Application of Artificial Intelligence (AI) Based Algorithm to Predict Contamination During Formation Fluid Sampling

Abstract

Reservoir fluid sampling using the wireline or LWD tools has long had a difficult time predicting the filtrate contamination flowing through the tool. When the conventional curve fitting method was developed, changes in measured dynamic fluid data were thought to be best explained by fitting the entire data to a power law equation. Results from simulation study shows that the complete change cannot be explained by a single exponential fit, especially the later time data which is most important for predicting contamination. Thus, the conventional method requires a selection of data just during the later time to give better results. The results are thus extremely dependent on the user and the data used to create the fit. A novel method was developed which uses an artificial intelligence (AI) based algorithm which was created using an extensive database of over 15,500 numerical simulations based on the fluid flow dynamic and accounting for most of the varying reservoir conditions for contamination prediction. Eight different models were developed and tested to provide not just accuracy of the results but also keeping in mind that this is a real-time decision-making tool, the time and computing power needed as well were considered. Multiple sets of fluid sampling data from wireline and LWD tools were run through the contamination prediction algorithm in this study. Each dataset had multiple sensor responses and for each set of results at least 3 sensor responses were selected to be run through the conventional as well as the new algorithm. These results were then compared to lab determined contamination values to determine the error from each method. Based on results from the multiple sets, this technique provides good predictions on the level of contamination not only in the conventional probes for which it was designed but also for focused sampling and sampling in LWD environments. This contamination prediction algorithm removes this user ambiguity and gives consistent and reproducible results.