A stochastic model to predict the success of fracture stimulations in the Cooper Basin
Engineering Honours Degree 2008
University of Adelaide
Since its introduction in the 1940s, fracture stimulation has grown to become a critical technology in the oil and gas industry. Despite this, some attempted fracture treatments still fail. Fracture treatments in the Cooper Basin are certainly no exception to this; the fracturing environment in the Basin is a challenging one. The subject of this report is the investigation and assessment of a newly suggested method to predict in advance the likely success of fracture treatments.
The method uses cluster analysis to characterise existing data in such a way that predictions using new data can be made. The clustering algorithm used was Multi-Resolution Graph-Based Clustering.
A set of logs was chosen as the data to be used for the model. Each depth in each well is treated as one data point for the clustering analysis. The algorithm calculates the suitability of each point to be the 'kernel' of a cluster. It then constructs clusters around the most suitable points. A set of dfferent cluster congurations, at different resolutions, is presented for the user to choose between. These clusters may be thought of as electrofacies. Each electrofacies has some range of production values associated with it, derived from existing data.
The prediction stage involves using the well log data from a new interval to assign each depth sample to an existing electrofacies. The production associated with this electrofacies is used to make a prediction.
The first phase of investigation used data from 23 Santos wells, from across the Cooper Basin. The objectives of this phase were to test the impact of two clustering parameters, to test the impact of the choice of model training data, and lastly to obtain an indication of whether the model was capable of generating results which are reasonable enough to warrant further analysis.
It was determined that changing the number of electrofacies used in the model had little effect on the results. Similarly, changing the number of nearest neighbours used in propagation achieved little more than a supercial 'smoothing' of the predicted log.
The choice of training data, however, was perceived to have a signicant effect on the quality of the results achieved. Data representivity, therefore, was identied as the crucial factor. To a lesser extent, the amount of training data available is also important.
In addition, the results which were being achieved were seen as promising enough to merit further investigation.
The second phase of investigation was intended to determine if using training data from within the same field as the interval for which prediction was required would satisfy the data requirements mentioned above.
The first investigation, using data from Field A, yielded frequency plots which looked reasonable. Inspection of the logs themselves, however, revealed some deficiencies. The electrofacies assignments were not as suitable as had been hoped. This was resulting in production prediction which was not optimal. The investigation in Field B yielded similar results.
It is thought that the unavoidable division of the data into producing and nonproducing zones, in combination with the unusual production log is causing diculties for the clustering algorithm due to the unevenly populated and strangely shaped clusters which may have eventuated.
The poor predictions of production may also be a result of lithological factors having a less substantial inuence on the fractures than previously thought.
To conclude, using cluster analysis to characterise well log data in order to predict in advance the success of fracture stimulation treatments, as described in this report, is in general not able to produce results of the accuracy required.