FAST AND ACCURATE PHASE EQUILIBRIUM CALCULATIONS FOR CONDENSATE SHALE GAS RESERVOIRS

Phase equilibrium calculations of complex fluids in shale gas reservoirs require the establishment of advanced numerical models that consider capillary effects, and the design of fast and reliable algorithms to handle the various components in the reservoir fluids in practical working conditions. In this study, we develop a thermodynamically consistent VT-type pore-scale flash calculation scheme based on realistic equations of state suitable for oil/gas reservoirs, e.g. the Peng-Robinson equation of state. The effect of capillarity has been incorporated in the scheme for a more accurate description of the thermodynamic properties of shale gas, and the diffuse interface model is applied to establish a dynamic evolution scheme in the phase equilibrium process, and a convex splitting method is used to model the evolution of compositional moles and volume. In order to accelerate the iterative flash calculations for realistic reservoir fluids containing a large number of components, a self-adaptive deep learning algorithm is developed in this paper with a novel structure to achieve wider applicability to various components in different fluids. The input and output features of the neural network are selected as the key thermodynamic features on the basis of thermodynamic analysis, and the network hyper-parameters have been carefully tuned to achieve a better performance on both accuracy and efficiency. Advanced deep learning technics resolving overfitting problems have been applied in our algorithm. The trained model significantly accelerates the conventional flash calculation based on iterative methods, while a good prediction accuracy has been preserved. Phase stability test and phase splitting calculations are automatically incorporated in our prediction, and we can significantly capture the effect of capillarity on phase equilibrium behaviors. Such a fast, accurate and reliable shale gas phase equilibrium calculation scheme using deep learning algorithms can provide an initial phase distribution field with physical meanings for subsequent multiphase flow simulations, while the number of phases can be also determined. The thermodynamic information and analysis can also be used as a thermodynamic basis for a multiphase numerical model with built-in physical conservation.