4D seismic and mulitple-point pattern data integration using geostatistics

4D time-lapse seismic data consists of observing the impact on velocity of changing fluid properties by examining the difference between successive 3D seismic surveys. Thus 4D seismic provides the possibility of monitoring fluid flow and managing reservoir production. Because of the different volume supports, the traditional direct point-to-point correlation between 4D seismic and 4D saturation variables is typically low (0.3). Multiple-point statistics, stemming from principal component analysis (PCA) and canonical analysis (CA), are proposed to extract pattern correlation between 4D seismic and 4D saturation time-lapse data. It is found that the pattern correlation of the two first PC's (one for seismic and one for saturation) is high around 0.7, which reflects the excellent visual pattern correspondence. Canonical analysis increases that pattern correlation to 0.8. Such improved correlation could be used to monitor time lapse saturation values from the 4D seismic data.; 4D seismic data is used to improve an initial geological model, representing a channel-type clastic reservoir. The discrepancy between the forward simulated and the observed 4D seismic data can be used to spot critical areas where the initial model needs to be improved. The correction method although heuristic is flexible and fast, therefore practical. The improved model still honors the initial geological scenario because all corrections are constrained by the training image. Reservoir performance prediction with the new geological model was improved significantly in terms of predicting water breakthrough and saturation distributions.; A structurally accurate reservoir model with a geologically consistent facies distribution is critical for reservoir performance prediction through flow simulation. Multiple-point geostatistical simulation algorithms, such as SNESIM (Single Normal Equation SIMulation) and FILTERSIM (FILTER-based SIMulation), can be used to generate alternative (stochastic) reservoir facies models conditioned to various data types while remaining consistent to any geological scenario depicted by a training image. However, many practical issues, such as simulation with non-stationary features, CPU and RAM costs, target statistics reproductions, are still lacking. In this thesis various methods are proposed to improve both the SNESIM and FILTERSIM algorithms while preserving the geological realism of the reservoir models produced.