Inside nanopores, solid-fluid interactions are of the same order of magnitude as intermolecular interactions of fluid molecules. This fact strongly modifies the thermodynamic properties of confined fluids with respect to bulk phases. Tight oil and shale gas reservoirs, where the proportion of micro (below 2 nm) and mesopores (between 2 and 50 nm) can reach more than 20% of the volume distribution, represent an environment with such problems and industrial challenges to hydrocarbon fluid pressure/volume/temperature (PVT) modeling. This study provides a detailed understanding of the thermodynamic behavior of confined fluid and reference data of the thermodynamic properties of pure components (methane, ethane, n-pentane, n-decane) and mixtures (methane/ethane, ethane/n-pentane) confined in graphite slit pores. Furthermore, a detailed explanation of the different pressures considered in a porous medium with nano-pores is given. The Gibbs Ensemble Monte Carlo (GEMC) NVT simulation is used for pure components instead of the more traditional Grand Canonical Monte Carlo ensemble (GCMC) simulation to get more precise results of liquid and vapor confined pressure avoiding the phase change location determination problem. The evolution of critical temperature and pressure versus pore radius is compared to literature correlations and confined vapor and liquid densities are calculated. A new and robust method in the GEMC ensemble called GEMC NPT Bubble point Monte Carlo (BPMC) completed with GEMC NVT simulations has been developed to get thermodynamic properties including pressures at equilibrium of confined mixtures. Pressure versus density diagrams, pressure versus molar fraction isotherms and examples of pressure versus temperature diagram for a specific composition are built. The phase envelope of the confined fluid is shifted and closes with respect to phase envelope of bulk fluid. The critical temperature and pressure are shifted from the bulk value to a lower value and the bubble point pressure is decreased as the dew point pressure is increased. With regards to the selectivity of the confined system compared to the bulk fluid, for the methane/ethane and ethane/n-pentane mixtures, the heavier component is preferentially adsorbed in the vapor phase and the lighter component is preferentially adsorbed in the liquid phase. All these results for pure components and mixtures provide relevant information concerning the understanding of the phase behavior in confined systems such as shale gas and tight oil reservoirs, emphasizing the difference from the bulk fluid. Furthermore all these data may be used as references for pore radius dependent equation of state (EOS) calibration.

The flash calculation with large capillary pressure often turns out to be time-consuming and unstable. Consequently, the compositional simulation of unconventional oil/gas reservoirs, where large capillary pressure exists on the vapor-liquid phase interface due to the narrow pore channel, becomes a challenge to traditional reservoir simulation techniques. In this work, we try to resolve this issue by combining deep learning technology with reservoir simulation. We have developed a compositional simulator that is accelerated and stabilized by stochastically-trained proxy flash calculation.We first randomly generated 300,000 data samples from a standalone physical flash calculator.We have constructed a two-step neural network, in which the first step is the classify the phase condition of the system and the second step is to predict the concentration distribution among the determined phases. Each network consists of four hidden layers in between the input layer and the output layer. The network is trained by Stochastic Gradient Descent (SGD) method with 100 epochs.With given temperature, pressure, feed concentration pore radius, the trained network predicts the phases and concentration distribution in the system with very low computational cost. Our results show that the accuracy of the network is above 97% in the metric of mean absolute percentage error.The predicted result is used as the initial guess of the flash calculation module in the reservoir simulator. With the implementation of the deep learning based flash calculation module, the speed of the simulator has been effectively increased and the stability (in the manner of the ratio of convergence) has been improved as well.

Tight oil and shale gas reservoirs have a significant part of their pore volume occupied by micro (below 2nm) and mesopores (between 2 and 50nm). This kind of environment creates strong interactions forces in the confined fluid with pore walls as well as between its own molecules and then changes dramatically the fluid phase behavior and its thermodynamic properties. Pressure-Vapor-Temperature (PVT) modeling of such fluids becomes therefore a challenge in order to get accurate production forecast reservoir simulations. Furthermore along the flow from the matrix to the well through the fractures, the fluid will pass through a very heterogeneous pore size distribution which will alter it differently according to the pore size and the spatial distribution. An important work has therefore to be done on developing upscaling methodology of the pore size distribution for large scale reservoir simulations. Firstly molecular simulations will be performed on pure components and mixtures in order to get reference thermodynamic properties at liquid/vapor equilibrium for different pore sizes. The comparison with commonly used modified equation of state (EOS) in the literature highlighted the model of flash with capillary pressure and critical temperature and pressure shift as the best one to match reference molecular simulation results. Secondly fine grid matrix/fracture simulations have been built and performed for different pore size distributions. The study has shown that the pore size distribution has an important impact on reservoir production and that this impact is highly dependent of the volume fraction of nanopores inside the matrix. Capillary pressure heterogeneity and pore radius dependent EOS cause gas flow slowdown or gas trapping inside the matrix and postponed gas flow apparition in the fractures during depletion which reduce the GOR (Gas-Oil Ratio) at the well. Coarse grid upscaling models have then been performed on the same synthetic case and compared to the reference fine grid results. The commonly used upscaling methodology of dual porosity model with average pore radius for the pore size distribution is unable to match the fine grid results. A new triple porosity model considering fracture, small pores and large pores with their own capillary pressure and EOS, together with MINC (Multiple Interacting Continua) approach, has shown very good match with the reference fine grid results. Finally a large scale stimulated reservoir volume with different pore size distribution inside the matrix has been built using the upscaling method developed here. The proposed triple porosity methodology is able to model the PVT of the confined fluid and its flow across a very heterogeneous pore size distribution up to the well through fractures in a large scale reservoir simulation.

No abstract

In this work, we present the development of a compositional simulator accelerated by proxy flash calculation. We aim to speed up the compositional modeling of unconventional formations by stochastic training. We first developed a standalone vapor-liquid flash calculation module with the consideration of capillary pressure and shift of critical properties induced by confinement. We then developed a fully connected network with 3 hidden layers using Keras. The network is trained with Adam optimizer. 250,000 samples are used as training data, while 50,000 samples are used as testing data. Based on the trained network, we developed a forward modeling (prediction) module in a compositional simulator. Therefore, during the simulation run, the phase behavior of the multicomponent system within each grid block at each iteration is obtained by simple interpolation from the forward module. Our standalone flash calculation module matches molecular simulation results well. The accuracy of the trained network is up to 97%. With the implementation of the proxy flash calculation module, the CPU time is reduced by more than 30%. In the compositional simulator, less than 2% of CPU time is spent in the proxy flash calculation. The novelty of this work lies in two aspects. We have incorporated the impacts of both capillary pressure and shift of critical properties in the flash calculation, which matches molecular simulation results well. We developed a proxy flash calculation module and implemented it in a compositional simulator to replace the traditional flash calculation module, speeding the simulation by 30%.