Abstract
One of the challenges faced by companies in the oil and gas industry is the difficulty in assessing and quantifying subsurface uncertainties when planning for hydrocarbon exploitation. A commonly employed approach is to use available exploration and appraisal data to produce a range of possible subsurface realisations, through which hydrocarbon production forecasts are generated. Prediction of hydrocarbon production from these simulations are then used to assess the viability of a planned development concept and the associated subsurface uncertainties. However, due to the sparsity of field data and unpredictability of underground geology coupled with the typically large dataset sizes, the ability to rapidly quantify prediction uncertainty and provide an overview of the range of underground geologies leaves much to be desired.
Here we show that the application of a network science approach to oil production prediction data provides an intuitive way to visualize and assess reservoir uncertainty. A network transformation utilizing Pearson correlation and mean absolute error as similarity measures were applied to a dataset containing time series predictions of oil production for 10 wells simulated in 50 different subsurface realisations. Realisations were generated using a synthetic reservoir spanning 20 years of production. It was found that the network representation enabled the inference of reservoir uncertainty by simple visual inspection. Additionally, network measures such as the beta index were used with results supporting their viability in quantifying uncertainty. The application of clustering algorithms to the resulting networks was also shown to simplify the time series into component characteristic subsurface realisations.
We propose a method to quantify subsurface uncertainties and create a simplified representation of the characteristic modes associated with a range of subsurface realisations, greatly reducing the time required to conduct a cursory analysis. Our results demonstrate how the application of network science ideas may be applied to provide new ways of analyzing production data and performace predictions. These methods provide an opportunity to further refine descriptions of uncertainty by incorporating stochastic elements into the network as well.
