A multichannel borehole‐to‐surface controlled‐source electromagnetic experiment was carried out at the onshore CO2 storage site of Hontomín (Spain). The electromagnetic source consisted of a vertical electric dipole located 1.5 km deep, and the electric field was measured at the surface. The subsurface response has been obtained by calculating the transfer function between the transmitted signal and the electric field at the receiver positions. The dataset has been processed using a fast processing methodology, appropriate to be applied on controlled‐source electromagnetics (CSEM) data with a large signal‐to‐noise ratio. The dataset has been analysed in terms of data quality and repeatability errors, showing data with low experimental errors and good repeatability. We evaluate if the induction of current along the casing of the injection well can reproduce the behaviour of the experimental data.

No abstract

No abstract

No abstract

There are two main methods to incorporate uncertainties in the production profiles. The first option relies on multiple geomodels and reservoir simulations that require long engineering time but help to assess accurately the low and high case profiles integrating the main geosciences and reservoir uncertainties. The second option is to use Decline Curve Analysis (DCA) with a reduced engineering time. Nevertheless, this method relies solely on production data, ignoring the geology and various production mechanisms. Process automation and machine learning workflows can then be of great help in this second case but remain generally limited, lacking transverse features for the geology and production mechanisms. In this paper we will present a hybrid approach where the low and high case production profiles were stochastically generated, derived from a deterministic basecase model simulation. This approach was tested on a giant carbonate field in the Middle East with a few million grid cells and hundreds of wells. The study was provided with a basecase history matched model and its associated forecast run. The interest of the method is to save time as the hybrid approach gives faster results compared to existing methodology while including geological and dynamic features to generate low and high cases for production profiles Machine learning based workflows were developed to calculate and extract from 3D timelapse models and 2D results all the important parameters (well features) driving well performance in the history but also in the forecast. Those important features were used to train a machine learning random forest model to predict the cumulative production (Np) of the wells. In addition to the ability to predict the Np, the model also gives the impact of every parameter on the cumulative production, allowing a ranking of the most impacting production mechanisms and geological parameters. The distance to the waterfront at the time the well was drilled was identified as the major parameter impacting the cumulative production; this feature and other key parameters were therefore used to generate the low and high cases. The final outputs from the study were the delivery of the low and high production profiles, taking into consideration the main production mechanisms and geological uncertainties identified and turned into features for the workflow. This work was done thanks to a multidisciplinary team composed of Reservoir engineers and datascientists from the FRF team.