Machine learning for underground gas storage with cushion CO2 using data from reservoir simulation

Abstract

Underground natural gas storage (UNGS) is a means to store energy temporarily for later recovery and use. In such storage operations, carbon dioxide (CO2) can be injected as cushion gas to improve the operating efficiency of the working gas and then be permanently stored in the same reservoir. A potential obstacle for widespread use of this technology is that the mixing of the different gases can lead to undesired CO2 production. Herein, we use a two-component flow model to simulate injection and withdrawal periods of methane (CH4) in idealized reservoirs containing CO2. First, we simulate cases with a single well for both CH4 injection and production. From 1200 simulations with systematic variation of reservoir temperature, porosity, permeability, height, and injection time, we find that the reservoir height and permeability have the most significant impact on the production time until the well stream reaches 1% mole fraction of CO2. In another set of simulations, we investigate the impact of well spacing in seasonal gas storage scenarios with separate wells for CH4 injection and production, while CO2 injection occurs from a third well. Based on the simulated data we construct artificial neural networks (ANNs) that describe the relations between the varied input parameters and the production time of CH4, well-block mole fraction and pressure. We conclude that trained and validated ANN models are useful tools to optimize important parameters for UNGS operations, including well positioning, with the aim at maximizing the amounts of delivered gas.