--> Innovative Log Reconstruction Workflows for Seismic Reservoir Characterisation Through Mitigation of the Challenges of Extremely Sparse Data Set

AAPG/SEG International Conference & Exhibition

Datapages, Inc.Print this page

Innovative Log Reconstruction Workflows for Seismic Reservoir Characterisation Through Mitigation of the Challenges of Extremely Sparse Data Set

Abstract

Abstract

Kuwait Oil Company is in early development phase of deep carbonate play, depth range of 13500-16500ft, in northern part of Kuwait. This play, consisting of eight fields has a regional cap in the form of a series of high-pressure salt-anhydride layers, locally known as Hith-Gotnia. This area also has mature simultaneous operations through shallow Cretaceous reservoirs. Because of the depth of the play and the presence of HP-HT overburden section, a number of casing strings are utilized before reaching the reservoir section, making it challenging to acquire reliable basic shallow logs, velocity and VSP data for seismic reservoir characterization work. To overcome this challenge of filling the missing shallow log section, a systematic workflow is applied to generate pseudo logs in the shallow part and carefully splice these logs with the logs from reservoir section.

An innovative workflow utilizing the multi resolution graph-based (MRGC) neural network approach was adopted for predicting the missing log data and replacing the poor quality log intervals. MRGC is a non-parametric method, combining normalized neighbouring index and ranking of a kernel representative index for initial clustering. The final ordering of cluster is performed by Coarse-to-fine Self Organizing Map. This takes advantage of both the approaches in detecting cluster in data set of any dimensionality and complex configuration. These clusters are used to carry out supervised training of the input data in the training wells. These trained clusters are then applied on the input data of wells needing log prediction. Similarity threshold method was used to check similarity of logs in all the wells for better choice of input data in model. Uncertainty of the input log is also managed by back propagation method to check ambiguity of each input and output. These predicted logs are further validated by comparing with the mud log information, in the zones not having open-hole logs like in Hith-Gotnia interval. The predictions are done separately for each field and for specific stratigraphic intervals to take care of the local variations.

Through this process prediction in 52 deep wells has been completed utilizing 10 training wells. Utilizing these reconstructed logs, it is now possible to have more robust synthetic seismograms from the surface for these deep wells. This, in addition to facilitating better seismic correlation gives needed reliable low frequency model for inversion work.