analysis techniques are performed to extract relevant input features from the temperature and production profiles; a novel parameterization scheme is also proposed to formulate the output variables that would effectively describe the detailed distribution of shale barriers. DL, such as convolutional neural network, together with other data analytics techniques are applied to capture the complex and nonlinear relationships between these input and output variables.
The feasibility of the developed workflow is validated using synthetic test cases. Salient features capturing the impacts of shale barriers are extracted. It is observed from the production time-series data that, as the steam chamber approaches a shale barrier, a decline pattern is noticeable until the steam chamber advances around the shale barrier. An obstruction in the steam chamber development can also be noted in the temperature profiles, as steam is trapped by shale barriers that are located reasonably close to the horizontal well pair. This observation is confirmed by comparing the petrophysical logs and the temperature profiles at the observation wells. Analyzing both temperature and production data could help to infer the size of shale barriers in the inter-well regions. Finally, the model outputs are used to generate an ensemble of heterogeneous SAGD realizations that correspond to the input production and temperature time-series data.
This study offers a complementary and computationally-efficient tool for inference of stochastically-distributed shale barriers in SAGD models, which can be subjected to detailed history-matching workflows. It is the first time that data-driven models are used to analyze both production data from horizontal production well pairs and temperature profiles from a vertical observation well for inferring SAGD reservoir heterogeneities. The results illustrate the potential for application of data analytics in reservoir modeling and flow simulation analysis. The developed workflow also can be extended to characterize reservoir heterogeneities in other recovery processes.