Development of a Neuro-simulation Tool for Coalbed Methane Recovery and CO2 Sequestration

The increase of atmospheric CO2 levels due to releases from anthropogenic sources is an important environmental issue. One of the most promising technologies to mitigate emissions of this greenhouse gas is to sequester the CO2 in geological formations for long time periods. For this purpose, coal seams are considered to be very attractive geological horizons because sequestration costs can be partially offset by methane production. However, optimization of a coal seam sequestration project is a complicated process, requiring accurate and efficient methods and tools for predicting methane recoveries and CO2 storage capacities. Because of the complex nature of coal seams, fluid flow through them, and other coal/fluid interactions, many parameters affect the outcome of a sequestration project. Our previous studies have indicated that at least fourteen parameters (e.g., face cleat and butt cleat permeabilities, Langmuir sorption constants of carbon dioxide and methane, sorption time constant) are important. Well patterns and injection and production pressures are parameters that can be controlled in attempts to optimize rates and amounts of carbon dioxide sequestration and methane production. Because of the large number of parameters involved, optimization of a field project design can require thousands of reservoir engineering simulations. This paper outlines the application of a neurosimulation methodology to predict the important performance indicators of a coal seam sequestration project. Although coal-seam sequestration projects are now being started in the United States, Japan, Poland, and elsewhere, very few data from mature field projects exist in the public domain. Therefore, a compositional coalbed methane simulator, was used to model the enhanced recovery of methane by CO2 injection. Results from the simulator were used to train artificial neural networks, utilizing the back-propagation learning algorithm. Neural networks were trained to predict the following production performance indicators: breakthrough time, cumulative CO2 injection, and cumulative CH4 production. Both the ability of the neural networks to fit the data sets used to train the neural network and the ability of the trained networks to predict other data sets were examined. The accuracy of the network’s predictions was thus successfully established. The networks trained in this work are capable of predicting project performance both quickly and accurately. They allow us to find the results of any of thousands of simulations that have been performed, and to rapidly predict results from combinations of parameter values that have never been simulated, without performing additional reservoir engineering computations.