Author(s): |
Gabriele Martinelli, NTNU (NO)
S. Tviberg, NTNU (NO)
J. Eidsvik, NTNU (NO)
R. Sinding-Larsen, NTNU (NO)
T. Mukerji, Stanford University (US)
|
Abstract: |
Modelling spatial dependency in a gas and oil field is a challenging problem, tackled with many different techniques. In recent years Markov Random Fields and Bayesian Networks (BN) have been used to model the spatial smoothness of some discrete parameters, such as rock types and saturations. One main problem of these models is integrating the expert geological opinion with real data coming from the field, in order to make the model more realistic and effective when used for decision making.
A step forward in this direction is the integration between these statistical spatial models and basin geological modelling techniques, such as Basin and Petroleum System Modeling (BPSM). BSPM computationally simulates the hydrocarbon-generation process and fluid-flow, coupled with sediment deposition, compaction and erosion, to calculate the volumes and the locations of HC available for entrapment, over a time horizon of hundreds of millions of years. The result is a time-dependent 3-D geological model that can forecast the basin-scale behavior of possible HC accumulations; this model is updated and calibrated using existing well logs, and geochemical data.
A detailed study about the translation of the data coming from BPSM into quantitative and probabilistic information is proposed, as a preliminary study for integrating the two modeling techniques. In order to integrate the two approaches it is first important to assess the impact that the input BPSM parameters and variables have on the model output. Inputs are typically a depth-structural map of the area of interest, the age history of the layers, source rock properties and several constituency parameters, such as heat capacity, thermal conductivity, porosity and permeability.
The main idea is to construct a BN model for decision making, consistent with the results of several BSPM outputs. The parameters of the BN conditional probability tables are learned from the BSPM output, while the structure of the network is usually a priori fixed following physical and geological arguments. Typically, we will have a hierarchy of parent nodes corresponding to major structural and constituency parameters (heat flows, rock properties, traps), and a number of child nodes corresponding to different accumulations. Decision theory based techniques are used to asses the value of information related to imperfect seismic tests, and explore the sensitivity of the original BPSM inputs. |
