Warm solvent injection (WSI) has been proposed as a promising alternative to steam-based methods for bitumen recovery, due to its potential to reduce greenhouse gas emissions and environmental footprint. It involves injecting heated vaporized solvent (low molecular weight hydrocarbons) to reduce the viscosity of bitumen via solvent diffusion and latent heat transfer. However, similar to its thermal counterparts, the WSI recovery response is also highly sensitive to the underlying reservoir heterogeneities, e.g., shale barriers; in particular, the conformance of solvent chamber advancement can be a severe concern in heterogeneous reservoirs. Therefore, it is essential to approximate and monitor the development of solvent chamber during production and to optimize the operations design. Conventional monitoring methods, such as 4D seismic, can be quite costly. This work proposes a novel approach involving machine-learning techniques to efficiently track the solvent chamber positions in heterogeneous reservoirs.
First, a detailed sensitivity analysis is performed to examine the impacts of shale barriers on WSI production responses, which include the oil rate and the evolution of solvent chamber. A set of synthetic simulation models for the WSI process are constructed. Petrophysical, fluid and operational variables representing typical Athabasca oil sands conditions are assigned. Different configurations of shale barriers with varying sizes, correlation length and proportions are assessed. Next, a large training dataset consisting of many heterogeneous models and their simulation results are assembled. The inputs features are extracted from the oil production based on several time-series analysis methods; the output parameters are formulated to represent the dynamic evolution of solvent chamber. Different dimension reduction and parameterization strategies are formulated and tested to represent the solvent chamber locations and interfaces. Convolutional neural network is implemented to dynamically track the solvent chamber positions by correlating the extracted inputs and outputs.
The simulation results confirm that the presence of shale heterogeneities would impede the development of solvent chamber, causing a reduction in oil rate. In particular, the shale barriers that are located closer to the well pairs would exert a more severe impact on production responses than those that are located at further distances from the wells. The application of machine-learning algorithms enables the locations of the solvent chamber as a function of producing time to be inferred and tracked reliably. The proposed workflow provides a practical workflow to estimate the real-time solvent chamber development corresponding to the WSI process in heterogeneous reservoirs from oil production and solvent profile directly.
The presented workflow offers a novel alternative to infer the development of solvent chamber in heterogeneous reservoirs from production time-series data directly. This type of analysis could complement many existing monitoring techniques to deliver a more comprehensive inference of the distribution of shale heterogeneities in solvent-based bitumen recovery operations. Production data is used directly to assess the conformance of solvent chamber advancement, which is an important consideration in operations design and real-time optimization.